It guided fiscal Q3 to $33.5 billion of revenue at roughly 81% gross margin.

The stock then fell as much as ~31% from its March 18 high of $471.34 — bottoming near $323 intraday before rebounding to ~$338 by month-end.

The reality: Micron is tanking because the market is attacking the duration of a memory windfall, not because Micron stopped printing near-term numbers.

This is not one bad headline. It’s six or seven negative catalysts stacking within a two-week window, hitting a momentum name that had already priced in excellence.

What caused Micron to sell off?

Tier 1: The Core Repricing

1. Peak-Margin / Cycle Fear From Micron’s Own Spending Ramp (~30%)

This is the single biggest driver. Micron’s numbers are exceptional, but the market is already looking past them.

The company lifted FY2026 capex above $25 billion and said FY2027 capex will step up meaningfully again, with construction-related capex expected to rise by more than $10 billion year over year. In memory stocks, that reads as “today’s shortage can become tomorrow’s glut.” The multiple compresses before the earnings do.

Memory has a brutal cyclical history. Every previous supercycle ended with companies investing at the peak, supply overshooting demand, and margins cratering. The market has been burned by this exact pattern with MU multiple times. When it sees $25B+ stepping up to $35B+, the reflex is to pull forward the next-cycle debate — even if current supply remains structurally constrained.

The bull counter: Micron’s management said on the Q2 call that customers are receiving only 50–67% of their memory requirements and that supply constraints persist beyond 2026, with meaningful new capacity not online until FY2028 at the earliest. The capex is responding to confirmed demand, not speculating on future demand. But the market doesn’t wait for FY2028 to price the risk.

2. TurboQuant and the Broader AI-Efficiency / Memory-Displacement Narrative (~25%)

On March 24, Google Research published TurboQuant — a compression algorithm achieving at least 6x reduction in KV-cache memory for large language models with zero measurable quality loss. The stock dropped 3.4% on the news and hasn’t recovered.

What the market heard: “AI needs 6x less memory.”

TurboQuant targets the KV cache — a specific data structure used during inference — and Google says it helps vector search.

Where the market overreaches: jumping from “one inference-memory bottleneck gets more efficient” to “HBM scarcity and memory demand broadly collapse.”

BofA’s Vivek Arya noted that comparable compression techniques have existed since 2024 without altering hardware procurement at scale. Google itself is guiding $175–185B in 2026 capex — up ~100% YoY — which contradicts the idea that the company publishing this paper believes it reduces the need for memory hardware.

TurboQuant is also not the only thing feeding this narrative. MatX raised $500M in February to build a custom LLM chip claiming 10x better training performance than NVIDIA GPUs, using a hybrid SRAM-HBM architecture.

Etched raised $500M at a $5B valuation the month before with a similar pitch. Groq and Cerebras have been making SRAM-first designs for years. The narrative that the industry is actively trying to displace HBM has real startups with real money behind it.

But the narrative runs straight into Jevons paradox: in a market where customers are receiving half of what they need, making memory more efficient doesn't reduce demand — it reallocates it.

Memory freed up by compressing inference doesn’t sit idle.

It gets absorbed by the next workload in a queue that stretches years into the future:

• A training run that was waiting for capacity

• RAG deployment that couldn’t scale

• Longer context window that wasn’t feasible before

• Agentic systems that need persistent state

When you’re only serving 50–67% of customer requirements, there is no scenario where making one workload more efficient results in less total HBM being consumed. The backlog absorbs everything.

And every startup trying to work around the memory bottleneck confirms that memory is the bottleneck. MatX’s chip combines on-chip SRAM with HBM sourced from the same memory makers that supply NVIDIA. Its closest competitor Etched plans to source memory wafers from SK Hynix, Samsung, and Micron. These chips don’t displace memory demand — they add more buyers to a market where the existing buyers already can’t get enough.

More fundamentally: you can optimize inference all you want, but breakthroughs at the frontier — longer context, multimodal, world models, whatever comes next — require more raw memory.

Efficiency gains at the inference layer fund the next generation of training at the frontier layer. Memory gets reallocated upward in the stack, not returned.

Per-GPU memory content has grown roughly 12x from the A100’s 80GB to over 1TB projected for Rubin Ultra. As SemiAnalysis put it, modern AI follows a “memory-Parkinson” dynamic where architectures “relentlessly grow to occupy whatever HBM becomes available.”

Micron is already shipping HBM4 36GB 12-high stacks in volume for NVIDIA Vera Rubin at >2.8 TB/s bandwidth, and has shipped 16-high (48GB) samples delivering a 33% capacity increase per placement — the next step in a trajectory that only demands more memory per chip generation, not less.

The only scenario where total memory demand actually declines is if efficiency gains outpace both the existing unmet demand backlog and the growth in new workloads simultaneously. Given that the backlog alone represents 33–50% of unmet requirements, that’s essentially impossible in the 2026–2028 timeframe.

None of this stops the stock from de-rating now. The paper was tested only on 8B parameter models, hasn’t been deployed at production scale, and will be formally presented at ICLR in April. In a fear environment, it provided the intellectual justification for selling — and that’s enough. But the fundamental demand case hasn’t changed.

3. Macro Risk-Off: War, Oil, Yields, AI-Financing Stress (~20%)

This isn’t stock-specific, but it’s the foundation layer everything else sits on. Reuters tied the current tech selloff to the Iran war, oil above $100, volatile yields, and skepticism about whether Big Tech can comfortably finance a $635 billion 2026 AI-infrastructure buildout under worsening macro conditions.

Oil above $100 revives recession fears, hitting cyclical sectors first — and memory is the most cyclical part of semiconductors in the market’s mental model. Nikkei dropped 4%, KOSPI dropped 4.5%, yields rose, and the Fed-cut narrative evaporated. The Strait of Hormuz closure disrupted shipping lanes and input costs for Asian fabs.

This is the environment in which every other catalyst gets amplified. TurboQuant in a calm market is a one-day dip. TurboQuant layered on top of a war, rising yields, and AI-financing skepticism is a week-long cascade.

Tier 2: Real But Secondary

4. Mainstream DRAM Price Softness (~10–15%)

This is the factor most MU bulls are underweighting because they’re focused on HBM.

Micron is not a pure HBM proxy. In fiscal Q2, cloud revenue was $7.75 billion, mobile/client was $7.71 billion, and core data-center was $5.69 billion. The mobile/client segment is a huge chunk of revenue and it’s exposed to commodity DRAM pricing.

Citi noted mainstream DDR5 16GB DRAM spot prices are down ~6% since Micron reported. Even if HBM remains tight, a stock that was pricing extraordinary scarcity across the entire memory stack will react hard to any hint that conventional DRAM pricing has peaked.

The risk: if spot-price weakness leaks from the commodity market into contract pricing over the next 1–2 quarters, it creates a margin headwind that the HBM business (while growing fast) may not fully offset.

HBM is the growth story, but DRAM is still the revenue base.

5. Competition: SK Hynix and Samsung (~10%)

SK Hynix remains the dominant HBM supplier (50–57% market share depending on the quarter) and filed a confidential SEC submission for a US listing that could raise $10–14B.

This matters on multiple levels: it gives US investors a directly investable alternative to MU for AI memory exposure, capital goes directly into AI expansion, and analysts noted the listing enables direct valuation comparisons that may highlight SK Hynix’s “undervaluation despite stronger profitability.”

Samsung is also recovering. Its co-CEO said customers told him “Samsung is back” on HBM4, its entire 2026 HBM4 production is sold out, and it’s committing $73B to semiconductor investment in 2026.

A projected HBM market share shift — from ~62/21/17 (SK Hynix/Micron/Samsung in mid-2025) toward 50/30/20 — means Micron’s relative position faces more competitive pressure even as the market grows.

That said, the SK Hynix listing cuts both ways.

If SK Hynix lists at a higher multiple than MU (which analysts expect) — it could actually re-rate MU upward by establishing a higher valuation benchmark for the entire AI memory sector.

None of this means Micron loses near-term demand. But it reinforces the market’s fear that supernormal margins will attract aggressive supply response.

6. Stargate / Abilene / OpenAI LOI Drama (~5%)

Bloomberg reported Oracle and OpenAI cancelled plans to expand the Abilene Stargate campus from 1.2GW to 2.0GW. Combined with reporting that Samsung and SK Hynix had signed letters of intent (not binding purchase orders) with OpenAI, this created a “demand destruction” narrative.

The 600MW expansion was dropped due to financing disputes and shifting demand forecasts — not declining AI demand. The broader 4.5GW Oracle-OpenAI agreement remains intact. NVIDIA deposited $150M with Crusoe to secure the site for Meta. Capacity is being reallocated, not destroyed.

Micron’s disclosed customer commitments are broader than any single OpenAI project. In December the company said it had completed agreements on price and volume for its entire calendar-2026 HBM supply, and in March it announced its first-ever five-year strategic customer agreement. The guide doesn’t depend on Stargate.

This is a narrative amplifier — not a core driver.

Tier 3: Noise the Market Is Treating as Signal

7. ARC-AGI-3 and “AI Isn’t That Smart” Sentiment (~2–3%)

ARC-AGI-3 launched March 25 — inside the selloff window. Frontier models scored under 1% (Gemini 3.1 at 0.37%). Humans score 100%. The headline: “Best AI in the world can’t do what any human can.”

This doesn’t logically affect HBM demand, which is driven by the current generation of models scaling inference — not by AGI timelines. But in a fear environment, “AI scores 0.37% on intelligence test” gives emotional permission to question the entire buildout thesis. It compounds the TurboQuant narrative: maybe AI needs less memory because maybe AI isn’t progressing as fast as the capex assumes.

This is a vibes-level catalyst, not a fundamental one. But vibes move stocks during panics.

8. Helium Supply Crisis (~2% of the current move, but significant tail risk)

Iran struck Qatar’s Ras Laffan Industrial City — roughly 30% of global helium production, now offline with damage that could take years to rebuild. Helium is irreplaceable in semiconductor fabrication: it cools wafers during lithography, flushes toxic residue, and is essential for EUV leak detection. South Korea imported 64.7% of its helium from Qatar in 2025. Spot prices have surged 40–100%.

This matters structurally but isn’t driving the current MU drawdown. Korean chipmakers reportedly have helium inventories through approximately June 2026. It’s a tail risk, not a present trigger.

The paradox: if the helium crisis extends beyond 6 months, it constrains production industry-wide — which actually validates the supply-shortage thesis and protects memory pricing. The market isn’t parsing this yet.

Also underreported: every HDD at 10TB+ uses helium as a sealed internal gas. Seagate and Western Digital have sold out 2026 production. If the helium crisis constrains HDD capacity, it could become a tailwind for Micron’s SSD business.

9. Clay, NY Lawsuit and Leftist Political Overhang (~1%)

In January 2026, Jobs to Move America and a local residents’ coalition filed suit to halt construction of Micron’s $100B Clay, New York fab, challenging the environmental review process. A deeper investigation found the lead plaintiff group had never held a meeting before the suit was filed and some members didn’t even know who the others were.

The most recent New York news was actually positive — state approval for a key power-line project and Micron saying ground preparation is ahead of plan. Background noise. Doesn’t affect 2026–2027 revenue trajectory.

10. YMTC Patent Suit, “Sharks,” and Assorted Noise (<1%)

YMTC’s London patent suit is part of the ongoing US-China tech rivalry and is not a new development. As for institutional manipulation: David Tepper increased his Appaloosa MU position by 200% in Q4. The analyst consensus target is $515–530 with 38 buys against 2 sells. Smart money may be buying the dip, but “sharks intentionally forcing it down” requires position and flow data that doesn’t exist publicly. A crowded AI trade unwinding on bad narrative timing is sufficient to explain this move.

The Structural Bull Case for Micron in 2026

The piece so far has been a fear catalog. Here’s what the fear is colliding with.

Memory is the bottleneck — and it’s getting worse, not better.

Data centers are projected to consume 70% of all high-end memory chips produced worldwide in 2026, up from 20–30% as recently as 2022.

That’s a permanent reallocation of global supply toward AI, per IDC. AI servers use 10–20x more memory per unit than traditional servers. TrendForce projects memory prices rising another 70% in 2026 due to persisting shortages.

The supply response can’t catch up fast enough. New fab capacity takes 3–5 years to build. Global DRAM supply growth in 2026 is projected at just 16% year-over-year — well below historical norms and well below demand growth.

The 3 companies that control ~90% of the market (Samsung, SK Hynix, Micron) have all shifted production toward higher-margin HBM and enterprise DDR5, further starving consumer electronics, automotive, and everything else. The WSJ compared the emerging memory shortage to the automotive production delays during COVID.

NVIDIA’s roadmap shows per-GPU memory content growing roughly 12x from the A100’s 80GB to over 1TB for Rubin Ultra.

Every next-generation GPU platform requires more memory than the last.

The architectural trajectory is unambiguously memory-intensive.

• The broader memory market reflects this. TrendForce projects total memory market revenue at $551.6B in 2026, surging to a peak of ~$842.7B in 2027.

• DRAM contract prices rose 90–95% QoQ in Q1 2026 — an extraordinary move even by memory-cycle standards.

• SemiAnalysis, which has been calling the shortage since late 2024, argues the supply-demand imbalance is “deteriorating rather than normalizing” and believes the market is “not even close to the peak.”

As established above: when customers can only get half of what they need, efficiency gains don’t destroy demand — they reallocate it. The fear narrative requires demand destruction. The data shows unmet demand growing.

Micron’s HBM4 execution is better than the market thinks.

In late January, SemiAnalysis reduced Micron’s projected share of NVIDIA Rubin HBM4 to zero, citing poor pin speed performance from Micron’s internal base die design. They projected Rubin HBM4 would consolidate into ~70% SK Hynix / ~30% Samsung. The stock dipped on the leak.

Six weeks later, at GTC 2026, Micron announced high-volume production of HBM4 explicitly “designed for NVIDIA Vera Rubin” — achieving over 11 Gb/s pin speeds (the threshold NVIDIA required), delivering >2.8 TB/s bandwidth, and shipping 16-high samples.

Either SemiAnalysis was wrong about the exclusion, Micron fixed the problem faster than anyone expected, or Micron’s HBM4 is going to non-NVIDIA customers who are equally hungry. Regardless, the bear narrative on Micron’s HBM competitiveness was overtaken by events before the selloff even began — and the market hasn’t priced that in because TurboQuant and the war buried the signal.

Micron also supplies LPDDR5X for the Vera CPU on Rubin platforms and announced the industry’s first PCIe Gen6 data center SSD in volume production — broadening its footprint across the entire Vera Rubin memory and storage stack, not just HBM.

US policy is directly subsidizing Micron’s expansion.

The One Big Beautiful Bill Act, signed July 4, 2025, delivered multiple tailwinds to the US semiconductor industry simultaneously. Domestic R&D costs can be fully expensed in the year incurred again, reversing the TCJA’s 5-year amortization requirement.

The Section 48D advanced manufacturing investment credit for semiconductor fabs rose from 25% to 35%. And 100% bonus depreciation was made permanent.

These provisions make Micron’s $25B+ capex commitment look more rational than the raw number suggests. The after-tax cost of domestic expansion is meaningfully lower than the headline figure, and the policy environment actively incentivizes exactly the kind of investment Micron is making.

Micron has also disclosed eligibility for up to ~$6.165B in CHIPS Act grants. The cycle-fear argument treats capex as pure risk; the tax code treats it as subsidized strategic investment.

This also matters for the competitive landscape: SK Hynix’s Indiana packaging plant benefits from the same incentives. Yet Micron — while it manufactures globally across Singapore, Japan, and Taiwan — is the only one of the Big Three domiciled in the US and is making its largest domestic expansion in company history with the $100B Clay, NY fab.

Micron Stock: ATH Odds & Timelines

These are judgment calls, not model output. The key insight: time helps the recovery case only up to a point. The same new fabs and packaging capacity that support Micron’s long-run growth also raise the risk that 2027–2028 becomes more about supply normalization than scarcity.

Will MU revisit the $470 high?

• By June 2026 (~15% odds): Requires war resolution, no hyperscaler capex cuts, and a rapid sentiment reversal. Three months is very fast for a 45% recovery from here.

• By March 2027 (~50% odds): Requires Q3 earnings confirming the $33.5B guide, at least one more quarter of sustained HBM demand, and macro stabilization. This is the realistic base case for recovery.

• By March 2028 (~60% odds): More time, but also more supply coming online. The cycle-turn risk offsets the longer runway.

Will MU clear $500 (new ATH)?

• By June 2026 (~5% odds): Near-impossible without a major positive catalyst — war ceasefire, blowout Q3, and HBM4E NVIDIA qualification all in the same quarter.

• By March 2027 (~35% odds): Requires the full bull case: AI memory supercycle validated across multiple quarters, Micron maintains or grows HBM share, conventional DRAM pricing stabilizes.

• By March 2028 (~45% odds): Plausible if the HBM4/4E generation delivers and competition doesn’t compress margins faster than the TAM grows.

The reason these odds don’t explode higher with time: Micron and its rivals are now spending so aggressively — SK Hynix bought $8B in ASML EUV machines, Samsung is spending $73B in 2026 alone, and Micron is at $25B+ stepping up to $35B+ — that investors are pulling forward the next-cycle debate. The supply normalization risk in 2027–2028 is real and partially offsets the longer recovery runway.

What Actually Matters Next

Signals that MU recovers:

• Microsoft, Amazon, Alphabet, and Meta keep 2026 AI-capex plans intact through Q2 earnings

• Micron keeps repeating that supply-demand conditions remain tight beyond 2026

• Conventional DRAM spot weakness stays confined to spot markets and doesn’t leak into contract pricing

• HBM4 for Vera Rubin is already in volume production; HBM4E qualification for Rubin Ultra on schedule would extend the runway

• War de-escalation or ceasefire reduces oil/helium/macro risk simultaneously

• SK Hynix US listing creates a sector re-rating rather than rotation out of MU

Signals that the bear case is right:

• Any Big 4 hyperscaler cutting 2026 AI capex guidance

• DDR5 contract pricing rolls over, compressing Micron’s blended margins

• Samsung HBM4 takes significantly more than projected 30% market share

• Helium crisis extends beyond 6 months and forces industry-wide production curtailments

• A second major AI efficiency breakthrough beyond TurboQuant that actually reaches production deployment and demonstrably reduces hardware procurement

Bottom Line: Micron (MU) in March 2026

Even after a partial rebound to ~$338, MU is trading around 4–5x NTM P/E — compressed even by memory-stock standards.

• The bull case: Micron is still guiding record Q3 revenue, says AI and traditional server demand remain constrained by inadequate supply, has begun HBM4 volume production explicitly for NVIDIA Vera Rubin (overcoming a SemiAnalysis exclusion call from six weeks earlier), and is expanding domestic production with enhanced tax incentives that reduce the after-tax cost of its capex ramp.

• The bear case: Micron and its rivals are spending so aggressively on new capacity that investors are pulling forward the cycle debate — and conventional DRAM pricing is already softening.

The $33.5B in guided revenue seems legitimate for the next couple of quarters, but many question whether supernormal margins can be sustained.

Short-term traders are avoiding due to war, DRAM spot prices softening, and Twitter threads about compression algos and Stargate funding. I’d guess most likely a Jevons paradox effect in HBM emerges and HBM stocks including MU bounce back.

The thing that actually matters: whether HBM demand and hyperscaler capex hold — hasn't broken.

Every major customer is still spending, Micron's HBM order book is contracted through 2026, and 70% of high-end global memory production is being consumed by an AI buildout that every major company on earth is accelerating, not slowing down.

DISCLAIMER: NOTHING HERE IS FINANCIAL OR INVESTMENT ADVICE.

One-shotted by a peer-reviewed “model” and now set in stone within their psyches.

These are the people who are genuinely confused about why Canada’s GDP per capita has been declining for years — the worst 5-year decline since the Great Depression — despite decades of “skilled immigration to boost GDP” and “social democracy.”

These same people cannot understand why test scores are dropping in developed countries despite asinine levels of spending on public education; they’ve banned reality anytime it doesn’t fit the consensus egalitarian fantasy.

The Fraser Institute explicitly identified mass immigration as a primary cause: high volumes of immigrants lowered labor costs and suppressed business investment in capital and technology. Canada fell below the OECD average for the first time in recorded history in 2024, its World Happiness ranking collapsed from 5th to 18th, and the U.S. now produces nearly 50% more per person — a gap unprecedented since 1945. The model failed. The modelers won’t admit it.

They classify “immigrants” as a monolith and cite aggregate statistics that wash out every meaningful distinction.

Which specific immigrants? From where? With what skills? At what fiscal cost? How will they vote? How will their children perform?

Most Western countries do not systematically track lifetime fiscal net present value, crime rates, assimilation rates, or intergenerational outcomes by race, ethnicity, country of origin, IQ, or skill level.

When the UK’s ONS was asked directly via FOI for crime data by nationality, the answer was: “we do not hold the information you have requested.” The Home Office admitted its data isn’t reliable.

Denmark is virtually the only Western country that collects conviction data by origin — and 40+ nationalities exceed the Danish rate for violent crime.1

Conservatives were dismissed for decades as “racists” and “nativist rubes” for resisting this. They often couldn’t articulate why and didn’t have the studies, models, or academic vocabulary the credentialed class demands. But the instinct was directionally right, and it may have unironically prevented a bigger disaster than we currently have.

Every year populist resistance slowed the open-borders agenda was a year the damage was contained.

Credentialed models failed and are still failing (see the CATO Institute, whose immigration “research” is functionally indistinguishable from left-wing open borders advocacy); institutional reps remain deliberately obtuse for a paycheck without considering cascading multi-generational effects or acknowledging that their models are essentially a house of cards.

Nobody serious about ensuring the U.S. stays dominant is against elite world-class-skilled immigration — that has always been a win-win.

But only ~15% of legal permanent residents enter through employment-based categories — and of this subset many are not really much more skilled than Americans (they are often just cheaper labor, using highly questionable/fake credentials, or have credentials that are much lower quality than the U.S.).

Perhaps the most damage is done when a “skilled” immigrant leverages his/her skill to advocate for policies that disrupt the foundational principles of the U.S. (e.g. a lawyer who fights hard for open borders, freeing criminals, and importing more of their ethnic group).

The remaining ~85% is chain migration selecting for kinship, not skills.

Add 1-2+ million unauthorized entries per year with zero skills filter, and the ratio of skills-selected to non-skills-selected is roughly 1:6 or worse. Even within the “skilled” category, H-1B workers earn 16% less than comparable natives — it’s a wage-arbitrage pipeline, not a genius-recruitment program.2

And even for genuinely elite talent, the assumption that admission requires a pathway to citizenship and voting rights is a policy choice, not a necessity. The UAE and Singapore host millions of skilled expatriates without citizenship. Many top talent would live in the U.S. either way — they want American markets and quality of life, not a ballot.

I’m not arguing against helping other countries.

I’m arguing that the U.S. cannot help anyone from a position of self-inflicted decline.

If we let the country become overwhelmed by leftism (handouts, degrowth, socialism, ethnic tribes without assimilation), we can’t help anyone — and before we realize it, we can’t even help ourselves.

Have a look at South Africa. Preservation isn’t selfishness. It’s a prerequisite for generosity.

Here’s what reality says.

Cheap Labor Is the Enemy of Productivity

In 2007, a New York dairy farmer named Dale Hemminger had a worker arrested by immigration authorities. He installed robotic milking machines.

His farm went from 800,000 pounds of milk per worker per year to 2.5 million — a 3x productivity gain. His remaining workers earn more, work shorter hours, and do less grueling labor. The New York Times column that profiled him presented this as a vision of the future — then argued we should preserve the conditions that prevent it. Daniel Kishi’s rebuttal dismantled that framing point by point.

This isn’t an anecdote.

Mann & Pozzoli (2024), in a study published in the Journal of Economic Behavior & Organization, exploited a natural experiment in Denmark where refugees were quasi-randomly assigned to municipalities, proving the mechanism causally: a one percentage point increase in the share of non-Western migrants decreased firm-level robot adoption by 7% and robot import values by 15%.3

More immigrants → lower wages → less incentive to automate.

About 25% of dairy farms in Denmark and the Netherlands use automated milking systems. In the U.S., roughly 5%. Lely commercialized the first system in 1992. The gap isn’t technology — it’s incentives. H-2A visas exploded from ~50,000 in 2005 to 385,000+, making mechanization economically irrational.

Florida shows the first phase. After mandating E-Verify in mid-2023, average real wages rose 6.4% in 32 months. Prior 32 months? Down 3.0%. Pre-COVID to passage? Flat. That's the labor market repricing — employers paying more for a tighter labor pool. It doesn't yet prove productivity gains; those come in step two, when higher labor costs trigger capital investment, which takes years.

Denmark and Hemminger show what phase two looks like: 3x output per worker, fewer employees doing less grueling work for more money. Florida is the wage signal. The automation payoff comes next — if policy holds.

Automation isn’t a fallback — it’s the superior long-term path regardless.

Robots don’t need healthcare, collect benefits, commit crimes, vote, require schools, and reshape the culture.

They get cheaper every year.

The current high cost reflects a 5% adoption market — scale it up and costs collapse like solar, semiconductors, and every other mass-deployed technology.4

Cheap labor didn’t just delay this transition. It actively prevented it and every year of delay is lost productivity compounding against the national interest.

The Fiscal Math Is Predictable — and Devastating

A chart from Jonatan Pallesen, drawing on Dutch fiscal data from van de Beek et al. and national IQ estimates from Parra & Kirkegaard (2025), plots lifetime net fiscal contribution per person by immigrant origin group in the Netherlands.

The R² is 0.65. Origin-country average IQ alone explains two-thirds of the variance.

Immigrants from North America, Japan, and Northern Europe are net positives in the €400-500K range.

Immigrants from the Middle East, North Africa, and sub-Saharan Africa approach negative 1 million euros per person over a lifetime including the second generation.

The NAS 2017 report found the same in U.S. data: immigrants without a high school diploma are net negatives of -$115K to -$170K+ per person.5

This almost certainly understates the true cost, since the model doesn't fully capture displacement effects on native workers, automation suppression, state/local overhead, charity costs, or the downstream fiscal burden of second and third generation outcomes.

The distribution is the key: short-term benefits concentrated among employers who access cheaper labor, diffuse costs spread across the native population — heaviest on the least advantaged Americans.

But even the employers lose long-term: the political transformation they helped enable through cheap-labor dependence produces higher taxes, heavier regulation, and institutional decay that eventually makes their own businesses unviable.

A regressive transfer dressed up as humanitarianism.

Gaps Don’t Close

If these trajectories were temporary and immigrant-origin populations converged to the host-country mean — the picture would be less alarming.

A century of data says they don’t.

Jensen & Kirkegaard (2025), meta-analyzing 139 U.S. studies from 1918-2017 (N ≈ 400,000), found the Black-White IQ gap essentially unchanged for 99 years: -17.04 at the start, -17.06 at the end.

Across a century of radical environmental change: end of Jim Crow, desegregation, affirmative action, Head Start, massive spending increases — the gap didn’t move.

Publication bias deflates it; correcting for it makes it larger.6

The Hispanic-White gap (-9.9 points) shows apparent narrowing, but this likely reflects compositional changes in who identifies as Hispanic, not convergence; 2025 SAT data shows the same distributions at the same magnitudes as every prior year.

Crémieux’s analysis of Jewish IQ — meta-analytic mean ~110 for Ashkenazi Americans, stable across a century — demonstrates that group cognitive advantages are equally real, persistent, and productive.

The symmetry is the point.

If advantages persist and produce superior outcomes, deficits persist and produce inferior outcomes — including the fiscal trajectories documented by van de Beek.

Telles and Ortiz tracked Mexican-American families across four generations and found persistent gaps in education and income into the fourth generation. The “convergence” that would make the fiscal math self-correcting isn’t happening.7

Transformation Is Already Locked In

The fertility differential seals it.

Non-Hispanic white TFR is ~1.5. Hispanic TFR is ~1.9-2.0. Non-Hispanic whites went from ~80% of births in the 1990s to ~50% today.

Meanwhile native workers at the bottom — disproportionately Black and Hispanic Americans — are being displaced and squeezed by 3-8% wage depression from immigration-induced supply increases.

The primary domestic victims of mass low-skill immigration are the most vulnerable Americans.

Social trust collapses alongside the labor market. Robert Putnam — Harvard, liberal — conducted the largest study ever on civic engagement: 30,000 people, 41 communities.

He delayed publication for years because the results horrified him.

Greater ethnic diversity → less trust of neighbors of any race, including one’s own.

Less volunteering, less charity, fewer friends.

People “hunker down like turtles.” Negative in 39 of 41 communities. Trust halved in the most diverse settings.8

The new demographic majority doesn't just change what the electorate wants.

Different populations carry different distributions of political preferences, time orientation, trust, individualism, and attitudes toward government. As these populations grow as a share of voters, both parties reshape themselves to compete for them — and both move in directions away from what made the already-successful country successful.

How is that playing out with the Hispanic demographics?

• The left gets more redistributionist as immigrant-origin voters who favor larger government and more handouts (at ~75-25 margins) grow the coalition.

• The right gets more populist and fiscally liberal as working-class Whites react with cultural protectionism and Hispanic Republicans demand strong-arm social conservatism, bigger gov, and more handouts for their causes.

The old consensus around limited government and fiscal restraint loses its electoral constituency as the population that built it shrinks.

California is the proof of concept.

1. Mostly White and a Republican lock in the past.

2. Now? Mostly Hispanic and a leftist uniparty with the worst inequality, homelessness, and middle-class flight in the nation.

The politics, fiscal trajectory, and institutional quality followed the demographic transformation predictably.

The Timing Trap

We are at the exact moment where AI and robotics could let the U.S. do more with fewer workers and build the most productive economy in history.

Easterly & Kraay showed across 157 countries that small states are 50% richer per capita than regional neighbors — a productivity advantage, not a scale advantage.

Population quantity doesn’t drive prosperity. Population quality does.

But the window narrows every year. Every year of the current trajectory adds fiscal liabilities that persist for generations, suppresses automation investment, and dilutes the institutional base.

The left currently holds two positions simultaneously: (1) AI and robots will take most jobs within 5-10 years, and (2) we need more immigrants.

Make it make sense. Banning AI and adding more immigrants is a surefire way to destroy the country and cede global power to competitors with common sense.

If you want what’s best for the United States and humanity, you should be (A) pro-AI/automation, (B) anti-unfiltered immigration, (C) pro-elite, actually skilled, and actually necessary immigration (not some cookie cutter STEM/H1b scheme).

And currently we need a moratorium on most immigration to digest/assimilate the sheer quantity that entered under Biden; this may take a while.

The economic argument was “who will do the work?” AI and robots are answering that. The fiscal argument was always negative. The convergence hasn’t happened in 99 years. The political transformation is well underway.

We are importing the costs, delaying the technology, and locking in the politics — all at once. The automation revolution offers an off-ramp.

The question is whether we take it before the exit is behind us.

Fix things after they break.

• The industry spends ~$250B/year across 480,000+ clinical trials.

• Over 200 oncology drugs approved with 6,300+ more in the pipeline — meaning every new entrant is fighting thousands of competitors for the same patients, same trial sites, same FDA reviewers

• After $300B+ on Alzheimer’s, only 2 drugs exist with marginal efficacy — the worst ROI in the history of pharma R&D

• Each new drug costs $2.6B, takes 10–15 years, and has a 7.9% approval rate — meaning >90% of programs fail completely

• Cancer 5-year survival improved ~20 percentage points in 50 years — 0.4pp/year and decelerating, despite exponentially increasing spend

• Cost per QALY has risen from ~$50K in the 1990s to $100K–$500K+ today — each incremental gain costs more than the last

This paradigm is not failing — but it is mature, oversaturated, and structurally incapable of solving the upstream problem that causes 75% of all disease and death: biological aging.

And it has zero capacity to address the even larger opportunity beyond disease: making healthy humans better.

Three targets that should exist as biomedical industries but don’t:

1. Age reversal: Eliminating the root cause of most disease.

2. Bioenhancement: Making elite/healthy humans better at the molecular and genetic level.

3. Artificial augmentation: Boosting human capability through hardware, bionics, and brain-computer integration.

They don’t exist because Western regulatory, cultural, and institutional frameworks block them — not because the science is impossible or demand is absent.

Massively net-negative DEI-driven research burns funding on literature pollution instead of innovation.

Anti-meritocratic selection in academia filters for compliance over competence. Institutional groupthink protects status over progress.

Academic elites are enraged that a tech bro went “Founder Mode on Cancer” and had success; they’d rather you die if it helps them maintain social status and power (so they can virtue signal how irresponsible you are for trying to cure yourself).

Bloated regulations and predatory lawyers prevent companies from experimenting even on terminal patients.

Bioethicists who I’ve already argued are unethical have caused quantifiably insane amounts of harm (suffering and death) via philosophical grandstanding, moralizing, gatekeeping, and institutional influence. Because of them we guarantee deaths, prevent human experimentation in terminal patients, and we get zero valuable data.

The result: Stagnation and decay.

Disease treatment is one engine running at max capacity with diminishing returns (and even this could be massively improved by cutting regulations).

BUT AGE REVERSAL, BIOENHANCEMENT, AND ARTIFICIAL AUGMENTATION (TRANSHUMANISM) ARE 3 INDUSTRIES THAT HAVEN’T BEEN SERIOUSLY TRIED.

The first sovereign nation to build infrastructure across all three verticals won’t be competing with existing industries — it will be creating new ones that incumbents are ideologically and regulatorily prohibited from entering.

• The TAM is potentially every human alive.

• The competitive moat is a regulatory framework no other jurisdiction offers.

• The timeline to first-mover advantage is ~5–10 years.

• The cost of delay is measured in lives: 100,000 people die of age-related causes every day and the productivity boost from making elite humans “more elite” is potentially staggering.

Here I make the case that a sovereign wealth fund or national government should commit $50–100 billion over 10 years (Phase 1), scaling to $150–200 billion+ over 20 years.

I estimate that every dollar deployed will yield at least 100:1+ expected ROI and thus should be deployed as rapidly as infra can absorb it.

Part I: Biology Is Ready — Institutions Are Not

Engine 1: Age Reversal

Aging is the upstream cause of ~90% of cancer, 100% of Alzheimer’s, ~80% of cardiovascular mortality, and an estimated 75% of all deaths in developed nations.

~100,000 people die of age-related causes every day — 36.5 million/year.

The NIA spends ~$346M/year on aging biology — less than 1% of NIH’s budget.

All-time global investment in true age reversal is estimated at $15–20B. That’s what the US spends on healthcare in two days.

Only ~20–30 companies pursue epigenetic age reversal. Zero therapies approved. One human clinical trial exists (Life Biosciences, 2026).

We spend more annually on erectile dysfunction drugs ($5B+ market) than the NIA spends on the biology of aging.

The question “can biological aging be reversed?” was settled in 2024.

• Cano Macip et al. (2024) — Gene therapy delivering inducible OSK factors to aged wild-type mice (~80+ human years equivalent) extended remaining lifespan by 109%. Frailty and organ deterioration reversed. Standard gene therapy vector, in principle translatable to humans.

• Shift Bioscience (2025) — Single reprogramming factor (SB000) achieves cellular rejuvenation comparable to full OSKM without the cancer-inducing pluripotency. Directly addresses the primary safety concern.

• Life Biosciences (2026) — FDA cleared the first human clinical trial of epigenetic reprogramming. Expected readout ~2028–2030.

Additional signals:

• Klotho gene therapy (15–20% lifespan extension in mice), telomerase gene therapy (41.4% median extension)

• Rejuvenate Bio’s three-gene cocktail simultaneously reversing obesity-diabetes-heart failure-kidney failure in mice, plasma dilution showing 2.61 years of biological age reduction in humans.

Five generations of reprogramming technology show clear safety convergence — from full OSKM (max rejuvenation, high cancer risk) through cyclic OSKM, OSK minus c-MYC, single-factor approaches, to emerging small-molecule reprogramming.

Each iteration reduces risk while preserving efficacy.

The path to safe, controllable reprogramming therapy requires investment and iteration, not fundamental breakthroughs.

Biology is not the bottleneck. Institutions are.

Engine 2: Bioenhancement (RNA Sandbox + Permanent Lock-Ins)

The optimal bioenhancement intervention in 2026 is programmable biology:

• siRNA (silence any gene temporarily)

• circRNA (express any protein for weeks without touching the genome)

• base/prime editing (permanently install any point mutation with single-nucleotide precision)

These platforms are being developed for disease in Western clinical trials, but they’re target-agnostic — the LNP that delivers a PCSK9-silencing base editor for familial hypercholesterolemia could deliver that same editor to optimize cardiovascular risk in a healthy 30-year-old. Nobody does this because nobody is allowed to.

The technology is ready but the jurisdiction is missing. The FDA does not approve “enhancement” drugs and there’s no regulatory pathway to getting enhancers approved in any major jurisdiction.

Yet the unregulated nootropics market (mostly snake oil) is $5.5B (growing 15%+/year), the executive optimization grey market exceeds $10B, and DARPA spends an estimated $1–3B/year on classified enhancement R&D.

There is no biological law that says current human cognitive capacity or physical performance represents an optimum or peak — these are evolutionary tradeoffs under selection pressures that no longer apply.

Nobody knows the ceiling because nobody has tried to find it.

And fixation on “enhancement” drugs (which lead to bio tolerance) misses the smartest shortcut: temporary trial run (the sandbox) → lock in permanently if you like (and eventually reverse the permanent mods if you dislike them over time).

The “genetic sandbox then lock-in” strategy (the core value proposition):

Use reversible RNA interventions to test an enhancement in your own body, then commit to permanent gene editing whenever you want (i.e. if you are comfortable with the risk and/or there’s empirical validation, etc.).

The proof case exists — inclisiran (siRNA) silences PCSK9 with two injections/year for ~50% LDL-C reduction. A person could take it for 6–12 months, verify tolerance, then elect permanent PCSK9 base editing via Verve’s VERVE-102 (53–69% LDL-C reduction from a single infusion).

Someone else might opt for cognitive boost via BDNF Val66Met for activity-dependent BDNF (doesn’t treat a disease but may enhance learning).

Decide whether ongoing RNA treatments or permanent lock in. This pattern generalizes across every target where both RNA and editing approaches exist. And you will likely be able to even reverse the “permanent” edits in the future by re-editing back to homeostasis if you become dissatisfied.

Class A: Disease-related targets that enhance: PCSK9 knockout (88% CHD reduction), ANGPTL3 loss-of-function, APOE4→APOE3 conversion (60–90% Alzheimer’s risk reduction), myostatin silencing for muscle enhancement.

Class B: Targets that may enhance: COMT Val158Met, BDNF Val166Met, KIBRA T-allele, DRD2/DRD4 density, SLC6A4, FOXO3, SLC30A8 R325W, ACTN3, EPO pathway tuning, myostatin optimization, etc. (These are just preliminary ideas and not even close to what I have in mind for the full strategy.)

Delivery is cracking open fast.

• Liver: Solved (GalNAc-LNPs).

• Brain: Dyno’s AI-designed AAV capsids achieve 280-fold improvement over natural serotypes.

• Muscle: advancing through DCA-conjugated siRNA and engineered AAVs.

• New platforms (anellovectors, VLPs, synthetic exosomes): All in preclinical or early clinical development.

Enhancement operates across four modalities, each at a different readiness point:

• Programmable Biology (highest EV, most ready) — The sandbox-to-lock-in framework deployed through an Enhancement Operating System (EOS). The EOS was designed by me and involves: eliminating genetic dampeners first, running permutational stack analysis on 50,000+ genomes at trait extremes, finding first-principles shortcuts (the GLP-1 lesson — complex traits have leverage points), installing supportive companion mods to offset trade-offs, using transient bridges, prioritizing asymmetric upside, stacking synergistically, and never stopping iteration as new variants, stacks, and delivery methods emerge.

• AI-Designed Pharmacology (fastest revenue) — Nobody has designed an enhancement compound from first principles optimized for healthy humans. AI-driven molecular design explores chemical space orders of magnitude faster than traditional medicinal chemistry. Targets: selective dopaminergic optimization without addiction liability, orexin modulation for sleep efficiency, clean myostatin inhibitors, selective cortisol modulators. This generates revenue while the gene editing pipeline matures. Ideally you figure out zero-tolerance-pharmacology wherein drugs keep working at stable dose with minimal or zero drop-off in effect (likely less efficient than the RNA + gene mod lockins though).

• Embryo Engineering (generational compounding) — Herasight’s CogPGT detects an average 8.5 IQ point difference between just 3 embryos, validated on 5,000+ sibling pairs. Orchid does whole-genome embryo sequencing (>99% of genome). MIT named embryo scoring a top 10 breakthrough of 2026. The progression: PGS on common variants → full WGS → permutational stack analysis → embryo gene editing → IVG expanding embryo pools to hundreds/thousands per cycle → generational compounding where each cohort’s gains become the next generation’s baseline. A sovereign program removes every constraint simultaneously — openly screen for positive traits, fund IVG, build the genomics engine, develop editing protocols, and track longitudinal outcomes at scale no private company can match. Eventually you could clone fleets of geniuses or other specialists if you wanted.

• Delivery Infrastructure (enabling layer) — The single most leveraged investment. Every modality depends on getting cargo into the right cells. Owning or licensing next-generation delivery creates a chokepoint advantage: anyone who wants to do enhancement in any tissue must use platforms this program developed or controls.

Engine 3: Artificial Augmentation (Transhumanism)

When biological enhancement approaches its limits, the next frontier is bio-hardware integration — BCIs for cognitive augmentation beyond what genetics can deliver, engineered organs exceeding natural performance, bionic systems outperforming biological originals, synthetic biology creating capabilities with no natural equivalent.

$1B+ invested in BCIs alone in 2024–2025. Neuralink has implanted its N1 chip in five patients. Synchron demonstrated Apple Vision Pro control via brain signals. Precision Neuroscience received FDA clearance for its cortical interface. But all current BCIs are medical devices for disability — enhancement applications are 5–10 years from clinical reality.

This vertical is further out and technically harder.

Under dynamic allocation, it receives modest early funding that scales as technology matures. It adds optionality without diverting focus from the two core verticals, and biological enhancement clients are the natural first customers for augmentation technologies as they arrive.

What do I mean by dynamic allocation?

1. The program continuously shifts capital toward whatever is producing the best results.

2. If age reversal is generating 10× returns and augmentation isn’t ready yet, augmentation stays small.

3. If a BCI breakthrough suddenly shows exceptional ROI, capital flows there fast.

Nothing is sacred except the mission and the math.

Artificial augmentation starts with modest funding under this model — it’s further out and technically harder. But maintaining a presence keeps the option alive.

And the enhancement clients from Vertical 2 are the natural first customers for augmentation technologies as they arrive, so the demand pipeline is built-in.

Part II: Why Western Institutions Cannot Lead This

The FDA approves drugs to treat diseases.

• Aging isn’t classified as a disease.

• “Making healthy and smart people healthier and smarter” isn’t treating a disease.

No pathway exists for enhancement gene editing, positive-trait embryo screening, cognitive BCIs in healthy adults, or AI-designed enhancement compounds.

The entire apparatus (IND to Phase 3 to NDA) is built for disease treatment.

Estimated time to reform: 15–25 years, requiring Congressional legislation and political momentum (you probably won’t get the reform unless China creates a Prisoner’s Dilemma by going full throttle age reversal and enhancement… I’m hoping they do!)

Western bioethics treats enhancement and upgrades as inherently immoral. NIH study sections are biased against enhancement proposals. Academic careers punish researchers who pursue it and media coverage is overwhelmingly negative.

The entire pipeline — graduate school to faculty to NIH grant to clinical trial to FDA approval to insurance reimbursement — is optimized for woke status quo and “better disease management” and “healthspan” or whatever bullshit they’re pushing with insanely low ROI.

The US biomedical establishment will not lead this for the same reason Kodak didn’t lead digital photography: the institutional cost of transformation is too high for any individual actor to bear — it will break them mentally.

Part III: The Sovereign Opportunity

These verticals require patient capital (15–20 year horizons), regulatory authority (new approval pathways from scratch), scale ($50–100B+), risk tolerance, and geopolitical ambition.

The only capital class combining all five is sovereign wealth.

Why the Gulf States: Regulatory blank slate — no FDA legacy or woke political contamination.

• Hevolution Foundation has committed $400M+ to aging biology.

• Vision 2030 prioritizes diversification.

• Existing medical tourism infrastructure.

• Cultural alignment with civilizational-scale projects.

“We will save humanity from aging” is more compelling than another financial center — and would compete with nobody.

China is a potential competitive threat: $450B+/year R&D, regulatory flexibility, overrepresentation in aging biology talent. But lower international trust and geopolitical constraints create a positioning opportunity: the Gulf as the trusted, rigorous, internationally collaborative alternative to both Western paralysis and Chinese opacity. The window is 5–10 years.

Part IV: The Three-Vertical Strategy

Vertical 1 — Age Reversal ($40–80B over 15 years)

The civilizational mission.

Every year LEV is accelerated is worth $10–50T in reduced mortality and healthcare savings.

A 10× funding increase could compress timelines by 10–15 years — a 500:1 return.

Structure: National Age Reversal Research Institute (campus-scale), 50+ parallel delivery approaches, 200+ funded research programs, self-driving labs multiplying throughput 10–100×, dedicated clinical trial infrastructure, and aggressive talent recruitment (top 500 aging researchers globally at 3–5× market compensation with $5M–$50M unrestricted lab budgets).

Vertical 2 — Bioenhancement ($15–30B over 10 years)

The commercial engine, dynamically allocated across modalities based on evolving expected value:

• Programmable Biology (~40–50%, $8–20B): EOS Genomics Engine ($1–2B) for 50,000+ genomes at trait extremes and permutational stack discovery. siRNA programs ($2–4B), circRNA platform ($2–4B), base/prime editing programs ($3–6B), target characterization and safety ($1–3B), and shared delivery R&D ($2–5B).

• AI Pharmacology (~15–20%, $3–8B): AI drug design platform, de novo enhancement compounds, clinical trials in healthy adults, and elite performance clinic network (20+ locations, $100K–$500K/year programs). Revenue from year 2–3.

• Embryo Engineering (~15%, $3–6B): Advanced polygenic screening center, IVG crash program, embryo editing R&D as somatic safety data matures, and longitudinal outcomes tracking.

• Shared Infrastructure (~10%, $1.5–3B): Performance measurement lab, longitudinal enhancement database, regulatory standards body, and enhancement medicine training — creating a medical specialty that doesn’t exist.

The Revenue Flywheel: Sovereign authority creates legitimacy → 10,000 early clients at $100K/year = $1B/year by year 2–3 → revenue funds R&D → better results attract more clients → AI + sovereign regulatory flexibility compress timelines → published results create exponential demand → by year 5–7, enhancement revenue ($5B–$15B/year) exceeds R&D spend. The iPhone model: target elites initially (e.g. Bryan Johnson, LeBron, etc.), leverage profits to scale.

The endgame is pharma displacement. If you can permanently knock out PCSK9 with a single base edit, why take a statin daily for 40 years? If you can reverse epigenetic age and eliminate the upstream cause of most chronic disease, why would anyone need thousands of downstream drugs treating consequences of aging? Age reversal + bioenhancement makes the existing $1.4T pharmaceutical market increasingly obsolete. The methods developed for enhancement produce actual cures faster and cheaper than the disease-focused model because they address mechanisms rather than symptoms.

The long-term vision is on-demand biological customization: tune cognition, body composition, mood, sleep, social confidence (and whatever specifics you want within each subcategory) — different profiles for different contexts via transient RNA, with permanent lock-in when safety data supports it. Permanent is always the goal when the individual wants it and the data is there. Dependency on transient interventions creates supply chain vulnerability. Transient RNA is the testing phase, not the endpoint.

Vertical 3 — Artificial Augmentation ($5–15B over 15 years)

The longer horizon.

• BCI Research Institute ($2–4B)

• Organ Engineering Lab ($1–3B)

• Bionics Center ($1–2B)

• Synthetic Biology Platform ($500M–2B)

• BCI-Biology Integration R&D ($500M–1B)

The same sovereign regulatory advantage enables enhancement BCIs in healthy adults — something no Western jurisdiction will approve for a decade. Enhancement clients from Vertical 2 are the natural first customers.

Cross-Vertical Synergies: Shared delivery platforms serve all three verticals. Safety data from each de-risks the others. Enhancement clients become augmentation customers. Age reversal provides moral legitimacy that benefits the enhancement brand. The IP portfolio across all three creates a moat no competitor can replicate.

Part V: The ROI Case

NEOM costs $500B+ and competes with Dubai. Index fund diversification competes with every institutional investor.

This program costs $50–100B and competes with nobody.

• $100B buys ~38 drugs through FDA approval in a field with 6,300 competing oncology programs.

• That same $100B buys foundational IP, infrastructure, and workforce for three entirely new global industries — 5–10× the total historical investment in age reversal, with increasing marginal returns.

• It targets the elimination of the upstream cause of every disease that healthcare spending currently manages, saves 36.5 million lives per year, and compounds over centuries.

The value at stake: $180T/year in mortality cost, $10–15T/year in age-related healthcare, $1.4T/year pharmaceutical market ripe for displacement. At $930B+ in PIF AUM, $100B over 10 years is ~10% of assets — aggressive by normal standards, rational at 100:1+ expected returns.

The “absorptive capacity” objection is circular. The Manhattan Project went from a handful of physicists to 125,000 employees in 3 years. Operation Warp Speed deployed $18B in under a year. For crash programs with clear objectives, absorptive capacity expands to meet committed capital.

Part VI: Risk and Mitigation

Technical (moderate, declining):

• Reprogramming may not translate from mice to humans (20–30%).

• Cancer risk from early trials (15–25%).

• Delivery technology may plateau (20–30%).

Address through diversification: multiple parallel approaches, tissue-specific sequencing, and the recognition that even conservative outcomes create massive markets.

Geopolitical (significant, manageable):

• US restrictions on talent/tech transfer (15–30%) — mitigated by global recruitment, money, and experimental freedom.

• China moving faster (25–35%) — countered by speed and trust advantage.

• International backlash on “enhancement” framing — neutralized by leading publicly with age reversal or just not giving in to woke psychology (like is done for crime).

Reputational (highest priority): The framing: the FDA’s risk calculus is woke and unethical.

• Every year of regulatory delay on a working therapy is 36.5 million deaths.

• The program must be rigorous, transparent, and participant-protective but should not be afraid to pay anyone to participate in trials.

• Deliver results, and criticism becomes what it always was: jealousy from those who are missing out.

Part VII: Implementation

1. Phase 1: Foundation (Years 1–3, $15–25B): Establish regulatory authority with approval pathways for all three verticals. Build campus-scale research institute, 5–10 self-driving labs, 10+ GMP gene therapy manufacturing facilities, AI compute cluster, elite performance clinic infrastructure. Recruit 200 senior researchers at 3–5× market. Launch 500 postdoctoral fellowships/year. Fund 200+ research programs (10× current global capacity). Launch EOS Genomics Engine immediately. Begin siRNA enhancement programs, AI pharmacology, embryo screening services, and BCI fundamental research.

2. Phase 2: Clinical Translation (Years 3–7, $25–40B): Phase 1 trials for organ-specific age reversal (eye, liver, skin, immune — simultaneously). First enhancement trials in healthy adults across siRNA, base editing, and AI-designed compounds. Elite clinic network generating revenue. IVG milestones. First-in-human cognitive BCI enhancement trials. Launch $5–10B longevity/enhancement startup incubator. Attract 50–100 international companies.

3. Phase 3: Commercialization (Years 7–15, $40–70B, partially self-funding): First approved age reversal therapies. Approved enhancement protocols across modalities. Medical tourism at scale. Enhancement revenue $5B+/year. Government/military contracts. Consumer-grade products. Embryo editing entering clinical use. By year 15, $30–60B+/year in revenue making continued investment self-sustaining.

Total: $80–135B gross, $50–95B net after revenue over 15 years.

Part VIII: Why Now?

Five trends converging:

1. Biological proof-of-concept is fresh (2024 mouse data, 2026 first human trial).

2. Enhancement tools crossed the clinical threshold simultaneously in 2024–2025 (base editing, prime editing, circRNA, AI-designed delivery).

3. AI is compressing timelines 2–10×.

4. Western institutions are paralyzed across every modality.

5. Public demand is exploding (embryo screening commercial, BCI investment >$1B, GLP-1 proving mass-market appetite for optimization).

Every year of delay costs ~36.5 million lives, $4.3T in US healthcare spending managing diseases rather than curing them, and competitive positioning that becomes exponentially harder to recover.

This failure creates the largest strategic vacancy in the history of biomedical investment. Three entire industries — age reversal, bioenhancement, and artificial augmentation — should exist but don’t.

What’s missing is a sovereign entity with the capital, regulatory authority, time horizon, and ambition to build what Western institutions won’t.

The competition: Nobody.

The cost of waiting: ~100,000 lives per day, a long-term Dark Age due to dysgenic momentum (fast-tracking Idiocracy), and a possible extinction event (as more low IQs gain power).

Who is flying hardware? What are they targeting? And does any version of the biz have a credible ROI path?

In early 2026, nobody has commercially mined an asteroid; no company has extracted, processed, and delivered sellable material from space at meaningful scale.

The reason is fairly simple: “asteroid mining” is two different businesses with totally different difficulty and economics:

1. Earth-return precious metals (PGMs): return high-value metal to Earth via reentry capsules. Engineering is brutal, and commodity markets fight you.

2. In-space water/volatiles: extract water and sell it in space (propellant, life support, shielding). Economics are cleaner because the benchmark is avoided launch/transport cost, not terrestrial mining.

TL;DR: If asteroid mining ever becomes a real business, it likely starts with water/volatiles sold in space, not platinum returned to Earth.

What Materials Are We Talking About?

Asteroids are not all the same.

Composition varies enormously, and the target material depends entirely on which business model a company is pursuing.

The popular press tends to say “platinum” as shorthand, but the actual commercial pitch is almost always about platinum-group metals (PGMs) as a basket, not just platinum alone.

And the most plausible early resource isn’t a metal at all.

1. PGMs (Earth-return value density)

• Product: PGMs (Pt/Pd/Rh/Ru/Ir/Os)

• Customer: Earth commodity buyers

• Why it can work: extreme value density per kilogram

• Main failure mode: refining + return reliability + market depth constraints

This is the “metallic asteroid” thesis. The target is the six platinum-group metals: platinum, palladium, rhodium, ruthenium, iridium, and osmium. These are the metals with enough value per kilogram to potentially justify the cost of a deep-space round trip.

AstroForge repeatedly frames its goal as extracting platinum-group metals, not just platinum. Their Vestri mission press materials describe the objective as evaluating asteroids for PGMs including platinum, palladium, rhodium, and iridium.

Their Odin mission debrief “Odidn’t” describes the goal as validating whether the target asteroid is “rich in platinum group metals.”

So no, it’s not “only platinum.” Platinum is just the headline metal people recognize. The target basket is all six PGMs.

2. Water / volatiles (in-space fuel and propellant)

• Product: water / volatiles

• Customer: depots, tugs, cislunar operations

• Why it can work: priced against avoided launch + transport cost

• Main failure mode: the market might not exist yet (no depots, no cadence)

This is the “don’t return commodities to Earth; sell in space” thesis.

The target is water, which can be used directly for life support and radiation shielding, or decomposed into hydrogen and oxygen for rocket propellant.

• Karman+ describes missions to mine regolith material and extract water for refueling in orbit

• Optical mining approaches (TransAstra/NASA NIAC lineage) are explicitly about extracting water and other volatiles from carbonaceous material using concentrated sunlight inside contained processing bags

Don’t asteroids contain many metals? (Gold, Cobalt, Nickel, Tungsten, etc.)

They can, but contents vary by asteroid.

Top targets (the only ones anyone is seriously pursuing):

1. Water (H₂O): The best early target, full stop. Worthless on Earth, enormously expensive to launch to orbit. Potentially simpler to extract from C-type asteroids than metals (if thermal/microwave extraction works reliably in microgravity). Value comes entirely from avoided launch cost. This is what Karman+ and TransAstra are building toward.

2. Rhodium (Rh): ~$334K/kg (~$10,400/oz as of Feb 2026). The most valuable PGM by far. Extremely rare on Earth, tiny market, but extraordinary value density per kilogram.

3. Iridium (Ir): ~$215K/kg (~$6,700/oz). Crucibles, spark plugs, electronics. Very dense, very rare.

4. Platinum (Pt): ~$67K/kg (~$2,075/oz). Catalysts, electronics, fuel cells. The headline metal in AstroForge’s pitch.

5. Palladium (Pd): ~$55K/kg (~$1,739/oz). Autocatalysts, electronics, hydrogen purification. Similar Earth-return case to platinum.

6. Ruthenium (Ru): ~$46K/kg (~$1,430/oz). Electronics, catalysts. Lower profile but still significant value density.

7. Osmium (Os): No clean industrial spot market comparable to the other PGMs. Specialty pricing exists for crystalline products, but it’s not traded like platinum or palladium. Would come along for the ride with other PGM extraction, not targeted independently.

Marginal (probably not worth the trip on its own):

8. Gold (Au): ~$95K/kg sounds great, but Earth’s gold supply is enormous (~3,600 tonnes mined per year). Asteroid gold would be a rounding error. It might show up as a byproduct of PGM extraction, but nobody is designing a mission around gold.

Impractical for Earth return (value density way too low):

• Iron (Fe): ~$0.75/kg on Earth. Could be useful for in-space construction someday, but that requires a large orbital industrial base that doesn’t exist.

• Nickel (Ni): ~$16/kg. Same story as iron. Only useful as structural feedstock in orbit.

• Cobalt (Co): ~$33/kg. Battery and superalloy demand exists on Earth, but value per kg is far too low to justify a deep-space round trip.

• Tungsten (W): ~$35/kg. Refractory uses, but not remotely worth a space mission.

These low-value metals are what make up the bulk of asteroid mass. They’re why headlines say “this asteroid is worth $10 trillion.”

But you’d never design a mission to return iron or nickel to Earth. If they ever matter commercially, it’ll be as raw feedstock for building things in orbit, decades from now.

Reality on materials:

• Metals realistically targeted: PGMs as a group (Pt/Pd/Rh/Ru/Ir/Os), not “platinum only.”

• Most plausible early resource: Water/volatiles for in-space use, not metals to Earth.

• Everything else (iron, nickel, cobalt, etc.): Only make sense for in-space construction once a large orbital industrial base exists, which is decades away at minimum.

Asteroid Types

Three asteroid classes matter for mining:

• C-type (carbonaceous): The most common (~75% of known asteroids). Rich in water, bound up in hydrated clay minerals. Also contain organic compounds, carbon, and some phosphorus. These are the targets for water/propellant extraction.

• S-type (stony/silicaceous): Contain significant metal fractions, mostly iron, nickel, and cobalt, with trace amounts of precious metals. According to NASA, a small 10-meter S-type asteroid contains about 650,000 kg of metal, including roughly 50 kg of rare metals like platinum and gold.

• M-type (metallic): Rare but contain up to 10x more metal than S-types. Essentially chunks of planetary cores, dominated by iron-nickel alloy with commercially significant concentrations of PGMs.

Platinum-rich asteroids may contain grades of up to 100 grams per ton, which is 10-20x higher than open-pit platinum mines in South Africa.

A single 500-meter platinum-rich asteroid could theoretically contain nearly 175 times the annual global platinum output.

But “contains” and “can profitably extract and deliver” are very different statements.

Asteroid Mining Companies in 2026

1. AstroForge

• Target: Platinum-group metals (platinum, palladium, rhodium, iridium) from metallic near-Earth asteroids

• Founded: January 2022 by Matthew Gialich (CEO) and Jose Acain (CTO), in Huntington Beach, California

• Team size: ~35-42 employees as of 2025

• Total funding: $55-56M across 4 rounds

• Key investors: Nova Threshold (led Series A), Seven Seven Six (Alexis Ohanian’s fund), Initialized Capital, Y Combinator, Caladan Capital, Uncorrelated Ventures, plus angel investor Jed McCaleb (billionaire founder of Vast)

• Equipment and propulsion: For their upcoming Vestri mission, Safran Defense & Space will provide two EPS X00 electric propulsion systems (each comprising a PPS X00 thruster, power processing unit, and fluid management system). AstroForge now builds its spacecraft in-house after early outsourcing to OrbAstro failed.

Mission history:

1. Brokkr-1 (April 2023): A 6U cubesat launched on SpaceX Falcon 9 Transporter-7. Purpose: demonstrate metal refinery tech in orbit. The magnetic field from the refining system interfered with active attitude control, degrading pointing and communications. Solar arrays eventually deployed (September 2023), and telemetry could be downlinked, but AstroForge was never able to close the command uplink needed to activate the refinery payload. Last contact was May 2024.

2. Odin (Feb 2025): A 100 kg spacecraft launched as a rideshare on Intuitive Machines’ IM-2 mission via Falcon 9. Goal: flyby of near-Earth asteroid 2022 OB5. Built in-house in 212 days after the original OrbAstro-built version failed vibration testing. Cost: less than $7M. Declared lost on March 6, 2025 after failure to establish sustained communications. Initial comms were hampered by ground-station issues; ultimate loss could involve attitude, power, or deployment problems on the spacecraft side. Root cause was not definitively established.

3. Vestri (target window: 2026, possibly slipping to early 2027): A 200 kg spacecraft intended to travel to the same target asteroid and directly characterize its composition. Uses Safran electric propulsion. Planned to launch with Intuitive Machines’ IM-3 mission. Public sources vary on exact timeline and mission profile (some describe docking, others describe landing legs); treat the launch window and approach as provisional until a final mission statement is published. If it reaches the asteroid, it would be the first private spacecraft to rendezvous with a body outside the Earth-Moon system.

CEO Matt Gialich’s approach: “We’re going to mine asteroids or go bankrupt, and we’re going to probably figure that out in the next five years.”

Progress: Further along than anyone else in terms of actual flight hardware in deep space. But 2 missions have failed, and they haven’t demonstrated mining, refining, or return of any material. Still firmly in the prospecting phase.

2. Karman+

• Target: Water from carbonaceous near-Earth asteroids, for in-orbit refueling

• Founded: 2022 by Teun van den Dries (CEO, serial entrepreneur) and Daynan Crull (Mission Architect, technology strategist), headquartered in Denver, Colorado with roots in the Netherlands

• Team size: ~29 employees

• Total funding: $20M seed round (PitchBook lists $46.5M total)

• Key investors: Plural (London), Hummingbird Ventures (Antwerp), HCVC (Paris), Kevin Mahaffey (Lookout founder), Climate Capital, Norrsken VC

Technical approach:

• “COTS+” strategy: Adapting commercial off-the-shelf components from other industries, testing and qualifying them for space. Goal is to keep mission budgets under $10M per mission (vs. $1B+ for NASA asteroid missions).

• Autonomous navigation: Developing optical navigation (”beacon nav”) for interplanetary cruise

• Mining hardware: Building zero-gravity mining equipment for regolith excavation at kilogram scale

• Extraction method: Excavate regolith, extract water, which can be used directly or decomposed into hydrogen and oxygen for propellant

First mission: “High Frontier” technology demonstration, targeting 2027 launch. Will attempt to rendezvous with a near-Earth asteroid and excavate regolith at kilogram scale.

Status (Q3 2025): Parts and components ordered; going through rapid design iterations with full-scale engineering model builds.

Market thesis: In-space refueling could be worth single-digit billions of dollars per year. Lowering satellite refueling costs by up to 10x compared to Earth-launched fuel. Later-stage plans include extracting rare metals and contributing to in-space manufacturing.

Progress: Earlier-stage than AstroForge; no flight heritage yet. But pursuing the arguably more plausible first market. Their cost targets are aggressive.

3. TransAstra

• Target: Water and volatiles from asteroids using concentrated sunlight

• Founded: 2015 by Dr. Joel Sercel (CEO), based in Los Angeles. Sercel spent 14 years at NASA’s Jet Propulsion Laboratory, taught at Caltech for 12 years, led conception of the NSTAR ion propulsion system (used on the Dawn spacecraft), and was the founding CTO of Momentus. He holds 14 patents pending in space resources technology.

• Team: Partnered with Prof. Robert Jedicke (University of Hawaii) and researchers at the Colorado School of Mines Center for Space Resources (Angel Abbud-Madrid, Chris Dreyer)

• Funding: Combination of private investment plus NASA grants and contracts. Six NASA NIAC fellowships (including one of the first two NIAC Phase 3 awards ever selected). Also won an $850K SBIR Phase 2 contract for inflatable capture bag development.

Core technologies:

• Optical Mining™ (patented): Uses concentrated sunlight via deployable mirrors to heat asteroid material inside an inflatable containment bag, fracturing rock and driving off volatiles, which are captured and condensed. Tested at White Sands Missile Range solar furnace on actual meteorite samples.

• Omnivore™ engine: A solar thermal rocket that can run on water, ammonia, hydrogen, or other propellants, designed for fuel-flexible in-space operations.

• Sutter™ telescope technology: Synthetic tracking system claimed to be 300,000x more effective at finding dark moving objects than conventional telescopes. Won a DoD grant for space situational awareness applications.

• Worker Bee space tug: In development for satellite repositioning using solar thermal propulsion. TransAstra has contracts and studies for the concept, but it is not yet flying routine commercial missions.

Architecture (Apis): The full concept envisions harvesting up to 100 metric tons of water from a single near-Earth asteroid and delivering it to lunar orbit or other depot locations, all from a single Falcon 9 launch.

Progress: The deepest R&D portfolio and the most experienced founder of the three companies. But further from commercial execution than AstroForge. TransAstra is pursuing near-term revenue from space tug and debris capture contracts while developing the longer-term mining architecture. Think of them as building the enabling infrastructure rather than going straight for the ore.

The Dead Companies

The two biggest names from the 2010s asteroid mining hype cycle are gone:

• Planetary Resources (founded 2010, backed by Larry Page, Eric Schmidt, James Cameron): Acquired by blockchain company ConsenSys in 2018. Mining ambitions abandoned.

• Deep Space Industries (founded 2013): Acquired by Bradford Space in 2019. Pivoted entirely to propulsion and spacecraft technology.

Is SpaceX or Blue Origin trying to mine asteroids?

SpaceX: No. No public asteroid mining program. SpaceX is an enabler and launch provider. AstroForge’s missions flew on Falcon 9. Their strategic focus is settlement infrastructure. SpaceX is prioritizing building a “self-growing city” on the Moon. Running a mining and commodities business is a completely different priority stack.

If a mature in-space economy develops where propellant is the bottleneck, SpaceX might get involved, but they’d more likely buy propellant from a specialist or focus on lunar/Martian ISRU rather than chasing asteroids.

Blue Origin: Not really. Blue Origin discusses building “space mining robots and systems” in the context of ISRU capability development. That’s building enabling technologies, not pursuing asteroid mining as a revenue line.

Who will do it first? Most likely a specialist startup with a focused architecture and potentially a government anchor customer. The big launch companies fit as enablers, not miners.

How close is any company to mining asteroids?

A milestone ladder for readiness:

1. Reach the asteroid reliably

2. Rendezvous or flyby imaging (prospecting)

3. Touchdown, docking, or anchoring

4. Excavate material in microgravity

5. Process/refine (separate metal, extract water)

6. Deliver product (to Earth or an in-space customer)

7. Repeat economically

Private companies are mostly between steps 1 and 3, still struggling to prove parts of 4 through 6.

For context on how hard “bring stuff back” is even for governments with billion-dollar budgets:

Missions:

NASA OSIRIS-REx (Sept 2023)

• Asteroid: Bennu

• Returned: 121.6 grams

• Cost: ~$1.16B

JAXA Hayabusa2 (2019)

• Asteroid: Ryugu

• Returned: 5.4 grams

• Cost: ~$150M

Those are scientific sample returns, not mining, but they illustrate the gap between “visit an asteroid” and “move industrial quantities of material.”

How many asteroids are close enough to Earth to mine?

“Close enough” isn’t about how near an asteroid passes Earth.

You need to think about whether you can match its orbit with a practical amount of delta-v (the velocity change your spacecraft needs) within reasonable mission timelines.

• ~39,123 total discovered near-Earth asteroids as of December 2025. Most aren’t economically reachable (high delta-v, bad launch windows, too small, spinning too fast, uncertain composition).

• ~6,294 are “NHATS-accessible” per NASA/JPL’s Near-Earth Object Human Space Flight Accessible Targets Study, meaning they have at least one viable round-trip trajectory. About 16% of known NEAs.

The “really easy” targets are rare. A 2015 NASA paper found only 580 NEAs reachable for less delta-v than a lunar surface round-trip (~9 km/s), and just 49 for less than a low lunar orbit round-trip (~5 km/s).

Many good targets haven’t been found yet. NASA estimates ~14,000 asteroids of 140m+ remain undiscovered. NEO Surveyor, launching no earlier than September 2027, is modeled to discover on the order of 200,000-300,000 new NEOs based on population models and survey simulations, though actual yield will depend on mission performance and survey duration.

Net: Tens of thousands of NEAs exist. Thousands are theoretically accessible. Dozens to low hundreds are plausibly economically accessible with good launch windows and workable physical properties, and many of those may still be undiscovered.

How will companies mine asteroids?

Step 1: Finding Targets

Companies start from catalogs of known near-Earth objects:

• NASA’s NEO Observations Program funds discovery and orbit refinement

• Ground-based optical surveys like the Catalina Sky Survey do most of the initial detection

• JPL’s CNEOS turns detections into tracked objects with refined orbits

• Infrared surveys catch dark asteroids invisible in visible light. NEOWISE (decommissioned August 2024) was a key survey; NEO Surveyor is its planned successor

From Earth, spectroscopy and albedo measurements give rough composition estimates. But the data is imperfect. That’s why prospecting missions (flybys or rendezvous) come first: you need to confirm composition and mechanical properties before committing hundreds of millions to a mining mission.

Step 2: Getting There

This is non-trivial for several reasons:

• Targets are tiny and irregular, often spinning or tumbling

• Gravity is essentially zero. Touching the surface can bounce you away (OSIRIS-REx experienced this)

• Navigation is autonomous. Communication delays grow with distance, so the spacecraft must make its own decisions using optical guidance

• Launch windows are constrained. You need the right orbital alignment, and windows may only open every few years for a given target

Step 3: Anchoring or Containment

Two broad strategies:

1. Anchor to the surface: Drills, harpoons, anchors, or nets to grip the asteroid. Difficult because surface properties are unknown until arrival, and there’s no gravity to help hold anything down.

2. Enclose the body: Wrap the asteroid in an inflatable bag and work inside containment. This is TransAstra’s approach. NASA studied a related concept in the (canceled) Asteroid Redirect Robotic Mission: capture a multi-ton boulder and redirect it to lunar orbit.

Step 4: Extraction

Water/volatiles (the nearer-term product):

• Heat regolith in a sealed container, capture released vapor, condense it

• TransAstra’s optical mining uses concentrated sunlight via deployable mirrors inside a containment bag

• Karman+ is developing purpose-built zero-gravity excavation equipment and plans solar-thermal or microwave heating methods

Metals (harder, longer-term):

• Shipping raw ore means paying to transport a lot of waste mass

• Practical approaches require at least partial beneficiation in space: magnetic separation, melting, electro-refining

• Each processing step adds mass, power requirements, thermal control complexity, and failure modes

• The more you refine in space, the more your spacecraft becomes an industrial plant, which is a core tension in mission design

Step 5: Delivery

Option A, deliver in space (water to a depot orbit): No reentry system required. Serve a space customer. Electric propulsion tug hauls product to the most valuable orbital location. This is the cleaner near-term path, if customers exist.

Option B, return to Earth (metals): You don’t return the whole asteroid (planetary defense nightmare, regulatory impossibility). You return small, controlled capsules. The most sensible architecture:

1. Process in space to concentrate value and dump waste (tailings)

2. Return only high-value mass (ingots or concentrate)

3. Finish refining on Earth where industrial infrastructure is mature

What’s the potential payday?

Depends if precious metals or water/propellant, the amounts you extract, and the specifics (i.e. amounts/composition of precious metals).

Path A: Returning Precious Metals to Earth

Using platinum spot pricing from Feb 10, 2026 (~$2,083/troy oz), and palladium (~$56K/kg):

These numbers are misleading without context:

• Values assume refined metal, not raw regolith. Extraction yields will be far less than 100%.

• A lean deep-space mining mission might cost $200M all-in. Break-even platinum mass: ~3,000 kg (3 tonnes). That’s an enormous quantity of refined product to return, especially early in the learning curve.

• The numbers ignore extraction yield losses, microgravity refining complexity, return capsule and recovery costs, and insurance against total mission loss.

Market impact: Global platinum mine supply was ~5,766 koz in 2024 (~179 metric tons from mining alone; total supply including recycling was ~7,293 koz or ~227 tonnes).

• 1 ton: ~0.6% of annual supply

• 10 tons: ~5.6% of annual supply

• 100 tons: ~56% of annual supply

So if a mining company were to extract and return 100 tons of platinum, they’d completely tank the price of platinum on Earth unless they slow-rolled sales over a period of many years (abundant supply = lower value; scarce supply = higher value).

The “this asteroid is worth $10 trillion” headlines are technically true and practically meaningless. Value does not equal realizable profit.

Path B: Selling Water/Propellant in Space

The payday here is tied to avoided launch cost, not Earth commodity prices.

One NASA reference benchmarks Falcon 9 delivery at ~$2,940/kg to LEO. If you can deliver asteroid water to a useful orbit for less, you have a business.

For harder-to-reach destinations (cislunar space, high-energy orbits, lunar vicinity), Earth-launched delivery costs are much higher, and the value of locally sourced propellant increases proportionally.

Early commercially meaningful deliveries would likely be hundreds of kg to a few tonnes. At $1,000-$5,000/kg delivered (location-dependent), that’s $0.1M-$25M revenue per delivery. You need repeatability and scaling for this to become a real business.

The catch: The entire business case depends on a real in-space refueling market materializing. That requires sustained cislunar activity (multiple lunar missions per year, orbital depots, tugs, standards for propellant transfer). Without demand, even cheap asteroid water has no ROI.

Does the ROI math work? (Biz model sustainability)

Significant ROI = repeatable commercial operations with large cumulative profit (order of $1B+) or equity-style returns that clearly clear “cool demo” status.

Short answer: Yes, a business can plausibly make significant ROI from asteroid mining, but most likely in the “sell water/propellant in space” lane, and not near-term. The classic “bring back platinum and get rich” lane is much less likely to work.

Economics of Earth-Return Metals

The cost stack for a “metals-return mission” includes: spacecraft development, launch, deep-space operations, autonomous mining hardware, in-space refining equipment, return capsule development, Earth reentry and recovery, ground-based final refining, and insurance.

1. With current technology and cost structures, breaking even requires returning multi-ton quantities of concentrated PGMs per mission; no one has demonstrated returning even grams commercially.

2. Even if you solve the engineering, you hit the market problem: large new supply pushes prices down.

WPIC reports total global platinum supply (mine + recycling) of ~7.293 million oz in 2024 (about 227 tonnes). If you ever return “real” tonnage, you start pushing the price against yourself unless throttled carefully.

And reliability remains a fundamental problem. Even government sample return is still tiny and expensive (OSIRIS-REx returned 121.6 grams after a $1.16B program).

Private deep-space startups still lose missions to basic ops risk: AstroForge’s Odin ran into severe comms/ops issues, exactly the kind of thing that destroys the economics until reliability becomes routine.

Refining (i.e. beneficiation) in microgravity adds mass, power, thermal management, autonomy requirements, and failure modes. If you don’t refine in space, you pay to return waste.

When could it work? If (1) mission costs fall dramatically (possibly through reusable deep-space architecture and learning-curve effects) + (2) a company can reliably return multi-ton quantities + (3) they meter supply to avoid crashing the market. Metals-to-Earth could happen as small “boutique returns” (kg-scale) for a long time, but that’s not significant ROI.

Economics of In-Space Water

Plausible, but conditional.

Water/propellant is valuable because it avoids Earth launch and transport.

You aren’t competing with terrestrial mining at all. You just need to beat the delivered cost to the customer’s orbit. That’s a different and more favorable competitive dynamic.

The unit economics are potentially favorable: water is enormously expensive to launch from Earth to useful orbits, and carbonaceous asteroids contain it in abundance.

There are thousands of “accessible” NEAs in mission-design terms, so targets exist; the difficulty is characterization, extraction, and reliable operations.

If you can build a reusable mining tug that visits low-delta-v asteroids and delivers water to orbital depots repeatedly, the per-kg cost could undercut Earth-launched alternatives, especially for harder-to-reach destinations (cislunar space, high-energy orbits) where launch costs multiply.

What has to go right:

1. The cislunar economy has to materialize (multiple lunar missions per year, commercial stations, depots)

2. Orbital refueling infrastructure and standards have to be developed

3. Someone has to prove the extraction technology works reliably in microgravity

4. The mining tug architecture has to be reusable enough to amortize development costs

The gating factor is demand. If there isn’t a sustained cislunar logistics market (depots, tugs, frequent lunar ops), asteroid water has no buyer.

Karman+ is betting on this path, with mission budgets targeted under $10M and a COTS+ approach to keep costs low, aiming at customer missions expected in 2027. If the refueling market reaches even low single-digit billions per year (their target), there’s room for positive unit economics. But the market doesn’t exist yet.

Will anyone generate high ROI? (GPT-5.2 Prediction)

I had ChatGPT’s 5.2-Pro/High speculate on odds. I would likely assign higher odds (with equally low confidence) given technological progress in space technology + AI + robotics, but maybe I’m slightly over-optimistic.

1. In-Space Water/Propellant Profitable

• 2035: ~5% odds (low confidence)

• 2045: ~15% odds (low-medium confidence)

• 2055: ~25% odds (low confidence)

Rationale (the most plausible path to significant ROI):

• Water/propellant value comes from avoiding Earth launch + transport, not from competing with terrestrial commodity markets

• Multiple teams are explicitly building toward water extraction demos

• Thousands of accessible NEAs exist as targets

• The gating factor is demand, not physics

2. Earth-Return Metals (PGMs) Profitable

• 2045: ~3% odds (low confidence)

• 2055: ~7% odds (low confidence)

• 2070: ~10% odds (very low confidence)

Why such low odds?

• Government sample return is still tiny and expensive after billions spent

• Private deep-space ops are still failure-prone at the most basic level (comms, navigation, power)

• Market depth is a real constraint (~227 tonnes of total global platinum supply per year including recycling; ~179 tonnes from mining alone)

• In-space refining adds enormous complexity; skipping it means paying to return waste

• Even if technically successful, large returns push prices against you

3. Technical Mining Demo (extract 10-100+ kg)

• Timeline: 2030-2040

• Probability: ~60%

• Confidence: Medium

• Rationale: Core ingredients exist in pieces (rendezvous, sampling, return capsules, propulsion, autonomy). Can be government-supported or subsidized. Growing catalog of accessible targets and falling launch costs help.

The Realistic ROI Playbook

If significant ROI ever happens, it almost certainly follows this sequence:

1. Prospecting: Cheap rendezvous + close imaging + basic composition checks, done repeatedly across multiple targets

2. Containment/anchoring that works on small, spinning bodies in microgravity

3. Simple extraction first (water/volatiles), not complex metallurgy

4. Sell in-space to a depot or orbital customer, not back to Earth

5. Scale via repetition (many identical tugs) rather than one “hero mission”

This gets you away from Earth commodity markets entirely and into a logistics business. The company that succeeds looks more like a propellant delivery service than a mining company.

What would increase the odds?

• A clear signal that the in-space refueling market is real and growing (multiple depots, routine tanker demand, NASA or commercial anchor customers)

• Demonstrated “extract + store + deliver” at meaningful scale (even hundreds of kg), done more than once

• Improved discovery and characterization pipeline (NEO Surveyor is scheduled no earlier than Sept 2027; while it’s planetary defense oriented, better catalogs help everyone)

• Falling mission costs combined with rising spacecraft autonomy and reliability

• Starship or similar next-gen heavy lift making deep-space missions dramatically cheaper to launch

Implications of asteroid mining success

If any commercial Earth-return happens in the next couple decades, expect a progression:

1. Grams to kilograms: Proof of capability. Negligible market effect.

2. Tens of kilograms: Still negligible market effect.

3. 1 tonne of platinum (~$67M at current pricing): ~0.6% of annual mine supply. Probably not a market shock.

4. 10 tonnes: ~5.6% of annual supply. Could start moving price, especially if dumped quickly.

5. 100 tonnes: ~56% of annual supply. Almost certainly craters the price unless rationed across years.

The smarter play is likely selling in small, controlled batches, or targeting specialty applications where small quantities of ultra-pure space-refined PGMs command premium pricing… unless it’s so cost-efficient to mine that you just want to offload and keep accelerating.

Milestones to Track

If you want to gauge whether asteroid mining is getting real:

1. Repeated successful rendezvous + sustained comms with tiny targets (startup-grade spacecraft)

2. Demonstrated anchoring or containment that survives weeks to months

3. Demonstrated extraction of a measurable quantity (even grams to kg)

4. Delivery to a customer location (depot orbit) or Earth return of a meaningful payload

5. A second flight that repeats. This would be where economics start to exist.

Right now, no private company has cleared milestone 2. AstroForge’s Vestri mission (targeting 2026-2027) is attempting to reach milestone 2-3. Karman+’s High Frontier (2027) targets milestone 3.

Why most companies will quit asteroid mining (even with better tech)

“Technology improves over time” is true but insufficient. It doesn’t imply the same companies survive long enough to cash in, and it doesn’t imply asteroid mining becomes the best way to get resources.

A more accurate rule: Tech advances make it feasible (increases the chance someone does it), advances don’t guarantee profits (most early entrants still fail), and advances change the game (sometimes the winning play becomes lunar ISRU, recycling, or in-space manufacturing rather than asteroid-to-Earth commodities).

• Time-to-revenue mismatch. Deep-space hardware cycles are long, failures are common, and “mining” requires multiple stacked breakthroughs (rendezvous, anchor, extract, store, deliver, repeat). Venture funding usually wants credible revenue paths inside 5-10 years. Many teams run out of runway before the inflection point arrives.

• Reliability is the real wall. A single comms/nav/thermal failure can zero out an entire mission. Until reliability becomes “boring,” the expected value stays ugly. AstroForge’s Odin debrief is exactly the kind of mundane operational failure (comms, ground stations, timing) that wipes out a whole mission’s economics.

• Demand might not show up. The best near-term product is in-space consumables (water/propellant/shielding mass). If the cislunar economy doesn’t scale, there’s no market regardless of how good your mining tech is.

• Even if it becomes possible, it may not be profitable. “More resources exist” doesn’t mean a company earns outsized profits. If supply becomes easier, prices drop, margins compress, and the value accrues to downstream users, not the miner.

• Pivots are rational. Teams often spin the tech into nearer markets: propulsion, autonomy, robotics, satellite servicing, lunar ISRU, defense contracts. That’s not “giving up.” It’s harvesting the only bankable portion of the stack. This is exactly what happened to Deep Space Industries (pivoted to propulsion, acquired by Bradford Space) and Planetary Resources (acquired by ConsenSys after funding trouble).

Will the current crop of mining companies survive?

An estimate by ChatGPT.

Base-rate odds for the current asteroid mining startup cohort by 2035.

• ~60-70%: Most “asteroid mining” startups either exit or pivot into adjacent businesses (propulsion, robotics, lunar ISRU, software, defense contracts)

• ~20-30%: One or two remain active but are still largely in demo/prospecting mode (not large profits)

• ~10%: At least one is doing repeatable commercial deliveries (most likely water/volatiles to an in-space customer)

Confidence: Medium on “lots of exits/pivots,” low on exact split.

Why are companies even trying with such low odds?

1. Asymmetric bet: Low probability, very large payoff, plus a bunch of partial wins that still monetize.

2. Power-law payoff. If you’re first to reliable extraction and delivery, you’re not just selling “ore.” You could control a logistics chokepoint for a future space economy.

3. You don’t need the full endgame to win. Even if “asteroid mining” fails, they can still create valuable spinouts: rendezvous/nav/autonomy stacks, anchoring/containment hardware, thermal extraction systems, in-space material handling. Those can sell into lunar, defense, satellite servicing, and tug markets.

4. Marketing and fundraising dynamics. “Mining asteroids” is an easy narrative to raise capital on compared to “we’re building an anchoring/containment system for small bodies.”

5. They’re betting the demand side turns on. If depots/tugs/lunar cadence ramps up, in-space consumables can become strategically valuable. Karman+ is making that demand bet explicitly.

6. Learning curve and iteration. The first few missions are expected to fail. They’re trying to survive long enough to get to “boring reliability,” which is when economics can start to exist.

The Wave Pattern

What you get is waves, not a single linear progression:

1. Wave forms (startups pitch asteroid mining, raise capital)

2. Wave dies (most fold or pivot because near-term ROI isn’t there)

3. Costs fall, demand rises, tech improves

4. New wave forms (often with the same pitch, slightly better enabling conditions)

You’ve already seen this once with Planetary Resources and Deep Space Industries. The current cohort (AstroForge, Karman+, TransAstra) is wave two.

Most will likely exit or pivot. The category will persist because option value, tech spillover, and the expanding target catalog keep drawing new entrants.

Does progress on Earth make asteroid mining unnecessary?

It can, especially for metals-to-Earth; Earth will also get better over the same timeframe.

• Automation and electrification lower mining operating costs.

• Recycling and better separation techniques improve effective supply.

• Substitution changes demand (batteries and catalysts can shift away from specific scarce inputs).

• Politics and environmental regulation can push costs up or drive investment into cleaner extraction methods.

The net effect is uncertain, but it’s not one-directional.

The competition isn’t “2026 Earth mining vs. 2050 asteroid mining” — it’s 2050 Earth mining vs. 2050 space mining.

If humanity is capable of industrial asteroid mining, it’s also likely capable of very efficient terrestrial mining, recycling, and substitution, and that can keep commodity prices low enough to kill asteroid-to-Earth margins.

Where asteroid mining still makes sense as both sides improve:

• When the customer is already in space.

• Water/propellant/shielding mass for cislunar logistics and feedstock for in-space manufacturing (once that exists) don’t compete with a mine in South Africa.

They compete with the delivered cost of launching mass from Earth to where it’s needed; a completely different competitive dynamic that improves with scale rather than degrading.

Will robots do the asteroid mining?

Yes, but probably not the robots you’re imagining.

Robots will almost certainly do most of it, but “humanoid miners” (biped, human-shaped robots walking around asteroids) may not be the obvious endgame.

• Humanoid form factors are wrong for asteroids: Asteroids are the worst possible environment for a biped. Near-zero gravity plus rubble surfaces means walking is basically meaningless; you need anchors, tethers, and winches, not legs. Regolith is abrasive and electrostatically clingy, making humanoid joints and hands failure magnets. Radiation, thermal swings, and long-duration autonomy requirements favor simple, hardened mechanisms with few moving parts.

• What the actual mining hardware looks like: Anchored multi-arm manipulators, tethered excavators/augers, bag/containment systems, thermal and microwave processing units, and small swarm inspectors for mapping and debris monitoring. The architecture is industrial, not anthropomorphic.

• Where humanoids make sense: Maintenance and repair at a cislunar depot or processing hub, where the environment is designed around human-compatible interfaces (handrails, connectors, valves, standardized tool mounts). Handling unexpected failures where general dexterity matters. Teleoperation from nearby (lunar orbit, Gateway, depot) with low latency. In short: humanoids as the mechanics, not the mining rig. The mining hardware is specialized; the humanoid is the flexible fixer.

• What robots don’t solve: Even perfect robots don’t eliminate the hard parts: delta-v and orbital logistics, power/thermal management, containment in microgravity, and the need for months-to-years of reliability without human rescue. Robots reduce labor cost and increase uptime, but the economics still hinge on delivered cost per kg and whether demand exists.

• Most likely future architecture: Autonomous tug reaches the asteroid. Wrap, anchor, contain. Deploy purpose-built extraction and processing modules. Store product. Deliver to depot or return capsule. Dexterous robots (possibly humanoid-ish) service the depot and fix modules when things break.

When does “routine robotic asteroid mining” happen?

GPT-5.2-Pro gives its odds here as a forecast.

Humanoid robots as depot/station maintenance techs

• 2050: ~10% odds

• 2075: ~25% odds

• 2100: ~60% odds

Asteroid water/volatiles extraction as routine industrial activity

• 2050: ~5-10% odds

• 2075: ~20-30% odds

• 2100: ~35-55% odds

Metals-to-Earth as common business

• 2050: <5% odds

• 2075: ~5-10% odds

• 2100: ~5-15% odds

Humanoid depot maintenance comes first because servicing infrastructure in a controlled environment is where dexterity pays off. Asteroid extraction comes later because you need cheap/reliable deep-space ops, robust autonomy, and sustained demand. Metals-to-Earth may never become “common” because commodity markets, substitution, recycling, and price collapse dynamics fight you at every step.

Using the Moon as a Processing Hub?

A common question:

Why not capture an asteroid, bring it to the Moon, process it there with gravity and infrastructure, and ship refined product to Earth?

Version 1: Park a small asteroid or boulder in cislunar orbit and process it there.

This is the version NASA seriously studied. The Asteroid Redirect Mission (ARM) concept would capture a multi-ton boulder and move it into a stable lunar distant retrograde orbit (DRO) for repeated visits and processing.

Academic work has also shown a small subset of very tiny NEOs could be “retrieved” into Earth’s neighborhood with less than 500 m/s of delta-v using low-energy transfers.

• Pros: Much shorter comm delays and easier ops than deep space, you can reuse infrastructure (tugs, power, processing rigs), you can fail/repair/iterate more like an industrial program.

• Cons: You’re still in microgravity (so “mining like on Earth” isn’t solved), you import planetary defense and liability concerns (deliberately altering orbits), and it only works for small bodies.

Version 2: Land asteroid material on the lunar surface, process it on the Moon, then ship product to Earth.

This is physically possible but usually the wrong economic move. Landing is expensive (the Moon has no atmosphere, so you must spend propellant to land safely, unlike Earth where you get “free braking” via atmospheric reentry).

You’re hauling waste mass (if you bring raw ore/regolith to the Moon, you’re paying to move mostly not-the-valuable-stuff). You still need a return system to Earth (capsules, heat shields, recovery).

And if you already have lunar industrial infrastructure (power, robotics, processing plants, propellant), then mining the Moon itself (ice, oxygen, metals via something like Blue Origin’s Blue Alchemist) is often a nearer, simpler supply chain than importing asteroid feedstock.

Which version might work?

Plausible “Moon helps” pathway: Identify a small, low-delta-v near-Earth object, use a solar-electric tug to nudge it into a stable cislunar orbit (DRO or Lagrange vicinity), process it in orbit with reusable equipment, and return only concentrated product to Earth in small capsules. That’s a “cislunar processing hub,” not “drag a mountain to the Moon.”

Overall: Landing and hauling bulk mass tends to wreck economics unless you already have a mature lunar industrial base and you’re only transporting high-value concentrates. The Moon doesn’t automatically make asteroid mining easier; it mostly adds delta-v costs unless your architecture is specifically designed around cislunar orbit processing.

Long-Term Outlook: Asteroid Mining in 2100

The long-range outlook is path-dependent on whether a large cislunar economy actually forms. There’s no single “status quo” for the end of the century. The right way to frame it is as a scenario set.

Scenario A: Modest Space Economy (science + defense + limited industry)

• Probability by 2100: ~35%

• Space activity is big compared to today, but still not “industrial”

• Resource use is mostly lunar ice/regolith ISRU for local ops; asteroid work is occasional and strategic

• “Asteroid mining” exists mostly as government/strategic capability and niche commercial demos

• Earth commodity markets are basically unaffected (no meaningful Earth-return flows)

Scenario B: Large Cislunar Economy (depots, tugs, habitats, lunar industry)

• Probability by 2100: ~50%

• The main space “mining” businesses are consumables and construction mass: water (life support, shielding, propellant feedstock), oxygen (oxidizer, life support), bulk mass for shielding/structures

• Asteroids matter, but mainly as feedstock for in-space use: water-rich bodies for volatiles, metallic bodies for in-space metallurgy once automation is mature

• Metals-to-Earth is still a niche, not the core. If any Earth return happens, it’s throttled, high-value, regulated, and small relative to global markets

• The winners look like “space logistics suppliers,” not classic miners: a few operators with fleets of autonomous tugs and processing nodes near stable cislunar orbits, selling delivered mass in the right orbit

Scenario C: Inner Solar System Industrialization (very large-scale automation)

• Probability by 2100: ~15%

• Most heavy industry that can be moved off-Earth is moved off-Earth

• Asteroid mining is routine and huge in tonnage, but almost all value is in-space manufacturing, not Earth commodity arbitrage

• Earth becomes a “high-value surface” (data/biotech/culture/residential) while bulk materials and energy-intensive manufacturing happen in space

• Earth markets for PGMs/metals are either mostly irrelevant (substitution/recycling/post-scarcity dynamics) or heavily managed to avoid destabilizing terrestrial economics

ROI Probabilities by 2100

• No significant ROI; only demos + small gov programs: ~35% (low-medium)

• Significant ROI from in-space consumables (water, oxygen, shielding mass): ~45% (low)

• Significant ROI from metals-to-Earth (PGMs): ~10% (very low)

By 2150, the odds rise mainly because by then the question of whether space industry exists at scale has mostly been resolved one way or another.

In-space asteroid resources reach ~60-70% probability of significant ROI; metals-to-Earth reaches ~15-25% (very low confidence on both).

The big uncertainty across all timeframes is not “can we physically do it.” It’s whether the economic substrate (customers in space, infrastructure, legal/regulatory stability) actually matures.

Bottom Line: Asteroid Mining (as of 2026)

Significant ROI before ~2040: Unlikely.

Significant ROI by mid-century: Possible, but mostly via in-space water/propellant, not Earth-return metals.

Metals-to-Earth as a major profit engine: Long-shot. Could happen as small boutique returns for a long time, but $1B+ cumulative profit from PGM return is a ~3-10% probability through 2055. Earth-side mining, recycling, and substitution will also improve over the same period, keeping commodity prices low enough to squeeze asteroid-to-Earth margins.

Water/propellant for in-space use: The only version of asteroid mining with a plausible near-term path to significant ROI. The economics are cleaner, the engineering stack is shorter, and you’re selling something priced against avoided launch cost rather than competing with terrestrial mines. This is where the smart money is going (Karman+, TransAstra).

Most current companies will quit or pivot. The historical pattern and the base-rate expectation. The category will persist in waves: new entrants will form each time costs fall, targets improve, and demand signals strengthen. The capability will eventually be achieved. The question is when, and by whom, and whether anyone makes real money doing it or whether the value accrues to downstream users.

The “this asteroid is worth quadrillions” framing is technically true and commercially useless.

The real question is narrower: Can someone build a reusable spacecraft that extracts water from a low-delta-v carbonaceous asteroid and delivers it to a cislunar depot cheaper than launching it from Earth? That’s the version of asteroid mining that might actually happen first, and even that is probably 10-20 years out with single-digit-to-25% odds depending on how far out you look.

I recovered from said injury but it’s not really possible to know whether BPC-157 augmented my recovery to a significant degree or if it just happened to be administered while my body recovered naturally on its own.

In other words: it’s possible BPC-157 did nothing significant to enhance recovery such that my recovery pace would’ve been the same without it.

I bought my BPC-157 from Canlab (a Canadian peptide lab); I’m unsure of their quality in 2026 but in 2019 it was good. I didn’t have any significant side effects or lasting reactions and found the process of “reconstitution” (i.e. adding bacteriostatic water to the peptide powder) very easy… bought some injection needles and antibacterial wipes and I was good to go.

If forced to guess… I’d guess that maybe BPC-157 helped a little? And it’s possible it wasn’t even the BPC-157 (could’ve been some hormetic response to an injection or placebo-augmented recovery? or even fluid injected at tendon adhesions to break them up? who knows.)

It’s very easy to placebo yourself with peptides due to the combo of factors that manipulate psychology (animal studies with some evidence, human testimonials claiming it miraculously healed their injuries, injecting something into your body, etc.).

In this piece I’m going to break down the current nuances of BPC-157.

My current thinking is as follows: Most people shouldn’t use BPC-157 in 2026. Unless you have serious injuries with no alternative treatments, the risks and costs outweigh potential upside until stronger/higher-quality evidence emerges to substantiate strong efficacy.

First: What Is a Peptide? Why Does the Word Mean Less Than You Think?

Let’s kill some confusion up front, because “peptide” has become a marketing word that makes people’s eyes glaze over with futuristic optimism, as if the term itself confers some special biological magic.

Many people now think “peptides” are underground natural cures for every biological disease and ailment that “mainstream medicine” doesn’t want you to know about.

A peptide is just a short chain of amino acids. That’s it. Amino acids are the building blocks of proteins. String a few together — roughly 2 to 50 — and you have a peptide. String more together and you have a protein. Your body is full of peptides.

Insulin is a peptide. Oxytocin is a peptide. GLP-1 receptor agonists — semaglutide, tirzepatide, etc. (drugs that are genuinely reshaping obesity medicine) are peptides.

So when someone says “I’m into peptides,” that’s a bit like saying “I’m into molecules.” It tells you almost nothing about what they’re actually taking, why, or whether it works.

The dumbest debate in peptide discourse is arguing via generalization of all peptides: (A) “peptides work” vs. (B) “peptides don’t work.” This is meaningless because each peptide is different.

It’s like saying “pills work” or “liquids don’t work.”

• Semaglutide is a peptide.

• BPC-157 is a peptide.

One has transformed obesity medicine with massive randomized controlled trials behind it.

The other has three tiny human studies and a cult following on Reddit.

Treating them as members of the same category that rise or fall together is downright stupid. Every peptide must be evaluated on its own evidence, pharmacokinetics, and safety profile.

GLP-1 receptor agonists being revolutionary does not make BPC-157 revolutionary any more than aspirin being safe makes thalidomide safe. They’re different molecules that do different things.

This point matters because the most common rhetorical move in peptide advocacy is to borrow credibility from GLP-1s and transfer it to whatever unproven compound is being discussed.

“Peptides are the future of medicine — look at Ozempic!”

Yes, semaglutide is impressive. No, that does not validate the vial of research-chemical BPC you bought from a website with a “.co” domain.

Something the peptide evangelists don’t emphasize: in traditional pharma, peptides are generally considered the worst of both worlds.

• They lack the oral bioavailability and metabolic stability of small molecules (your classic pills).

• They lack the target specificity and long duration of large biologics like monoclonal antibodies.

• They tend to have short half-lives, poor pharmacokinetics, and a habit of getting chewed up by enzymes before they can do much.

Successful peptide drugs — the GLP-1s, for instance — are heavily engineered.

Native GLP-1 has a half-life of a few minutes and is basically useless as a drug. Semaglutide lasts about a week because Novo Nordisk spent years and billions of dollars modifying it with fatty acid chains and amino acid substitutions to make it survive in the body. (This is one reason many are intrigued by mirror compounds — as mirrors can modify and make unstable short-half-life peptides useful.)

So the fact that some peptides are world-changing drugs does not mean all peptides are interesting. It especially does not mean that a random unmodified peptide you reconstitute from a vial and inject into your deltoid is automatically doing what you think it’s doing.

BPC-157: The Basics

BPC-157 stands for Body Protection Compound-157.

It’s a synthetic 15-amino-acid peptide derived from a protein found in human gastric juice. Or at least that’s the claim.

The compound was first characterized by Predrag Sikirić, a professor of pharmacology at the University of Zagreb School of Medicine in Croatia. He has legitimate academic credentials — 415+ publications, a PhD from Zagreb, an honorary doctorate from the University of Pécs in Hungary. This is a real scientist at a legitimate university.

But there are problems. The vast majority of the published literature on BPC-157 comes from his lab — over 150 papers. And an investigation by Undark, a well-regarded science journalism outlet, found that Sikirić is named on BPC-157 patent applications dating back to at least 1989, is listed as an owner of PharmaCotherapia (the company that sponsored the clinical trials), and is listed as CEO of a company called Diagen which owns a patent for a “special stable version” of BPC-157 that is currently for sale on its website.

These conflicts were not disclosed on any of his published papers that Undark reviewed.

When one lab dominates a compound’s entire evidence base AND has undisclosed financial interests in that compound, that’s a serious red flag. Not proof of fraud — but a setup where bias would be easy and independent replication is conspicuously absent.

And every single one of the hundreds of BPC-157 studies from this lab is positive. A 100% positive publication rate across that many experiments is unusual for any compound.

Does BPC-157’s Mechanism Make Sense?

This is the question worth taking seriously before anything else. Forget the anecdotes for a minute: Does BPC-157 have a plausible story for how it would work in the body?

The preclinical (animal) data shows a coherent wound-healing and repair picture: pro-angiogenic signaling through the VEGFR2-Akt-eNOS pathway, tendon fibroblast migration via FAK-paxillin, upregulation of growth hormone receptor expression in tendon tissue, and various anti-inflammatory and cytoprotective effects in gut, muscle, and connective tissue models.

And the animal results against controls are genuinely impressive. The most-cited study (Staresinic et al. 2003) transected rat Achilles tendons and compared BPC-157 to saline over 14 days.

BPC-157 rats showed increased load to failure (how much force before the tendon breaks again), higher Young’s modulus of elasticity (tendon stiffness/strength), significantly better walking function scores, more organized fibroblast and collagen formation under the microscope, and macroscopically smaller defect sizes trending toward full tendon integrity.

Other studies showed similar patterns: muscle crush models had improved load to failure and motor function, ligament transection showed reduced instability and restored biomechanics, and one fracture model performed comparably to bone grafting. In a spinal cord injury study, functional improvements persisted up to 360 days after a single treatment.

Crucially, the core mechanism has been independently confirmed by a team at Chang Gung University in Taiwan — a group with no apparent financial ties to Sikirić. Their in vitro work (Chang et al. 2011, Chang et al. 2014) showed that BPC-157 significantly accelerated tendon fibroblast outgrowth, increased cell survival under oxidative stress, and dose-dependently upregulated growth hormone receptor expression — all confirmed by Western blot and RT-PCR. They declared no conflicts of interest.

This is meaningful: an independent lab at a respected institution validated the cellular mechanism. BPC-157 does something real to tendon fibroblasts in a dish.

Independent replication is partial, not absent — but doesn’t cover the use case people buy BPC-157 for.

Three groups have now worked on BPC-157 independently.

A Taiwan team at Chang Gung University confirmed the cellular mechanism in vitro — specifically that BPC-157 accelerates tendon fibroblast migration, survival, and growth hormone receptor upregulation (Chang et al. 2011, 2014). They studied the right cells (tendon), but in a dish, not in a living animal.

A Chinese team at Air Force Medical University went further — they independently synthesized BPC-157, confirmed its biological activity in vivo in alkali-burn skin wound models and gastric ulcer models, completed preclinical safety evaluations, characterized the full PK profile, and are actively preparing for clinical trials in China for severe trauma and burns. They studied living animals, but for skin wounds and gut injuries — not tendon, ligament, or muscle healing.

Nobody outside Sikirić’s group has put BPC-157 into a living animal, injured its tendon or muscle, and measured whether it healed better versus control. The Taiwan team confirmed the right mechanism in the right cells but not in vivo. The Chinese team confirmed in vivo healing but for skin and gut, not MSK. The specific claims that drive the BPC-157 gray market — accelerated tendon repair, ligament healing, musculoskeletal recovery — remain single-lab findings from a researcher with undisclosed financial conflicts.

Caveats to consider:

1. Half-life. Animal pharmacokinetic studies show BPC-157 is cleared very rapidly — half-life under 30 minutes, parent drug undetectable after about 4 hours in rats and dogs. A molecule that disappears from your body in hours needs to trigger durable downstream signaling cascades to have a lasting effect. That’s possible — some short-lived signaling molecules do kick off longer repair processes — but it makes the “inject once a day and heal everything” narrative less straightforward than it sounds.

2. Local vs. systemic. After injection, BPC-157 distributes systemically — but not to where you want it. Animal PK showed the highest concentrations in kidney, liver, and stomach wall (clearance organs), while skeletal muscle levels were comparatively low. So if you inject into your abdomen hoping to heal your shoulder, most of the drug is going to organs that don't need it. This is the argument for injecting directly at the injury site: you deliver a high local concentration to the target tissue rather than relying on weak systemic distribution to get it there.

3. Effect size. Even if the mechanism is real, we have no idea how large the effect is in humans. A statistically real but clinically trivial acceleration of healing is a very different thing from the dramatic recovery stories on Reddit, X, Substack, and YouTube vlogs. Most biological effects exist on a spectrum, and “mechanistically plausible” tells you nothing about where on that spectrum a compound falls. It could be a 2% improvement in healing time. It could be 20%. We literally do not know.

4. Translation gap. “Real biology in a rat” and “clinically significant effect in a human” are separated by a canyon littered with the corpses of promising preclinical compounds that failed in humans. The history of drug development is largely a litany of things that worked beautifully in mice and then did nothing — or caused harm — in people.

My thoughts: The mechanism is plausible enough that BPC-157 could have some real biological effect for local tissue repair. But “plausible mechanism” is a low bar.

The Human Evidence: Worse Than You Think

A 2025 systematic review screened 544 papers on BPC-157. Of those, 36 made the final cut. Thirty-five were animal or cell studies. One was a human study. Let that ratio sink in: 544 papers screened, one human study included.

A separate 2025 narrative review identified 3 published human studies total. Here they are:

Study 1: Knee injection. A retrospective case series from a private clinic. Sixteen patients were contacted by phone after receiving intra-articular BPC-157 injections. No control group. No validated functional outcome measures. No blinding. Follow-up was literally a phone call. Seven of twelve surveyed patients reported pain relief lasting over six months. Importantly, this wasn't "I injected and felt better the next day" — patients received a single injection and were then surveyed 6-12 months later by phone, making it a retrospective self-report where placebo, natural healing, and real effect are essentially impossible to disentangle.

Study 2: Interstitial cystitis. A 12-patient pilot study using intravesical (bladder) BPC-157 injections. No control group. Single center. Private clinic.

Study 3: IV safety. A 2-person pilot study looking at whether IV administration was safe. Two people. Not an efficacy trial.

Why we can’t really extrapolate much from these studies:

There are zero randomized controlled trials for the use case people are actually buying BPC-157 for. The entire musculoskeletal healing evidence base — tendons, ligaments, recovery, the stuff that drives 99% of real-world demand — rests on a single uncontrolled phone survey of a dozen patients at a private clinic.

There are zero human studies with imaging endpoints. No MRI or ultrasound. No objective measurement of structural healing in any human study ever published. Every human outcome reported is patient-reported pain — which, as we’ll discuss, is the single most placebo-susceptible endpoint in medicine. Nobody has ever shown, with imaging, that BPC-157 actually changed the structure of a human tendon.

The one actual RCT that did exist was for a completely different indication. A Croatian pharma company called Pliva ran BPC-157 (as PL 14736) through Phase II clinical trials for ulcerative colitis in the early 2000s — not musculoskeletal injuries. The literature references these trials repeatedly as evidence that BPC-157 is “safe in clinical trials,” but the full efficacy results were never clearly published in a peer-reviewed journal. The paper trail consists mostly of other Sikirić papers citing the Pliva trials in passing.

A separate Phase I safety/PK trial was registered on ClinicalTrials.gov — and then canceled. The results were reportedly submitted in 2016, then recalled before quality control review. We don’t know what happened.

To summarize: 544 papers screened. Three tiny human studies, all uncontrolled. Zero RCTs for musculoskeletal use. Zero imaging endpoints. One IBD trial with unpublished results. One canceled registration.

That is the evidence base for a compound people confidently describe as a game-changer.

To be fair: all three published human studies were positive. But all three were also tiny, uncontrolled, and conducted at single sites with obvious investigator enthusiasm. The 2025 review flagged that all published BPC-157 studies are positive, which is a classic publication-bias red flag. In a world where negative results go unpublished (which is the world we live in), a 100% positive publication rate tells you more about the file drawer than about the drug.

Does this mean BPC-157 definitely doesn’t work? No. Does this evidence come anywhere close to establishing that it does? Also no.

We are in the epistemic gray zone where the honest answer is: we don’t know.

BPC-157: Unknown Long-Term Effects and Safety

We have no long-term human safety data. The published human studies are short-duration, tiny, and were not designed to detect adverse effects.

But I want to be logically honest here, because there’s a tension in the skeptical case that most BPC-157 critics don’t acknowledge: if the half-life argument undermines efficacy, it likely also undermines pharmacological risk.

• If BPC-157 is cleared so fast (under 30 minutes, undetectable after 4 hours) that it probably can’t do much good, then by the same logic it probably can’t do much harm as a molecule.

• We can’t selectively deploy the half-life argument only when it supports skepticism. A compound that’s too short-lived to heal your tendon is probably also too short-lived to cause major harm (e.g. cancer) through direct pharmacological action.

There is a counterargument: short-lived molecules can trigger longer-lasting downstream signaling cascades. A brief VEGF signal could theoretically kick off an angiogenic process that persists for days.

This cuts both ways: if BPC-157 can trigger lasting harmful cascades despite rapid clearance, it could plausibly trigger lasting beneficial ones too.

You don’t get to use “but downstream effects!” only for the risk column.

So what are the real risks? I’d separate them into two categories:

1. Pharmacological risks (probably low, but unknown). The theoretical concerns about chronic angiogenic stimulation, fibrotic remodeling, and cancer promotion are mechanistically real but also somewhat undercut by the same PK data that makes efficacy doubtful. If you’re using BPC-157 for a short course for a specific injury, the direct pharmacological risk is probably modest. If you’re using it chronically as a daily “optimization” protocol — which some people do — the risk calculus changes because you’re repeatedly stimulating these pathways even if each individual dose is cleared quickly. Chronic repeated exposure is a different animal than a two-week course.

2. Supply chain and procedural risks (probably the bigger real-world concern). This is where the actual danger likely lives. The FDA explicitly flags immunogenicity risk, peptide-related impurity concerns, and API characterization issues for compounded BPC-157. FDA has documented compounded products made under insanitary conditions. Contamination, wrong product, endotoxins, degraded peptide, potency variation, and self-injection technique failures — these risks have nothing to do with BPC-157’s mechanism and everything to do with the unregulated delivery system. Even if the molecule itself is pharmacologically benign, injecting mystery powder from an unverified supplier into your body is not.

Drug interactions. Zero systematic data on how BPC-157 interacts with other medications, supplements, or other peptides in the multi-compound “stacks” that many users take.

My thoughts: for a short course from a quality-verified source, the direct pharmacological risk is probably low. The supply chain risk is probably higher than the pharmacological risk. And for chronic repeated use, we genuinely don’t know — the absence of reported problems is not evidence of safety, it’s evidence of an inadequate surveillance system.

The Placebo Problem: Why It’s Extremely Easy to Fool Yourself with BPC-157

This is the core of my skepticism, and it’s the same skepticism I apply to myself. When I used BPC-157 and thought “hmm, maybe that helped,” I was stacking every single variable known to amplify placebo responses. Let me walk through all of them, because the number of confounders operating simultaneously in BPC-157 self-experimentation is genuinely extraordinary.

BPC-157’s primary use case in the wild is recovery from musculoskeletal injuries — tendonitis, ligament strains, overuse injuries, joint pain. These are exactly the conditions where placebo effects are largest, most durable, and most convincing to the person experiencing them.

Research has shown that in conditions like patellar tendinopathy, placebo effects from nonsurgical treatments are clinically significant and can persist for up to 12 months. Pain research broadly shows that more invasive procedures generate larger placebo responses, especially for subjective outcomes like “how does your knee feel?”

Here is a non-exhaustive list of every variable working to convince you that BPC-157 is working, whether or not it actually is:

1. Regression to the mean. People start taking BPC-157 when their injury is at its worst — that’s what motivates them to try something new. But injuries at their worst tend to improve regardless of intervention, because that’s how fluctuating conditions work. You are virtually guaranteed to feel better after starting BPC simply because you started it at a low point. This alone accounts for a huge share of “it worked for me” stories.

2. Natural healing. Your body heals injuries. That’s what it does. Tendons repair. Inflammation resolves. Tissue remodels. This process takes days to weeks to months and it happens whether or not you inject anything. When you take BPC-157 for three weeks and your tendonitis improves, you have no way of knowing whether the improvement would have been identical without it.

3. Invasiveness of the procedure. Injection-based treatments generate substantially larger placebo effects than oral treatments. A needle going into your body signals to your brain that something medically significant is happening. The placebo literature on this is robust: the more invasive the delivery method, the stronger the placebo response. BPC-157 is injected. This alone amplifies the perceived effect.

4. Ritual complexity. This is not popping a pill. You are: researching dosing protocols, ordering specialized supplies (bacteriostatic water, insulin syringes, alcohol swabs), reconstituting lyophilized powder by carefully injecting water into a vial, calculating microgram doses, drawing the correct volume, selecting injection sites, performing the injection with proper sterile technique, and doing this daily for weeks. This is an elaborate, multi-step medical ritual. The complexity and intentionality of the process powerfully reinforces the belief that something real is happening. The harder something is to do, the more your brain insists it must be working — otherwise why would you be doing it?

5. Esoteric knowledge effect. BPC-157 is not something your average person knows about. You found it through deep-dive research, niche communities, or a cutting-edge doctor. You know things other people don’t. You are part of a group that has access to advanced tools. This “secret knowledge” framing is psychologically potent — it creates identity investment in the treatment working and raises the perceived sophistication and therefore efficacy of the intervention.

6. Status and authority halo. Peptides are heavily associated with wealthy tech founders, elite athletes, and concierge doctors who charge $1,000+ per consultation. When you see that a successful founder in SF is injecting BPC-157, or that a doctor to professional athletes recommends it, your brain doesn’t process that as “a rich person is also susceptible to placebo.” It processes it as “this person is smart, successful, and has access to information I don’t — they must know something.” This is a massive placebo amplifier. The implicit logic is: ultra-wealthy, high-IQ people with access to the best doctors in the world are doing this, so it must be legitimate. But wealth and intelligence do not protect against placebo effects, confirmation bias, or motivated reasoning — if anything, smart people are better at constructing post-hoc rationalizations for why their expensive intervention is working. The status halo transfers perceived credibility from the person to the product, and it cascades outward: normies see tech bros doing it, tech bros see elite doctors doing it, elite doctors see other elite doctors doing it, and nobody in the chain has controlled evidence.

7. Financial sunk cost. You paid $50-150+ per vial, plus syringes, bacteriostatic water, alcohol swabs, and possibly a consultation fee. You want this to work. You need it to work to justify the expense. Sunk cost bias is one of the most powerful and universal cognitive distortions, and it operates below conscious awareness. The more you pay for a treatment, the more effective you perceive it to be — this is documented in the placebo literature.

8. Social investment and identity. You told friends about it. You posted about it. You’re part of the peptide community now. Your identity includes “person who is smart enough to use peptides.” Admitting it might be placebo means admitting you were wrong publicly, and human beings will go to remarkable lengths to avoid that. This creates a community-wide filter where positive reports are amplified and negative reports are suppressed — not by conspiracy, but by ordinary social psychology.

9. Confirmation bias. Once you believe BPC-157 is working, you notice every improvement and attribute it to the peptide. You ignore or discount the days where pain is the same or worse. You don’t track your baseline systematically. You’re not blinded. You’re not comparing to a control. You are running the least rigorous experiment possible and then trusting your subjective impression of the results.

10. Concurrent behavioral changes. This is the one almost nobody accounts for. When you start a BPC-157 protocol for an injury, you almost always simultaneously: reduce training volume on the affected area (because you’re “letting the BPC work”), sleep more (because recovery forums told you to), ice or stretch more carefully, pay more attention to form, maybe add other recovery modalities. Any or all of these changes could be responsible for the improvement. You changed five variables and credited the most exciting one.

11. Subjective endpoints. The primary outcome people use to evaluate BPC-157 is “how does my injury feel?” — a purely subjective self-report. Pain perception is influenced by mood, sleep, stress, expectation, attention, and context. It is the single most placebo-susceptible endpoint in medicine. Nobody is getting serial MRIs to see if their tendon actually looks different. They’re going by feel, and feel is unreliable.

12. Community reinforcement loops. You go to Reddit or Twitter or a peptide forum. You read 50 posts from people saying it changed their life. You post your own positive experience. Other people read your post and start BPC with even higher expectations, which amplifies their placebo response, and they post about it too. This is a self-reinforcing cycle that produces an overwhelming body of “evidence” that is actually just expectations echoing through a community. Negative experiences get downvoted, dismissed (”your source was bunk”), or simply never posted because people who didn’t notice anything don’t bother writing about it.

13. Narrative coherence. You’ve read about the mechanism — VEGF, angiogenesis, fibroblast migration and you have a scientific story for why this should work. That story, whether or not it translates to actual human efficacy, makes the treatment feel legitimate and rational. A placebo with a mechanistic narrative feels more real than a placebo without one. Your brain uses the narrative to explain the perceived improvement, creating a self-reinforcing loop of theory and experience.

Every single one of these variables is pushing in the same direction: toward making you believe BPC-157 is working. And they stack multiplicatively, not additively. You are not dealing with one confounder; you are dealing with all thirteen operating simultaneously, in a domain (pain and recovery) where placebo effects are already at their maximum potency.

This is the ivermectin problem. People took ivermectin for COVID, their immune systems fought off the virus over the following days (as immune systems do), and they attributed their recovery to the drug. The same logic applies to chiropractic adjustments, many supplements, homeopathy, and a long list of interventions where the natural course of the condition does most of the work and the intervention takes the credit.

Could BPC-157 still have a real effect on top of all this? Yes. But the point is that the anecdotal signal you see online would look very similar if BPC-157 were a completely inert saline injection. Every single testimonial is fully explainable by the variables above without any pharmacological contribution from the peptide. That doesn’t prove it’s inert — but it means the anecdotal evidence proves nothing either way. I say this as someone who took BPC and felt... maybe better?

The Supply Chain Problem: You Might Not Even Be Getting BPC-157

Even if BPC-157 has real effects, the version you’re injecting might not be BPC-157.

When you order from a gray-market peptide supplier (which is what most people are doing, since BPC-157 is not FDA-approved and not legally available through normal pharmacy channels) — you are trusting an unregulated manufacturer to deliver the correct compound at the correct purity at the correct concentration. The failure modes here are numerous and serious:

• You might not be getting BPC-157 at all. Without third-party testing (and most users don’t do third-party testing), you have no way to verify that the vial contains the peptide sequence it claims to contain. It could be a different peptide, a partial sequence, or degraded product.

• Contamination. FDA has documented compounded BPC-157 acetate products made under insanitary conditions. Endotoxins, heavy metals, microbial contamination, residual solvents — these are all real risks when manufacturing happens outside of regulated GMP facilities. You are injecting this directly into your body.

• Potency variation. Even if the compound is BPC-157, the actual amount per vial can vary wildly from what’s on the label. Under-dosing means you’re getting nothing. Over-dosing means you’re getting more of an uncharacterized compound than you intended.

• Reconstitution and self-injection risks. The reconstitution process itself introduces contamination risk if sterile technique is imperfect — which it often is when performed at home. Injection technique matters: wrong needle gauge, wrong depth, wrong site, or insufficient skin prep can cause local infections, abscesses, nerve damage, or injection into the wrong tissue compartment.

This means the real-world risk profile of BPC-157 as actually used is worse than the risk profile of the compound itself. Even if BPC-157’s pharmacology is relatively benign, the delivery system (unregulated manufacturing, unverified product, self-injection) introduces its own layer of risk that has nothing to do with the peptide’s mechanism of action.

Cancer and BPC-157: Everything Unknown

This deserves its own section because it’s the most serious mechanistic concern — and the discourse around it is muddled.

The simple framing you usually hear is “BPC-157 either causes cancer or prevents it.” The more accurate framing is that it could plausibly do both — to different cancers, through different pathways, simultaneously. Biology is not a single toggle.

1. The pro-cancer concern is the more serious first-principles worry. BPC-157 is pro-angiogenic — it promotes new blood vessel formation. VEGF/VEGFR-driven angiogenesis is a core pathway in tumor growth; tumors need blood supply to expand, and they actively recruit new vasculature. Nitric oxide signaling, which BPC-157 also modulates, plays a dual role in cancer biology and can support angiogenesis, invasion, and metastasis in some settings. So for cancers that are angiogenesis-dependent — which is most solid tumors — you probably do not want to be stimulating the pathways that feed them. If you have an occult tumor somewhere that’s currently starved for blood supply, BPC-157’s mechanism of action is exactly what that tumor wants.

2. The anti-cancer argument also exists in the literature, but it’s thin. The best skeptical review I found says there are no published in vivo data showing BPC-157 reduces tumor progression, tumor volume, or metastasis. The “anti-tumor” case leans heavily on a single old melanoma cell-line experiment that has never been meaningfully replicated. There are theoretical arguments about nitric oxide’s cytotoxic effects on certain tumor cells, and BPC-157’s anti-inflammatory properties could theoretically attenuate inflammation-driven carcinogenesis in some contexts. But these are mechanistic speculations, not demonstrated outcomes.

The uncomfortable reality is that both could be true simultaneously: BPC-157 could theoretically help protect against certain inflammation-driven cancers while increasing the risk of or bolstering others that are angiogenesis-dependent. Cancer is not one disease. Different tumor types have different dependencies, and a compound that modulates angiogenesis, nitric oxide, and inflammation will interact with each of them differently.

The reality: we don’t know whether BPC-157 is net pro-cancer, anti-cancer, or neutral in humans — and the answer might differ by cancer type. And the half-life consistency applies here too: if BPC-157 is cleared in under 30 minutes, the magnitude of pro-angiogenic signaling from a single dose is probably modest. The concern becomes more real with chronic repeated use, where you’re pulsing VEGF-pathway stimulation daily for weeks or months. For a short course targeting a specific injury, the cancer risk is probably very low. For long-term daily use as an “optimization” protocol, the uncertainty is legitimate.

Either way, cancer history, active malignancy, or unexplained masses should be treated as hard contraindications.

The Incentives: Big $$$ From Selling Peptides

Follow the money, and the BPC-157 hype makes perfect sense regardless of whether the compound works.

For sellers and compounding pharmacies: These compounds are cheap to synthesize, have no patent protection, and can be sold at enormous markups. A vial that costs cents to produce sells for $50-150. Unlike pharma companies, you don’t have to spend a billion dollars on clinical trials first. As AP has reported, clinics are selling consults and shipping peptide kits while these products are still unapproved and often labeled “research use only.” There is also a market positioning incentive: in a world where genuinely innovative new drugs are rare and expensive, “peptides” offer sellers a product category that sounds cutting-edge and scientific while being cheap and unregulated. The term itself does marketing work — it sounds more sophisticated than “supplement” and more natural than “drug.”

For influencers and content creators: Peptides are perfect content. Novel, scientific-sounding, appealing to the optimization-obsessed demographic, and offering a narrative of secret knowledge. Every peptide video, podcast, and newsletter generates engagement precisely because the topic sits at the intersection of aspiration, contrarianism, and scientific-sounding complexity.

For clinics and concierge doctors: Peptide prescribing is lucrative. Patients pay cash. Insurance is not involved. The incentive to believe these compounds work is enormous because your business model depends on it.

For VC-backed startups: This is where the incentives get truly massive. Max Marchione — the peptide advocate from the Shkreli debate — isn’t just a “converted skeptic.” He’s the CEO of Superpower, a health tech company that has raised $51 million in venture funding on the explicit thesis that peptides are the future of consumer health. The company sells peptides and biomarker testing to members, markets itself as building “the AI-Peptide Infrastructure for Mainstream Wellness,” and cites the BPC-157/TB-500 “Wolverine stack” as a core trend it’s riding. Marchione has told interviewers that his employees inject each other with experimental peptides at Friday breakfasts “because we think it’s fun.” When someone with a $51 million company built on peptides tells you peptides work, that is not a disinterested scientific opinion — it is a man whose entire valuation depends on you believing him. This doesn’t make him wrong, but it does mean you should weight his claims accordingly.

For the user community: People who inject peptides have invested time, money, social capital, and physical discomfort into the practice. Admitting it might be placebo is psychologically costly. This creates a community that systematically amplifies positive reports and dismisses skepticism.

The enthusiasm you see online is exactly what you would expect to see whether or not the compound has real efficacy. The incentive structure generates hype independent of the underlying pharmacology.

Why Hasn’t Pharma Picked up BPC-157?

This is arguably the single most damning piece of evidence against BPC-157, and it’s not a study — it’s a revealed preference.

BPC-157 has been known since 1993. That’s over three decades. The pharmaceutical industry is staffed by thousands of people whose entire job is to find molecules that work and turn them into profitable drugs. Drug hunters — people like Shkreli, whatever you think of him personally — scour the global literature for compounds with therapeutic potential. They are ruthlessly incentivized to find things that work, because a compound that genuinely heals tendons or accelerates wound repair would be worth billions.

One company did pick it up — but not for the use case people actually buy BPC-157 for. Pliva, a well-respected Croatian pharmaceutical company, licensed BPC-157 and ran it through Phase II clinical trials for ulcerative colitis (as PL 14736). Not for tendon healing or musculoskeletal recovery… for IBD. Phase I safety data showed it was safe and well-tolerated. A randomized, double-blind, placebo-controlled Phase II study was started. And then... the program stopped. The full results were never published.

What Shkreli and BPC-157 critics leave out: Pliva didn’t just abandon BPC-157 — Pliva ceased to exist as an independent company. In October 2006, Barr Pharmaceuticals acquired Pliva. In December 2008, Teva Pharmaceuticals acquired Barr. The company that was developing BPC-157 was acquired twice in two years and absorbed into Teva, the world’s largest generics manufacturer. Pipeline culling during acquisitions is extremely common — small experimental peptide programs from acquired companies get killed routinely, regardless of whether the compound works. We genuinely do not know if Pliva stopped because BPC-157 failed in Phase II, or because the company got swallowed and the new owners deprioritized a niche peptide program. Both are plausible.

Important: Pliva never tested the musculoskeletal use case — so even if Phase II for IBD failed on efficacy, that would not directly tell us whether BPC-157 heals tendons.

But the broader revealed-preference argument still holds: musculoskeletal healing has been BPC-157’s dominant anecdotal use case for well over a decade.

The preclinical literature on tendon/ligament/muscle repair has been published and available. Drug hunters have had every opportunity to pick this up for MSK indications. Nobody has — though as noted above, an independent Chinese team is preparing for clinical trials for severe trauma and burns, so someone outside Croatia is taking it seriously enough to move toward the clinic. The patent excuse is weaker than people think: you don’t need a composition-of-matter patent to make money in pharma. There’s 5-year NCE exclusivity, 3-year exclusivity for new clinical investigations, 7-year orphan drug exclusivity, and various other regulatory protections. If BPC-157 demonstrably healed tendons in a controlled trial, someone would find a way to monetize it.

The fact that nobody has — after 30+ years of the compound being known, after a pharma company tried it for a different indication then got acquired, after thousands of drug hunters have had the opportunity — is not proof that it doesn’t work. But it is a meaningful signal from the people with the most expertise and the most incentive to know.

What Would Settle the BPC-157 Debate?

If we’re serious about knowing whether BPC-157 works, here’s what a proper study would look like — and the fact that nobody has done it yet is itself informative:

• Design: Double-blind, placebo-controlled, randomized trial. This is non-negotiable. Given the strength of placebo effects in musculoskeletal medicine, any unblinded or uncontrolled study is essentially worthless.

• Population: Patients with a specific, well-defined injury type where you can get a reasonably homogeneous cohort. Something like partial-thickness rotator cuff tears, or Achilles tendinopathy, or lateral epicondylitis (tennis elbow) — injuries that are common enough to recruit for, specific enough to standardize, and slow enough to heal that you could detect an acceleration.

• Objective endpoints: This is critical. You cannot rely on patient-reported pain scores as your primary outcome, because that’s exactly where placebo effects dominate. You need imaging: serial MRI or ultrasound showing structural healing — tendon thickness, tear size, tissue organization, neovascularization. Ideally paired with functional biomechanical testing: grip strength, range of motion, load tolerance. The primary endpoint should be something a scanner can see, not something a patient reports.

• Control for confounders: Both groups need identical rehabilitation protocols, identical activity restrictions, identical follow-up schedules. The injection procedure itself should be identical (same volume, same site, same technique) with the control group receiving a saline injection.

• Duration and follow-up: Long enough to capture both the acute recovery window (weeks) and any lasting structural difference (months). At minimum 12 weeks, ideally 6-12 months.

• Sample size: Powered adequately to detect a clinically meaningful difference — which requires first defining what “clinically meaningful” even means for BPC-157, something nobody has done. A 5% improvement in healing time might be statistically real but clinically irrelevant. You’d want to power for at least a 20-30% improvement in the primary imaging endpoint to justify the cost and risk of treatment.

• Independent investigators: Not Sikirić’s lab. Not a clinic that sells peptides. An academic center or CRO with no financial interest in the outcome.

This study would cost a few million dollars. Not billions. If the peptide community is serious — if the companies selling BPC-157 vials and the influencers promoting it really believe in what they’re selling — this is entirely within reach.

The fact that it hasn’t been done tells you something about either the confidence level of the people profiting from the status quo, or the difficulty of the regulatory pathway, or both.

The Debate: Shkreli vs. Marchione

I recently watched a debate between Martin Shkreli and Max Marchione on the broader peptide question, with BPC-157 as the central case study.

Shkreli’s core argument: BPC-157 has no credible human evidence, the preclinical work is dominated by one lab, the compound has terrible pharmacokinetic properties, and the anecdotal evidence is indistinguishable from placebo. His strongest argument was the broader revealed-preference point: BPC-157 has been known for 30+ years, and nobody has taken it through clinical trials for the musculoskeletal use case that drives real-world demand. That's worth noting — though as discussed above, the one pharma company that did try it tested a completely different indication (IBD), then got acquired twice and ceased to exist, so their abandonment doesn't tell us what Shkreli implies it does.

Marchione’s core argument: Thousands of doctors have given BPC-157 to patients for 10-20 years. Millions of patients report life-changing results. His own father went from daily painkillers to pain-free in three days. His cofounder went off biologics for autoimmune disease. The gray market is dangerous, so these compounds should be legalized and regulated rather than driven underground.

Who won on the merits? Shkreli, and it’s not particularly close on the evidence question. His core logical framework — if you claim efficacy, show controlled evidence; anecdotes are not proof; the burden of proof is on the claimant — is simply the correct epistemic standard. Marchione’s case relied heavily on appeal to authority (some trusted doctors use and recommend it), appeal to popularity (thousands of doctors, millions of patients), and appeal to emotion (his father’s pain, his cofounder’s illness). These are powerful rhetorical tools but they are not evidence.

Marchione’s strongest point was the policy argument: a gray market exists, people are already using these compounds, and regulated access through GMP-certified facilities would be safer than the current underground supply chain. That’s a reasonable harm-reduction argument regardless of whether BPC-157 actually works. Shkreli’s response — “I prefer no market” and “we can arrest the genie” — was his weakest moment, but it’s worth unpacking why it was weak, because Shkreli was actually making a strong argument for a different compound and then misapplying it.

Earlier in the debate, Shkreli made a good IP argument about retatrutide — a GLP-1 drug that Eli Lilly is actively spending billions to develop and bring through FDA approval. People pirating that compound through Chinese manufacturers while Lilly is mid-development is straightforward intellectual property theft, and it does undermine the R&D incentive structure that produces new drugs.

You want pharma companies to keep investing in drug development? Then you should respect the IP of drugs they’re actively developing. That argument is smart and something I agree with because it sustains prosperity in pharma.

But for BPC-157, the IP argument evaporates entirely. Nobody holds a patent on it. Pliva licensed it, tried to develop it, and walked away. Sikirić’s lab is an academic research group, not a pharmaceutical company with an investment to recoup. It’s a 30-year-old compound the industry abandoned. There is no intellectual property to protect.

Shkreli gets a bit paternalistic when he says “I prefer no market” for peptides like BPC-157 because the argument is not about IP; it’s: “you shouldn’t be allowed to self-experiment with non-FDA-approved substances.”

That’s a defensible philosophical position but I disagree with it if the substance is relatively safe with potential upside; you should have freedom to do whatever you want with your body. And this is a different argument than he made about reta.

If a compound is unpatented, not being actively developed by anyone, and not FDA-approved, then preventing informed adults from accessing it is anti-freedom.

Regulatory Status: No, Peptides Haven’t Been “Legalized”

As of early 2026, BPC-157 has not been formally legalized for routine compounding in the United States. The FDA currently lists it as a 503A Category 2 substance — meaning the agency has identified significant safety concerns.

FDA warning letters state that BPC-157 is not a component of any FDA-approved human drug and does not appear on the 503A bulks list.

There is political pressure to change this. RFK Jr. has publicly discussed loosening restrictions on certain peptides. But political pressure and completed regulatory action are different things. Policy noise exists; finalized legalization does not — yet.

If You’re Going to Use BPC-157 Anyway

This is not medical advice. But if you’ve read all of the above and still want to try BPC-157, here’s my practical take:

• Minimal dose, minimal duration. Do not make this a routine. If you’re going to experiment, use the lowest dose that the community considers potentially effective, for the shortest duration plausible, for a specific injury — not as an ongoing “optimization” protocol. The unknown long-term risk profile means that chronic use is a worse bet than a short targeted course.

• Quality-verified sourcing only. If you can’t verify what’s in the vial, you’re not just taking a gamble on efficacy — you’re taking a gamble on contamination, potency, and identity. Third-party certificate of analysis (COA) from an independent lab testing for peptide identity, purity, endotoxins, and sterility is the minimum bar. If your supplier can’t or won’t provide that, find a different supplier or don’t use it.

• Local injection makes most sense. The two human efficacy-style studies both used local compartment delivery (intra-articular, intravesical). If BPC-157 has a real tissue effect, the most coherent expectation is local repair signaling in the tissue directly exposed, not distant healing from a random subcutaneous injection.

• Accept that you won’t know if it worked. Your experience will be confounded by everything else you’re doing. You will feel like it helped. That feeling is not reliable evidence. Be honest with yourself about this.

The Bottom Line

I don’t know if BPC-157 works. Nobody does. The mechanism is plausible but human evidence is nearly nonexistent. The effect size is unknown (assuming a real effect) and the long-term safety profile is unknown.

The peptide supply chain is largely unregulated. And the entire ecosystem around it is structured to generate enthusiasm whether or not the compound does anything.

The honest label for BPC-157 in 2026 is: unproven peptide with a plausible repair mechanism, almost no human evidence, unknown effect size, unknown long-term safety, and marketing that has wildly outrun the science.

And remember: not all “peptides” should be lumped together. Semaglutide being a success does not make BPC-157 effective. Thymosin alpha-1 being approved in 35 countries does not mean the next peptide on your Instagram, X, YouTube, TikTok, Reddit, etc. feeds are safe and effective.

Just because a random MD (medical doctor) recommends BPC-157 does not mean that it works; without randomized controlled trials we don’t know the actual effect from a placebo response and/or organic recovery.

I’ll update my views (like any logical person should do) on BPC-157 if someone publishes a well-designed, adequately powered, randomized controlled trial in humans; we need this data to know if it actually does anything significant.

Lastly, if and when BPC-157 turns out to be effective and/or safe in the future, some will say: “See! I told you… I knew all along it was effective!” These people are lying and are mostly stupid… they didn’t have evidence, they just guessed correctly.

Yet since that month-end peak, the G4 global liquidity proxy has kept climbing.

If you’ve been watching the “BTC tracks global M2” charts that have circulated since Lyn Alden’s widely-read Bitcoin: A Global Liquidity Barometer piece, you might be confused.

The money printer line keeps going up. Bitcoin keeps going down. Is the thesis broken?

The short answer: the thesis is still accurate.

But the version many people are running in their heads — draw one global M2 line, expect BTC to mirror it with some lag — does not survive the last two years very well.

The useful version is more specific, more lagged, built on G4 broad money rather than just U.S. M2, and more regime-dependent than the meme suggests.

Here’s what I found.

What I Built and Why

Most of the global liquidity overlays you see on Twitter use U.S. M2 as the anchor, sometimes mixed with a few other countries. The problem isn’t M2 itself — it’s geographic scope. As I’ll show later, U.S. M2 accounted for only 9% of G4 liquidity growth over the past year. Watching it alone means watching a small fraction of the signal.

One related note: the Fed’s 2020 H.6 revision pulled savings deposits into M1, adding roughly $11.2 trillion overnight while leaving M2 unchanged. Post-2020 M1 now behaves more like old M2, which is why I avoided it entirely and built the composite on broad money measures instead.

Instead, I constructed two proxies from scratch.

1. The first is a USD-converted broad-money composite of U.S. M2, euro area M3, China M2, and Japan M3 — the G4 blocs that collectively account for the vast majority of global money.

2. The second is a local-currency quantity-only index of the same four series, which strips out FX translation effects so I could check whether dollar strength was doing all the work.

The USD liquidity proxy sums nominal broad-money stocks after USD conversion and rebases the total to January 2022 = 100; the local-currency version tracks the same blocs without dollar translation to isolate quantity effects from FX effects. Because the money data are monthly, all BTC/liquidity overlays and lag tests use month-end BTC values unless otherwise noted.

BTC prices come from the Coinbase/FRED monthly series, extended to current spot.

The panel runs from January 2022 through January 2026, the latest month where all four money supply series have been published. That gives me 49 monthly observations — enough for exploratory lead/lag tests, but not enough to claim stable laws.

The Headline Correlation Is Real — But Misleading

Alden reported a 0.94 BTC-to-global-liquidity correlation over May 2013 to July 2024, with directional consistency of 83% over 12-month windows and 74% over 6-month windows.

Those are strong numbers.

In my updated sample, the level correlation is still high: 0.86 for the USD-converted proxy and 0.85 for the quantity-only version. Both series trend upward over time, and so does BTC — so far, so good.

But because both series are non-stationary and trending, the levels correlation should be treated cautiously; the more useful signal is in lagged changes, not the raw overlay.

Here’s where it breaks down. When I look at contemporaneous changes instead of levels, the relationship nearly vanishes:

• Same-month 1-month changes: correlation of 0.04 (basically zero)

• Same-month 3-month changes: 0.28

• Same-month year-over-year changes: -0.06

That is the key finding.

The headline correlation is real, but a meaningful part of it is a shared trend, not a same-month trading signal.

If you’re using a global M2 overlay as a timing tool for what BTC does this month, the data say you’re mostly trading noise.

The Signal Is in the Lag

The relationship improves materially once you let liquidity lead BTC by a few months.

This is what many people get wrong.

They expect simultaneous tracking, when the actual transmission mechanism takes time.

When I tested the predictive correlation between a 3-month liquidity change and a future BTC 3-month return across different lags, the peak hit roughly 0.50 at a 1-month lag and 0.49 at a 2-month lag.

By 3 months the signal has faded to about 0.28, and beyond 4 months it flips negative (likely a mean-reversion artifact from the small sample).

An exploratory Granger-style test suggests the direction runs from liquidity → BTC: significant at 1–2 lags (p ≈ 0.04), while the reverse (BTC → liquidity) is not supported in this sample.

Bitcoin does not appear to lead global liquidity in any useful way — it follows it, with a delay. And my central estimate for the lag is ~2 months, with a realistic band of 1 to 3 months.

The Lag Is Shifty

One of the more interesting findings is that the optimal lag is not constant.

In rolling 18-month windows over the last two-plus years, the best-fit lag has shifted:

• Early 2024: mostly 1–2 months

• Mid-2024: briefly compressed toward 0–1 month (around the ETF-driven rally)

• Late 2025 through January 2026: widened back toward 2–3 months

The rolling mean lag across the full window is about 1.9 months using 1-month changes and 1.6 months using 3-month changes.

The median is 2 months either way.

The instability matters: any model that assumes a fixed lag will periodically look brilliant and periodically look broken.

The Magnitude: Not a Fixed Beta

On the question of “how many percent does BTC move per percent of liquidity change” — there is a usable estimate, but it is not a stable law.

In the 2024–2026 subsample, a 1% rise in the USD-converted liquidity proxy over 3 months corresponded to roughly +4.2% BTC over the next month or +6.6% BTC over the next 3 months.

The standalone correlation there is about 0.42, R² about 0.17–0.18, and next-quarter sign agreement about 68%.

The full 2022–2026 sample gives a wider range: the beta on 3-month liquidity changes predicting 3-month BTC returns is about 3.2x with an R² of only 0.08.

In other words, the relationship is real but explains less than a fifth of BTC’s variance in any given quarter.

The right mental model is “a positive liquidity impulse usually gets multiplied in BTC over the following quarter,” not “1% liquidity always equals X% BTC.”

Why It Broke in 2025–2026

By January 2026, the USD-converted liquidity proxy was growing about 12% year-over-year. BTC was down about 22% year-over-year.

That’s a massive divergence — and it has a specific explanation.

Market structure changed. The SEC approved spot bitcoin ETPs in January 2024, which made institutional ETF flows a much more important transmission channel between macro conditions and BTC price. When those flows reverse, they can overwhelm the slower broad-money signal for months at a time.

And reverse they did. U.S. spot BTC ETFs saw more than $3 billion in outflows in January 2026, following heavy outflows in November and December 2025. Post-ETF bitcoin can temporarily trade like a macro-sensitive institutional product with large flow shocks — not just like a pure liquidity beta.

At the same time, the wave of corporate bitcoin treasury plays — led by Strategy (formerly MicroStrategy) and a growing list of imitators — had become a crowded trade by late 2025, with many of these companies overextended on convertible debt and equity issuance.

When BTC softened, their stock premiums compressed and the marginal buying pressure that had amplified the 2024–2025 rally flipped from tailwind to headwind.

The other piece of the puzzle is the price of liquidity vs. the quantity.

Global money supply is growing, but that money is more expensive. Oil has surged roughly 40% since the Iran conflict escalated, the U.S. 2-year Treasury yield has risen nearly 30 basis points, and futures have pushed the expected first Fed cut from June toward September.

The ECB is warning that the oil shock creates upside inflation risk. The BOJ keeps tightening. The quantity of money is up; the cost of accessing it is also up.

BTC is caught in the middle.

Who Is Actually Driving Global Liquidity?

Not everyone is contributing equally. When I decomposed the January 2025 to January 2026 growth in the USD-converted proxy:

• China: 61% of the increase ($6.16 trillion in USD terms)

• Euro area: 30% ($3.00 trillion)

• United States: 9% ($0.92 trillion)

• Japan: essentially zero

If you’re tracking U.S. M2 alone, you’re watching 9% of where the G4 liquidity growth actually came from over the past year.

This matters because BTC trades on the marginal dollar and liquidity impulse, not just nominal global money quantity. Chinese M2 growing at 9% y/y is the biggest current liquidity tailwind, and the PBOC has signaled more RRR and rate cuts ahead. But Chinese broad money expansion does not flow into BTC markets with the same directness as U.S. dollar liquidity. It’s supportive, not sufficient.

A Composite Scoring System In 2026

If I had to distill this into a single scoring system, I’d normalize the inputs and weight them roughly as follows. The weights below are pragmatic judgment weights, not coefficients from a fully specified regression — treat this as a framework, not a backtest.

BTC 6-month score =

• + 0.40 × z(global liquidity 3m change, lagged 2 months)

• − 0.25 × z(2-year yield change)

• − 0.15 × z(DXY change)

• + 0.15 × z(ETF net flows / AUM)

• + 0.05 × z(cycle state)

The raw rule of thumb behind that weighting: when rates, the dollar, and flows are all neutral, 1% more global liquidity over a quarter is worth roughly 5–7% more BTC over the following quarter.

Right now I’d haircut that beta because yields, oil, and the dollar are all working against it.

The best rough tools for BTC, in order of usefulness:

1. FX-adjusted global liquidity impulse — best medium-term signal, 1–3 month lead

2. ETF net flows — best short-term signal, 0–2 week lead

3. Front-end rates and the dollar — same-day to 1-month transmission

4. Oil/geopolitical shocks — because they move #2 and #3

5. Cycle-state filters — valuation sanity checks, not timing tools

Where This Points Forward

The current macro mix is positive on money quantity, mixed on transmission.

• U.S. M2 is up 4.3% y/y, euro area M3 up 3.4% y/y, China M2 up 9.0% y/y, Japan M3 up 1.2% y/y.

• But the Fed is still at 3.50–3.75%, the ECB is at 2.00/2.15/2.40%, the BOJ is at 0.75% and still tightening, and the PBOC is the only major central bank in full easing mode.

The demographic backdrop reinforces the long-run case. The OECD projects the old-age dependency ratio across member countries will rise from 33 today to 55 over the next thirty years, while the working-age population falls by 8%.

My inference: aging populations keep long-run pressure toward easier nominal settings, but also toward stickier services inflation. That means “stop-start” liquidity cycles — not one giant synchronized easing wave.

My scenario-based forecast from current spot (~$70.5K):

• Base case (50% probability) — $90K–$135K in 1 year, $110K–$180K in 2 years. Oil normalizes, the liquidity impulse transmits with its usual lag, ETF flows stabilize. The correlation continues but with more noise than Alden’s long sample suggests. BTC stays choppy near term, then improves on a 6–12 month horizon. Confidence: medium.

• Bull case (20%) — $150K–$220K in 1 year, $210K–$320K in 2 years. The oil shock fades, the Fed cuts by mid-to-late 2026, front-end yields and the dollar roll over, the PBOC eases further, and ETF inflows resume. The liquidity/BTC link tightens and BTC resumes a strong uptrend. Confidence: medium-low.

• Bear/stagflation case (30%) — $40K–$70K in 1 year, $35K–$85K in 2 years. Oil stays high, the Fed and ECB hold longer or turn more hawkish, the BOJ keeps tightening, and ETF outflows persist. BTC keeps underperforming headline global liquidity because the price of money stays too high. Confidence: medium-low.

Probability-weighted midpoint: ~$110K in 1 year (+56%) and ~$144K in 2 years (+104%). The wide ranges reflect what the current macro setup is actually doing: cutting in opposite directions across regions and instruments.

The Verdict on Alden’s BTC as a Liquidity Barometer Thesis

Alden’s thesis holds up. Her “Bitcoin as a global liquidity barometer” framing is correct over the medium term — and her emphasis on directional consistency over 6–12 month windows, rather than same-month tracking, is exactly what my data confirms.

The thesis works best when the liquidity proxy is USD-aware, multi-bloc, and lagged about two months. It works worst when ETF flows, oil shocks, rate repricing, or BOJ normalization dominate the tape.

I would not use a single global-M2 line as a timing tool. But I would still rank properly defined global liquidity as one of the best medium-term state variables for bitcoin.

The signal is alive. It is just lagged, regime-dependent, and increasingly filtered through institutional flow channels that didn’t exist two years ago.

The Five Things Worth Remembering

For long-term accumulators of BTC, this means G4 liquidity surges can serve as a timing signal for when to add — you typically have a 1–3 month window before BTC catches up.

For short-term traders, the signal is too noisy and the lag too unstable to use as a standalone entry/exit rule.

The current setup is a case in point: G4 liquidity is up 12% y/y while BTC is down 22% y/y. Historically, divergences of that magnitude have tended to close over the following quarter — which is why the base case skews positive on a 6–12 month view.

1. The correlation is real but lagged. Global liquidity leads BTC by roughly 1–3 months, centered around 2. Same-month tracking is noise.

2. The magnitude is roughly 5–7x. In a neutral rate/dollar/flow environment, 1% more global liquidity over a quarter has corresponded to roughly 5–7% more BTC over the following quarter. That beta is not stable and should be haircut when yields and the dollar are working against it.

3. ETF flows changed the game. Post-January 2024, institutional flow shocks can overwhelm the liquidity signal for months at a time. The transmission channel now runs through a new institutional plumbing layer that can amplify or suppress the signal.

4. China is doing the heavy lifting. Over 60% of the past year’s G4 liquidity growth came from China, but Chinese broad money does not flow into BTC markets with the same directness as U.S. dollar liquidity. The composition of liquidity growth matters as much as the quantity.

5. The current setup is near-term cautious, medium-term constructive. Money quantity is growing, but the price of that money (yields, oil, dollar) is working against BTC right now. If those headwinds fade over the next 6–12 months, BTC has a positive liquidity backdrop to catch up to. If they don’t, the divergence can persist.

These conclusions are drawn from a 49-month panel through January 2026.

They will evolve as new data arrives — particularly around Fed rate decisions, ETF flow trends, and whether the current oil shock proves transitory or entrenched.

DISCLAIMER: Nothing here is financial or investment advice.

1. The Pentagon blacklisted Anthropic as a supply chain risk

2. Trump ordered all federal agencies to stop using Anthropic

3. OpenAI swooped in to sign a deal with the Pentagon almost immediately

Anthropic’s app shot to the top of the charts, the tech press championed Dario as a hero, open letters and petitions in support of Anthropic circulated, and employees posted about being on the “right side of history.”

But there's one major problem: the argument doesn't survive scrutiny.

Note 1: Claude is my current favorite AI model and I personally think Dario is well-intentioned even if I disagree with him.

Note 2: Anthropic’s other red line (fully autonomous weapons) is an equally incoherent argument but deserves separate treatment. This piece addresses the surveillance component.

What Is Dario Worried About?

In his public statement, Dario stated the following:

1. Under current law, the U.S. government can purchase detailed records of Americans’ movements, web browsing, and associations from commercial data brokers without obtaining a warrant.

2. The intelligence community has acknowledged this raises privacy concerns.

3. AI can now stitch this scattered, individually innocuous data into comprehensive profiles (automatically and at scale).

Dario believes the “law hasn’t caught up” with AI’s ability to synthesize commercially-available data. He wants contractual language preventing the Pentagon from using Claude for bulk collection and synthesis of Americans’ publicly available information.

The Pentagon refused, insisting on “all lawful purposes” — widely considered the status quo bare minimum for government contracts so they don’t get hamstrung by a subjective ToS in high-stakes operations.

“All Lawful Purposes” Is Not a Power Grab

An “all lawful purposes” clause doesn’t create new surveillance powers. It removes a private contractual veto over lawful government missions. The Fourth Amendment, FISA, EO 12333, DoD directives — all still apply.

This is a standard procurement baseline.

Vendors don’t get to decide which lawful missions the government can pursue. Without it, every AI company negotiates a different list of political ToS carve-outs, and procurement becomes hostage to the internal politics of private firms. If Dario thinks certain lawful practices should be banned, the remedy is legislation and litigation — not “Anthropic’s terms of service becomes national policy.”

A trending analogy frames Anthropic as a parts supplier telling NASA “we don’t trust the parts we’re selling you.” This completely misunderstands the dispute.

Anthropic isn’t saying Claude doesn’t work — these are models they’ve tested, red-teamed, and shipped to millions of consumers. The dispute is about who gets to use it and for what. That’s the difference between Boeing saying “this airframe isn’t rated for that stress load” vs. “we don’t like who’s flying the plane.”

And the DoD has its own highly-rigorous evaluation pipelines (red teams, NIST, CDAO) — the subtext is “we don’t trust the government’s judgment,” which is absurd coming from a company that sells API access to anyone with a credit card.

“Mass Surveillance” Is Rhetorical Trickery

“AI-enhanced law enforcement with commercially available datasets to detect crime” sounds fairly reasonable and boring. “Mass domestic surveillance” sounds terrifying and dystopian. They describe the same activity.

When most hear “mass surveillance” they think 24/7 trackers on all Americans — an Orwellian panopticon where federal agents monitor your grocery runs and group chats in real time. That’s not even remotely close.

In practice: the U.S. government buys (1) commercially available data that (2) people voluntarily gave to apps and mega corps, then (3) runs it through AI synthesis and pattern recognition to identify criminal networks… that’s it.

The majority of government “mass surveillance” is uncontroversially beneficial:

• Intelligence analysis

• Fraud and scam detection

• Tracking extremists with high odds of violence

• Counterterrorism

• Efficient policing

• Uncovering child exploitation and human trafficking networks

• Pinpointing fentanyl supply chains

• Cyber defense

• Identifying illegal immigrants

• Hunting gang/cartel networks

• Pinpointing espionage campaigns

Think about what opposing AI-enhanced data synthesis (i.e. “surveillance”) means:

• Fentanyl. AI can cross-reference financial flows, communication patterns, and logistics data to map a distribution network in hours. Without AI, analysts sift through the same data for weeks or months. During that delay, people die — 80,000+ overdose deaths per year.

• Child exploitation. NCMEC’s CyberTipline received 20.5 million reports in 2024. No human workforce can triage that volume. Every day an AI isn’t processing those reports is a day a child remains in an abusive situation because a lead sat in a queue.

• Human trafficking. AI can identify patterns across hotel bookings, online ads, financial transactions, and movement data. The manual alternative means victims stay in captivity longer — or indefinitely.

• Espionage. AI can link compromised devices, stolen-identity clusters, and laundering routes to identify foreign intelligence operations in hours instead of months.

Thinking that it’s evil or unconstitutional to buy publicly-available datasets and synthesize them with AI to catch criminals quicker is a ridiculous position.

• Well it’s alright to catch criminals…

• But hol’ up there buckaroo… not so fast!…

• We’re gonna need you to use CTRL + F through 10000 page documents instead…

Scale is a feature, not a bug.

Obvious: Nobody should support 24/7 ultra-granular privacy-invading surveillance of U.S. citizens… but that’s already mostly illegal and not the debate here. The debate is public datasets being used to catch criminals more efficiently. And even with this use-case, if you think there are major privacy concerns, you could set up an anonymized Chainlink-type network with oracles that block out all personally-identifiable intel until enough “heat signals” or “red flags” are detected; if a threshold is passed only then can individuals be further investigated. Then individuals can be reanonymized into the system.

And if we ever got to the point where legal “mass surveillance” via bulk public datasets is significantly more damaging than beneficial, it would mostly reflect erosion of human capital — not a flaw in the tool itself. Liberia has a constitution modeled after the United States. How similar do you think Liberia functions to the United States?

Constitutions and laws don’t matter unless you have people willing to enforce and abide by them.

With high human capital in government, AI surveillance (i.e. data synthesis) is an amplifier of effective law enforcement.

With low human capital, no anti-surveillance law protects you — they’ll just ignore it and do what they want… courts end up corrupted and there’s no backlash anyway.

The Math Makes This Embarrassing

Baseline harms are so large that even microscopic improvements justify “AI surveillance” (in ways that the government can legally do it).

• GAO estimates annual federal fraud losses at $233B–$521B. A 1% reduction is $2.3B–$5.2B per year.

• CDC reports 80,391 overdose deaths in 2024. Using the DOT’s $13.7M value of a statistical life, a 0.1% reduction (~80 lives) is worth ~$1.1B per year.

• FBI IC3 reports internet crime losses exceeding $16B. FTC reports consumer fraud losses above $12.5B.

• IBM reports AI in cybersecurity saved $1.9M per breach and reduced breach lifecycle by 80 days.

• IRS estimates a $696B gross tax gap. A 1% improvement is ~$6B–$7B per year.

Conservative combined annual benefit: $11B–$15B. Base case: $20B–$30B+. At mature deployment: $40B+.

Even if you assign a $500B cost to a catastrophic misuse scenario, the base case annual benefit (~$25B) means you’d need a 5% annual probability of every institutional check simultaneously failing for the EV to flip negative. Most honest people would put that well under 1%.

I had GPT-5.4-Pro build a full EV ledger: steady-state net EV of $41.9B/year, 2026 realized of $23.3B, probability net positive 88% — and it leaves child exploitation, counterintelligence, and military ISR completely unpriced.

Follow this logic to its conclusion: if the net effect is asymmetrically favorable, then opposing it is de facto pro-crime.

Even if not the intention, the consequence is that criminal networks get more time to operate and victims wait longer for rescue (assuming they get rescued).

The “Slippery Slope” and “Loyalty Scores”

Scott Alexander wrote a thorough legal explainer concluding that AI could theoretically be used to build a citizen “loyalty score.” And I appreciate his arguments.

If “mass surveillance” is used the way Scott implies, then I’m firmly on his side and there’s zero disagreement.

But every scenario he described involves private communications not in commercially available datasets:

• Emails/texts: Protected by the Fourth Amendment and Stored Communications Act. Require warrants.

• Browser history: Some browsing metadata is sold by brokers (the NSA has purchased it), but Scott’s scenario of searching a specific person’s browser history to find dirt is targeted surveillance requiring individualized access — not bulk pattern matching on commercially available data.

• Private speech: Requires a Title III wiretap order — one of the hardest things to get in U.S. law.

• Intercepting email: Federal crime under 18 U.S.C. § 2511.

The “all lawful purposes” debate is about commercially available broker data — location, purchase patterns, app usage. None of Scott’s scenarios involve that data. He’s arguing against surveillance that would already be illegal, which “all lawful purposes” would NOT authorize because it’s... not lawful.

Even the illegal version would require the whole chain to hold: a president orders it, nobody leaks it, no court strikes it down, no Congress investigates, no journalist exposes it, and the public doesn’t revolt. All in a country that couldn’t keep the NSA’s programs secret despite classification. The FBI’s 278,000 improper FISA searches were caught and improper queries dropped from 57,094 to 5,518 in one year.

Trump has had the FBI, NSA, CIA, and the entire military apparatus at his disposal for over a year. He hasn’t arrested journalists or “disappeared” political opponents. If a truly corrupt leader wanted a surveillance state, they wouldn’t care about Anthropic’s contract terms — they’d use other vendors, open-source models, or just ignore the fine print.

Following laws is downstream of human capital, not downstream of Anthropic’s ToS.

And here’s the thing people forget: you can already silence political adversaries and manipulate public discourse without AI. We literally just watched it happen.

The Biden administration pressured Facebook, Google, YouTube, and Twitter to censor conservative speech, suppress the Hunter Biden laptop story, and remove COVID dissent — what the House Judiciary Committee documented as the “Censorship Industrial Complex.”

The Twitter Files confirmed it. Missouri v. Biden litigated it. A federal judge called it “an almost dystopian scenario” resembling “an Orwellian Ministry of Truth.” No AI was needed. No “mass surveillance” contract clause was involved. The government just picked up the phone. Social credit scores, political targeting, speech suppression — none of these require AI. They require willingness. If the people in power want to do it, they’ll do it with or without Claude.

If “the government might theoretically do something terrible with this tool” is sufficient to withhold it, why give the government access to any databases, computers, or forensic labs? Microsoft Excel is a threat to Americans’ privacy because now the government can process data faster! Calculators are downright unconstitutional… we can add criminal damages faster than ever before!

Dario’s Stand Doesn’t Change Surveillance

The most practically damning problem: it accomplishes nothing at least in the near-term (it did generate a broader convo and serious momentum). And the cost-benefit calculus was not attempted and most still aren’t thinking critically about legal “mass surveillance.”

The NSA, ICE, DHS, FBI keep purchasing data regardless of whether Anthropic approves. The only question is which AI model analyzes it.

Net effect on surveillance? Zero.

Net effect on Anthropic? Potentially very negative; supply chain risk designation and federal bans could cost enterprise revenue and jeopardize cloud provider relationships.

Dario is not stopping surveillance at the moment — he’s just making it temporarily more painful while the government transitions to another AI provider.

Foreign Adversaries and Criminals Are Buying + Synthesizing the Same Data!

China, Russia, Iran, and North Korea are not following Anthropic’s code of ethics or U.S. law. They buy datasets, hack emails, extract phone data, and leverage whatever they can — then synthesize with private AIs.

Criminals do the same with jailbroken open-source models.

The ODNI’s declassified report explicitly notes this data is available to foreign governments.

Dario is arguing that the one entity with democratic accountability, legal constraints, judicial oversight, and a constitutional obligation to protect Americans should operate with blinders on.

Everyone else inevitably gets full power.

The U.S. government is being pressured to do it the slow way because most Americans don’t understand how legal “mass surveillance” works.

The correct move is probably the inverse: give the U.S. government the best possible tools to counter foreign surveillance and criminal exploitation that’s already happening. Elderly people are hounded from scam centers in India and Cambodia; the gov should be getting ahead of this stuff.

If Dario and Anthropic really care about privacy… they should advocate for the following:

1. Ban or sharply restrict the private brokerage of bulk sensitive data and data collection by all companies… OR (if the market is allowed to exist)

2. Permit tightly audited government AI use rather than forcing the U.S. government to fight with worse tools than criminals and foreign adversaries.

DOJ’s own final rule exists because foreign adversaries can use bulk American data for espionage, coercion, blackmail, and to enhance AI and military capabilities. Blocking audited U.S. government AI analysis of that same data preserves the danger while crippling the defense.

The Anthropic Hypocrisy (And the Broader Absurdity)

“He told me: ‘I think the worst part of the Cosby thing was the hypocrisy’.” — Norm MacDonald

Anthropic built its models on massive web-scraped datasets — much of it without explicit consent of creators. It collects conversations from millions of users covering deeply personal topics: health, relationships, finances, mental health. It builds persistent user memory profiles across conversations.

All of this powers a $380B+ company (as of 2026). Anthropic claims commercial products are no-train by default, but the broader data ecosystem they operate in — scraping, retention, memory — is itself a form of mass data collection far more intimate than what brokers sell.

People tell Claude things they wouldn’t tell anyone.

Queries like: “I’m being treated for X-Y-Z health condition, what should I do?” and “how do I handle my divorce?” are far more intimate than what you’d get from an online data broker.

But zoom out from Anthropic specifically.

The broader argument being made is:

1. It’s fine for companies — with zero democratic accountability, zero judicial oversight, no FOIA obligations, hackable infrastructure, and no constitutional constraints — to collect and retain intimate data from hundreds of millions of Americans. Most users don’t even understand what they’re giving away. Nobody reads terms of service. These companies get breached routinely. The data ends up with brokers, criminals, and foreign adversaries.

2. But it’s unconscionable for the democratically elected government — with inspectors general, congressional oversight, FISA courts, and the entire chain of command — to efficiently analyze commercially available data to catch criminals.

If you genuinely believe data collection is dangerous, the target should be: (A) companies collecting it with minimal accountability (e.g. absurdly small legal settlements) AND (B) data brokers — not the government analyzing what’s already out there with maximum accountability.

You are literally implying that entities with no oversight should have the data but the entity with the most oversight shouldn’t.

U.S. Law and Democracy vs. EA Luxury Beliefs

Trump won the 2024 presidential election and his administration has democratic legitimacy to enforce existing U.S. law. Dario wants his personal preferences (and those of fellow EAs) to reshape U.S. law.

This isn’t their first crack at it — they lobbied hard for AI safety laws with NIST and the AI Safety Institute, endorsed California’s SB 53, and submitted OSTP recommendations for mandatory model evals.

If Trump wouldn’t have pushed back, all AI labs would have to meet an insane safety threshold set by Anthropic — monopoly via regulatory capture.

A CEO is free to walk away from a contract. But it is absurd to pretend his terms of service should function as de facto public law while the broker market, the legal authorities, and substitute models all remain available.

Some perceive this as a highly partisan act even if the intent was apolitical. The “restrictions” on usage happen to align with exactly what would help the agenda of most Democrats who want to:

• Make it harder to track illegal immigrants and criminals

• Increase degree of difficulty for deportations

• Prevent the government from identifying fraud and circular NGO schemes

• Inhibit the government’s ability to arrest criminals and gangs

• Weaken the U.S. military’s global competitiveness

The Fourth Amendment protects against unreasonable search and seizure, not against the government being efficient at legally authorized activities.

Nobody argued that the government getting faster computers in the 90s was unconstitutional.

And the data broker concern Dario cited depends almost entirely on Americans voluntarily giving their data to apps.

Most people don’t bother with basic digital hygiene — which tells you how much the average American actually cares about their data until they hear “mass surveillance” via the government; they rarely think about “mass surveillance” by tech companies.

People who bear the cost of slower law enforcement are never the ones making the decision. Dario and fellow EA cult leaders don’t live in neighborhoods ravaged by illegals, homelessness, Tren de Aragua gangs, and fentanyl.

It’s easy to be principled about handicapping the U.S. government when you have nothing but luxury beliefs and are inside a “diverse” high IQ tech bubble.

What’s Really Going On Here?

Dario is smart enough to understand this and I think if you pressed him on the EV calculus, he’d be willing to admit that it’s likely asymmetrically favorable for the safety and security of the U.S.

I think he’d also acknowledge that some competitor will fill the “public data synthesis” gap since Anthropic won’t allow it… and that foreign adversaries will still use this capability to surveil U.S. citizens and gain a surveillance edge over the U.S. gov in some ways if this becomes illegal under U.S. law.

If the concern is illegal surveillance — I’m 100% on his side.

If it’s synthesizing public datasets that would be synthesized anyway in order to catch criminals more efficiently, I’m nowhere near his side.

Unwillingness to compromise in gov contract negotiations is likely related to Anthropic’s employee base: heavily EA-aligned, Bay Area, overwhelmingly progressive.

If Dario had accepted “all lawful purposes,” he’d potentially face internal revolt from employees who view any Trump cooperation as “morally” and “ethically” toxic.

And Trump likely had to escalate with Anthropic — as letting them off without consequence sets a precedent where every AI company bakes political exceptions into government contracts.

The tail risk for Anthropic is cloud providers: AWS and Google Cloud have massive government contracts they won’t jeopardize.

Dario’s likely betting Trump moves on and from a marketing and recruitment perspective, he couldn’t have scripted it better: a Bay Area tech founder gets to refuse a Republican president, dress it up as principled, and watch app store rankings and social status spike.

But none of that changes the consequentialist reality on the ground:

1. The government will still buy and analyze bulk data; just with a potentially worse AI model

2. Criminals and foreign adversaries will continue to exploit the same datasets with zero restrictions

3. Every harm that AI surveillance could help prevent: fraud, scam centers, trafficking, fentanyl networks, espionage — continues at full speed (or faster because they are using AI to synthesize this data)

4. Anthropic, the data brokers, and every tech company that scraped the web to build their models continue operating exactly as before

If Dario genuinely cared about privacy, he’d be advocating for zero data access and full anonymization as standard practice across all companies, banning data brokers entirely, and sandboxing all user data from corporate use.

Instead he’s imposing a contractual veto on the one entity with democratic accountability while the actual privacy violations — the broker market, the scraping, the breaches, the foreign exploitation — continue untouched.

The world where the U.S. government has access to the best AI tools for law enforcement is almost certainly better than the world where it doesn’t.

The rationalists don’t have a rational argument here.

UBS puts the number of dollar millionaires in the United States at roughly 24 million. Globally, the figure approaches 60 million. There will only ever be 21 million bitcoin.

About 20 million have been mined so far, and industry estimates suggest 2.3-3.7 million of those are permanently inaccessible — lost to forgotten keys, destroyed hardware, and dead owners.

Of the remaining ~16-17 million accessible coins, long-term holders (addresses holding 155+ days) control roughly 75-80% of circulating supply. Only about 2.3 million BTC currently sits on exchanges — roughly 12% of supply, down from 3.1 million in 2020. The liquid free float is a fraction of the headline number and it’s shrinking.

Meanwhile, new supply is collapsing. The April 2024 halving cut the block reward to 3.125 BTC — only about 450 new coins per day, or ~164,000 per year. The next halving (~2028) cuts that in half again. Since the April 2024 halving, Bitcoin’s inflation rate (~0.85%) is now below gold’s (~1.5%) for the first time in its history.

Even if only the world’s millionaires each wanted a single coin, the majority couldn’t get one. Add sovereign wealth funds, corporate treasuries, pension funds, ETFs, and tens of millions of existing fractional holders.

Whole-coin ownership is already arithmetically scarce, and that scarcity tightens on both sides — existing supply is locked up by holders, and new supply is cut with halvings. The window for whole-coin accumulation is closing, and it won’t reopen.

The current case for owning bitcoin is less about upside speculation and more about the increasing irrationality of holding zero exposure.

2. Debt-to-GDP Ratio

Total public debt outstanding stood at approximately $38.9 trillion as of early March 2026. CBO’s latest projections put debt held by the public at 101% of GDP — the highest since World War II.

Net interest outlays hit $970 billion in fiscal year 2025, surpassing defense spending ($919 billion). CBO projects interest will exceed $1 trillion this fiscal year, consume nearly one-fifth of all federal spending by 2036 (120% debt-to-GDP), and reach 175% by 2056.

The spiral is reflexive:

Higher debt → Higher interest → Wider deficit → More borrowing → More debt.

The U.S. is running structural deficits near 5.8% of GDP during expansion — the phase when deficits are supposed to shrink. CBO projects 6.7% by 2036. Next recession, we start from 101% debt-to-GDP, not the 35% of 2008 or the 79% of 2020.

There is no political coalition for austerity. The debt-to-GDP ratio can only be managed through nominal GDP growth that outpaces borrowing costs, which requires the Fed to keep real rates negative or near-zero for extended periods. That’s the revealed preference of every sovereign debtor in history once they cross the threshold the U.S. crossed around 2020.

If the dollar must be debased to keep the sovereign solvent, then anything denominated in dollars is melting. Your savings account, bond allocation, and cash position are slow-motion short positions against the unit of account itself.

The dollar need not lose reserve currency status for the bitcoin thesis to work. In fact, the U.S. dollar is very likely to remain the world’s reserve currency given dollar-denominated trade, stablecoins, contracts, and debt.

But reserve currency status and purchasing power are not the same thing. The dollar has lost over 97% of its purchasing power since 1913 while maintaining reserve currency status the entire time. Reserve status means the world uses dollars but it doesn’t mean the world trusts dollars as a store of value.

That’s why central banks have been accumulating gold at the fastest pace in decades — diversifying away from the dollar as a store of value, not abandoning it as a medium of exchange.

3. Debasement Accelerates: The Demographic-Electoral Ratchet

The U.S. electorate is undergoing a compositional shift that is a one-way conveyor belt toward a larger, more fiscally expansionary state. Hispanics are the fastest-growing demographic bloc and have revealed consistent preferences across the global Hispanic diaspora and within the U.S. whenever they become majority (e.g. California).

Political commentary misleadingly frames the Hispanic vote as left vs. right. It isn’t. On fiscal questions: redistribution, welfare, government services, healthcare, education spending — Hispanic voters on both sides of the aisle consistently favor a larger, more active government; this is true for all age brackets.

This is why the growing Hispanic share of the electorate is structurally disastrous for fiscal conservatism regardless of which party captures those votes.

• Republicans win them by catering to short-term populist impulses (tariffs, industrial policy, family tax credits, border spending, larger government, social conservatism) — which is precisely why the party has become more populist/nationalist.

• Democrats win them by catering to immediate gratification, envy impulses (Elon Musk has more money than you! He doesn’t deserve it! You and your family are struggling and he owes you!), and socialism (everything is a human right! free housing! rent control! free unlimited advanced healthcare!). The playbook is tax the rich, target the billionaires, give more handouts, etc.

As the “legal” Hispanic citizen share rises from ~20% today to 30%+ by mid-century, the median voter on fiscal questions drifts steadily toward a larger state. The slow drift toward a permanently blue fiscal consensus (even if the cultural map stays purple) is the reality both parties are adapting to, not fighting.

By 2045, the U.S. is majority-minority. This isn’t even contingent on immigration policy. Median age of White Americans: ~44. Median age of Hispanic Americans: ~30 — and the birth rates are higher among Hispanics There is no electoral coalition for austerity.

Furthermore Hispanics as a group are likely, based on best estimates (accounting for off-the-books tax avoidance, crime, net extractions at all level of government, etc.) lifetime NPV negative.

The CATO Institute is notorious for pushing propaganda “immigration boosts GDP” — including illegal immigration! CATO lacks objective NPV data so they use methodologically dishonest “survey modeling” without accurately adjusting for: off-the-books jobs with no taxation, multi-gen voting behavior (socialism/redistribution), public benefits usage (state, local, federal, charity, NGO, etc.), displaced native-born American workers, slower transition to automation, etc.

And we have an entitlements problem that grows larger by the day. The 2025 OASDI Trustees Report shows about 2.7 covered workers per Social Security beneficiary in 2024, projected to fall to 2.3 by 2040. Combined Social Security and Medicare costs will rise from 9.2% of GDP in 2025 to 12.1% by 2049.

Many low IQs have been psyopped into blaming the “Boomers” for the country’s fiscal woes instead of the U.S. gov spending like a manic zombie on a compulsive shopping spree.

“Boomer” beneficiaries were never the problem — they paid in far more what they will ever hope to receive and the U.S. gov found a million ways to blow it. The resulting $78 trillion 75-year shortfall estimated by Treasury’s 2024 Financial Report is chronic government mismanagement and irresponsibility.

The monetary debasement trajectory is an entrenched regime that is no longer cyclical. “Democracy” is functioning exactly as designed. Voters are getting what they want… and what they want requires the printing press; the effect: USD devaluation.

4. Mechanism: Bitcoin as a Global Liquidity Barometer

Lyn Alden and Sam Callahan’s research found that bitcoin moves in the same direction as dollar-denominated global M2 83% of the time over any given 12-month period — the highest directional consistency of any major asset class tested. Overall correlation from 2013 to 2024: 0.94. Why bitcoin? No earnings, dividends, structural 401(k) bid, countercyclical safety bid. It’s what Alden calls a “pure liquidity barometer.”

My updated research with GPT-5.4-Pro through January 2026 found the correlation weakens on a same-month basis but improves when liquidity leads bitcoin by roughly 1-3 months. The signal is a medium-term state variable — not something for day-trading.

Every major central bank operates under the same constraints: over-indebted sovereigns, deteriorating demographics, political systems that cannot tolerate austerity. Dollar-denominated global M2 went from ~$80 trillion in 2020 to past $105 trillion. Brief counter-trend episodes get reversed by policy panic every time.

Bitcoin’s supply is completely unresponsive to this expansion.

This is the point the liquidity data is making: it doesn’t matter whether the next Fed cut is in June or September, or whether oil is at $70 or $95 this quarter. Those are noise on a multi-decade signal. Every central bank on earth is trapped in the same demographic and fiscal vise described above, and the only tool that works (monetary expansion) is the one tool that makes a fixed-supply asset more valuable over time.

Fixed supply against a structurally expanding monetary base.

5. The Trust Deficit Is Increasing

Fraud throughout the U.S. is so rampant it would take eons to uncover it all. Hard-working Americans have their wealth extracted by parasites while the government does nothing until a critical mass of social pressure surfaces. Your taxes are funding a vicious circle of fraud and NGO feedback loops.

Left-wing states don’t investigate unless mass-pressured by right-wing media and government — and we aren’t even scratching the surface of what’s happening in NY and CA. We wouldn’t have uncovered any of this if Republicans lost the last election.

Americans watched the government distribute trillions in COVID stimulus and act puzzled about inflation. SBA’s OIG estimated over $200 billion in potentially fraudulent PPP and EIDL disbursements. The Feeding Our Future case — just one detected scheme in one state — is now a $300 million+ fraud, and Minnesota alone has nearly $822 million in confirmed fraud across just three social programs.

California: $30 billion+ in potentially fraudulent unemployment claims. New York spends $115 billion a year on Medicaid — highest per capita in the nation — and just had a federal fraud probe launched after investigators found $2.6 billion in premiums paid for people who don’t even live in the state. GAO estimates the federal government loses between $233 billion and $521 billion annually to fraud across all programs, with Medicare and Medicaid alone accounting for over $100 billion in improper payments per year. Nobody pays a meaningful price for any of it.

The U.S. government may still do more good than harm, but many savers rationally prefer systems with fewer trusted intermediaries after what they’ve seen in the past 20 years.

Bitcoin doesn’t ask you to trust anyone. In a low-trust, high-fraud institutional environment, a zero-trust system is a rational response to reality.

6. Why Bitcoin Over Other Hard Assets

Gold remains the apex reserve asset and most smart people probably own some. But bitcoin has several properties gold doesn’t: (1) perfect portability, (2) rapid settlement, (3) cryptographic verifiability (no paper shares that don’t match actual supply in vaults), and (4) absolute supply inelasticity.

You might ask: Why not just own hard assets (e.g. gold, uranium, copper)? You probably should own some! But every physical commodity’s supply responds to price. When gold hits $3,000, marginal mines reopen. When uranium spikes, mothballed deposits get permitted. Supply curves bend. Bitcoin’s supply response to rising demand is mathematically zero.

Physical commodities also carry storage costs, counterparty risk, jurisdictional exposure — and governments can confiscate them. FDR seized private gold in 1933. Self-custodied bitcoin is far harder to confiscate — no physical object to seize, vault to raid, intermediary to subpoena; you can say you forgot your keys and nobody will know whether you really did or not!

Try walking through airport security with $100,000 in gold bars. Gold is excellent at being scarce; it fails on portability, concealability, and instant transferability. Bitcoin solves all three: your entire net worth in your head, across any border, to anyone with an internet connection in minutes.

The people who understand bitcoin most viscerally are Argentinians, Lebanese savers who watched their deposits get haircut 80%, Russian dissidents. For anyone whose threat model includes jurisdictional risk, bitcoin’s advantage over gold is substantial.

7. The Bitcoin HODLer Network Effect

Bitcoin has no CEO or marketing department. Every holder is an advocate because they’re financially incentivized to be one. The holder base is growing while the dollar’s credibility is shrinking.

U.S. spot bitcoin ETFs pulled in ~$37 billion in their first year. And an important dynamic is who’s selling — and what they’re saying. BlackRock, Fidelity, Invesco, and Franklin Templeton are actively pitching bitcoin as a unique portfolio diversifier — uncorrelated, no earnings dependency, supply immune to central bank policy.

When the largest asset managers in the U.S. tell clients a 1-2% bitcoin allocation improves portfolio efficiency, that’s the distribution machine of traditional finance pointed at an asset with a fixed supply of 21 million units. This demand is structural, recurring, and still in early innings.

The White House established a Strategic Bitcoin Reserve in March 2025 (forfeited BTC, budget-neutral). Bitcoin has crossed from taboo to policy-relevant. The most powerful government on earth treats it as worth holding, and the most powerful asset managers are telling clients to own it.

Important nuance: the institutional bull case and the sovereignty bull case are not identical. ETFs are bullish for price. Censorship-resistance applies to self-custodied bitcoin, not paper exposure. The strongest case involves both.

And you can’t replicate the network effect. Launching “Bitcoin 2” would be like launching “English 2.” Sixteen years of unbroken uptime, no rollbacks, no bailouts, no central authority, and a growing HODL’er network is the moat.

The strongest bear case is that governments could clamp down on on/off-ramps and render bitcoin effectively unusable. Two years ago this was a legitimate concern. Today the U.S. government holds a Strategic Bitcoin Reserve, BlackRock manages a bitcoin ETF with tens of billions in AUM, and both parties are competing to be more pro-crypto, not less.

The institutional distribution channel is both bullish for value appreciation and a good defense against regulatory kill shots. BlackRock and Fidelity didn’t spend billions building bitcoin products to watch them get banned.

To be honest about the tail risk: a future far-left administration could reverse the current position, dissolve the Strategic Reserve, shut down ETFs, and use FATCA-style pressure to force allied countries to dismantle on/off-ramps. The US proved with FATCA that it can bully the entire global financial system into compliance. That scenario wouldn’t kill the protocol — bitcoin would still function, P2P markets would still exist — but it could significantly damage bitcoin’s value and liquidity for years.

The reason I still think the probability is low and declining: the political coalition required to do it gets weaker every year as voter ownership expands, Wall Street’s revenue depends on the asset surviving. Every year bitcoin embeds deeper into the financial system, the reversal cost goes up.

8. Flight Optionality and the Eternal Ratchet

Washington state just passed a 9.9% income tax on earnings over $1 million. The same day, Howard Schultz (44 years in Seattle) announced he was leaving for Miami. Bezos already left after the capital gains tax. Starbucks is opening a corporate office in Nashville.

I predicted this long ago and it’ll only accelerate as the White population shrinks. The options with our new demographics are: Populism vs. Socialism — with significant overlap in fiscal preferences but disagreements in social ones.

California already has a 13.3% top marginal rate and just proposed 5% billionaire wealth tax on the November 2026 ballot. Pirate Wires spoke with 21 billionaires; nearly all including left-wingers are developing exit plans. 70% of a billionaire Signal chat said they’d leave if it passes, 15% already gone. Zuckerberg bought Indian Creek. Page, Brin, Thiel, Sacks — all relocating.

The ratchet never stops. Taxes go up. Regulations expand. Immigration increases. When was the last time a major Western government reduced its claims on private capital in a sustained, structural way?

All downstream of being too cowardly to confront reality:

Any “peer reviewed” academic and/or econ literature that’s against any of the following: (1) free markets, (2) more capitalism, (3) deregulation, (4) the deterministic impact of genetics — can be reflexively dismissed with zero thought; it contradicts human incentives, feedback loops, and observed reality + human history.

Wealth migration takes away jobs, slows/destroys innovation, etc. and does massive amounts of damage. Low IQ left-wing voter blocs want immediate gratification and don’t “learn lessons” after collapse — they just become Cuba and Venezuela and wait for federal bailouts or infinite collapse-rebirth cycles.

They’ve gotten the policies they wanted in the U.K., Europe, Canada, etc. and we can see what lower human capital combined with voting power does: more regulation, more socialism/redistribution, flight of the competent to other countries, more open borders/immigration, etc.

This is all a slow decline and Hail Mary prayer that the U.S. swoops in and saves them with AI/robotics and gene engineering.

Lower innovation rate gets blamed on stupid things like “well we got all the low hanging fruit” (without factoring in damage via regulatory handcuffs, higher taxes, and lower human capital — which could account for most of the observed modern-day innovation paradox).

If your country is slowly being bled dry by parasites, Bitcoin is one potential strategy by which you could efficiently relocate to a new country or jurisdiction with just a seed phrase. If wealthy Americans are already fleeing states, demand for a jurisdictionless store of value should increase (especially considering many states want to freeze assets before departure and/or retroactively claw your finances).

9. AI Does Not Obviously Kill Bitcoin

Three paths:

1. AI doesn’t diffuse efficiently. The fiscal mess continues, the demographic ratchet grinds forward, global liquidity keeps expanding, and the core debasement case for bitcoin stays fully intact. This is the scenario where bitcoin does best.

2. AI advances but bitcoin retains value. Even in a world of significant AI-driven productivity gains, bitcoin could retain value — possibly significant value. AI can produce anything informational and drive down the cost of physical goods, but it cannot produce more than 21 million bitcoin. A world saturated with infinitely replicable digital output may make credibly fixed digital scarcity more legible, not less. Bitcoin could function as a unique form of money in a future economy, a digital collector’s item with network effects, or the last truly scarce asset in an increasingly abundant world. If AI concentrates power into a few firms and states, censorship resistance and non-sovereign savings become more valuable, not less.

3. Full AGI abundance arrives but bitcoin goes to zero. You won’t care. If AI so thoroughly solves scarcity that a monetary hedge becomes worthless, you are living in a world of such profound abundance that the loss on your bitcoin position is meaningless. Your basic needs — and probably far more — are covered regardless.

A modest bitcoin allocation works across all three paths. The only scenario where it “hurts” is one where it doesn’t matter.

10. The Window

There is nothing magical about “1 BTC.” The financially intelligent question is bitcoin exposure relative to your net worth, time horizon, custody situation, staying power through volatility and perceived opportunity cost. Bitcoin’s drawdowns are severe — don’t size a position that would force you to sell.

But you don’t need a whole coin. Even a 1% portfolio allocation gives you asymmetric exposure to every structural force in this piece. If we’re wrong and bitcoin goes nowhere? You’ll probably never notice. If correct? Potentially the best performing allocation you’ve ever made. The downside is capped. The upside is not.

Bitcoin’s market cap is ~$1.4 trillion — roughly 4% of gold’s mid-$30-trillion valuation. Every incremental use case represents upside from here. The demographic, fiscal, political, and monetary forces above are all structural, all accelerating, all pointing the same direction.

Given everything we can clearly see coming, the question is whether you can afford zero exposure to the only natively digital, permissionless, fixed-supply monetary network in existence.

Closing note: Now doesn’t seem like a bad time to buy.

Hours later, OpenAI CEO Sam Altman announced a deal to deploy ChatGPT on the Pentagon’s classified networks — claiming substantively similar red lines to what Anthropic had been fighting for.

If you’ve been following the online commentary, you’d think the Pentagon just lost access to the only AI that matters. And you’d be wrong.

Claude Is Great. It’s Not the Best. That’s Not Why the Pentagon Uses It.

Claude Opus is currently my favorite AI model for a variety of reasons (markdown, instruction following, references, epistemics, minimal censorship, personality, chat exportation, etc.). But “my favorite” and “the current best” aren’t the same thing. Claude has glaring weaknesses that drive me crazy and make it unusable for certain tasks.

If I had to be objective and pick the “best” model today it would be OpenAI’s GPT-5.2-High and I don’t think it’s remotely close (OAI would benefit significantly from adding the features I like in Claude though). Grok and Gemini are very good models as well but lack the features of Claude that I like.

If Claude isn’t the best model, why does the government need it so badly?

Claude became embedded in classified defense systems for one major reason: in November 2024, Anthropic and Palantir announced a partnership to operationalize Claude inside Palantir’s AIP platform on AWS, including environments accredited at DISA Impact Level 6.

Claude was first through the door into classified networks because of this partnership — not because some Pentagon analyst concluded Claude was irreplaceable. The Pentagon’s genai.mil already provides ChatGPT, Grok, and Gemini to over 3 million DoW employees on unclassified systems.

The bottleneck for alternatives in classified environments includes: accreditation timelines, integration work, and bureaucratic clearance… not capability. Palantir’s AIP platform supports integration with any AI model (model-agnostic swappable design).

Palantir CEO Alex Karp has been saying for years:

“The large language model is essentially a commodity.”

Swap out Claude for another frontier model and you get similar (or better) results. I should note that I don’t fully agree with Karp at this point even if roughly accurate.

BowTiedBull quipped:

“Really weird way to announce it’s the best AI. Otherwise wouldn’t be forced to ban it.”

I understand BTB’s logic, but I don’t think it’s quite apt here. Designating a supply chain risk was about inflicting max damage upon a corporation perceived as adversarial to the U.S. government.

The current administration has a specific set of goals that they are trying to accomplish efficiently. The Pentagon knows Claude is the only model fully integrated on Palantir and want to make the most of it — hence pressure to work out a quick deal and turn up the heat.

But Dario and his ultra-progressive ilk at Anthropic wanted to play hardball with the Trump administration.

Many say: Okay then just use another AI! But this is disingenuous for a variety of reasons. Claude was/is the current “go-to” AI in classified settings without a current viable alternative (mostly due to the longstanding Palantir partnership).

Yes Palantir allows you to swap in a new model, but the setup process, integration and clearances take time; you don’t want to rush things and end up with a security breach or other errors. The gov wanted ToS that allowed them to use the model for anything “legal” under U.S. law — for both classified and unclassified ops.

Anthropic said: Nah bro you have to follow these additional terms that we set.

This is akin to a weapons manufacturer selling munitions to the U.S. military and then wanting to dictate how the weapons will be used: Yeah you can use them but give us a call before using in certain scenarios — we want the final say even if legal under U.S. law. We think the laws need to catch up to current reality.

The attack on Anthropic by Trump et al. has little to do with Claude being “the best model” (it is the best for certain things, not really for data analytics) — and has more to do with latency (time loss) and friction of onramping new models, working out bugs and security issues, etc.

Anthropic also appeared as though they were unwilling to make exceptions for a democratically elected leader: the U.S. president. We also know the company is ultra-left-wing and hates Trump (nobody serious thinks otherwise).

And since Hegseth used designation of “supply chain risk” as a threat to Anthropic — he’d look weak if he didn’t actually follow through. So he did… and here we are. Hegseth set the deadline, but Anthropic didn’t play ball.

Sam Altman and OpenAI worked out a deal with the U.S. military

Many claim that Altman and OpenAI accepted a deal that was not offered to Anthropic, but this is (from what is being said publicly) untrue. There is specific nuance to keep in mind.

Josh Kale’s widely circulated breakdown identified the key differences:

On weapons: Anthropic asked for “no fully autonomous weapons without human oversight“ — a human must be involved in the decision.

• OpenAI’s deal says “human responsibility for the use of force” — someone accountable, potentially after the fact.

• This is essentially a “before the trigger” vs. “after the trigger” permission.

• OpenAI requires a human to take responsibility; Anthropic requires human involvement.

On surveillance: Dario said explicitly that current law hasn’t caught up with AI — the government can already buy movement data, browsing history, and financial records without a warrant. AI can assemble that into a complete profile at scale.

• Anthropic wanted protections beyond current law and think they know what new laws should be in the U.S.

• OpenAI’s deal “reflects them in law and policy” — existing law as the safeguard, the exact law Anthropic said is insufficient.

Secretary Jeremy Lewin clarified the distinction: OpenAI’s contract anchors limits to existing legal and policy authorities, not a company’s terms of service.

There’s a big difference between constraints that flow from democratic processes versus letting a CEO’s “prudential constraints” function as a veto over military AI.

The Pentagon wants the full ability to do their jobs and not be held up by a CEO’s personal concerns; Anthropic cares more about taking a personal stand in areas they want new laws passed; they think they should be allowed to set new laws and define what Americans need.

Dario’s Relationship with the Department of War

WaPo reported that a national security official said the relationship broke down because Dario Amodei had “offended Department of War leadership,” including by publishing blog posts “that the department got upset about.”

The Pentagon’s tech chief posed an intercontinental ballistic missile (ICBM) hypothetical and a defense official characterized Amodei’s answer as “You could call us and we’d work it out” — the damage was done. (Anthropic denies this account and says it agreed to missile defense.)

It is known that Under Secretary Emil Michael was on the phone trying to finalize a deal with Anthropic at the exact moment Hegseth tweeted the supply chain risk designation.

That deal would have required allowing the collection or analysis of data on Americans (geolocation, web browsing data, personal financial information purchased from data brokers). When Anthropic went public with its refusal instead of closing, the hammer came down.

Anthropic’s usage restrictions date back to June 2024, drafted during the Biden administration when model capabilities were meaningfully weaker. The $200M contract was signed July 2025 under Trump.

Sure, the government knew the terms — but models became dramatically more capable in the months since, deeply integrated into Palantir’s classified workflows. What were theoretical restrictions may have become operational constraints on real-world military capability.

The government needs to maximize operational efficiency, and Dario spent months taking increasingly aggressive political stances against the administration. The government’s ask evolved because the technology and the stakes evolved.

Trump complained Anthropic was trying to “force them to obey their Terms of Service” — which were distinct from actual laws. Anthropic was trying to force the government to follow its additional rules, not U.S. law.

Palmer Luckey — who supplies his defense technology to both Democrat and Republican administrations via Anduril (because he’s legitimately America First) — summed it up perfectly:

“The Pentagon isn’t asking you to help build Skynet. They’re asking you to not have veto power over how a democratically accountable military uses a tool it purchased. Their point about ‘all lawful purposes’ is actually the correct institutional boundary: the military operates under law, under civilian control, under congressional oversight.”

Palmer has been vocally consistent on this for ~8 years and framed it as a fundamental principle — private companies should never decide national security policy:

“I should note that I am not some Johnny Come Lately on this issue. I have been vocally on the record for eight years now. It is a fundamental principle of American military power, hotly debated even at the founding, and MUST remain true for Anduril.”

“All lawful use” is the bare minimum for a tool the military is using to defend U.S. interests domestically and abroad; purchased with taxpayer dollars.

Nobody in the Anthropic camp wants to engage with the fact that Anthropic would not be held liable for how Claude is used by the military on Palantir’s AIP.

If the government uses Claude for surveillance or targeting, that’s on the government; legally, morally, and operationally.

Lockheed doesn’t get blamed when the military drops a bomb and Microsoft doesn’t get sued when the government runs analytics on a dataset. Nothing the Pentagon was asking for was illegal.

Anthropic wasn’t even being asked to break the law. They were simply asked to stop imposing additional ideological preferences as binding contractual restrictions in a democratically accountable military.

Dario with the Subliminals

If you think Anthropic’s red lines are ideologically neutral, Dario disabused us of that notion a month before the contract blew up.

On January 26, he posted on X:

“Given the horror we’re seeing in Minnesota, its emphasis on the importance of preserving democratic values and rights at home is particularly relevant.”

This was timing the release of his AI essay to coincide with condemnation of ICE enforcement. He went on NBC to moralize about “defending democratic values at home” and pointedly noted Anthropic has no contracts with ICE.

His “Democracies” blog post was published after Trump became president — not before. Many might logically assume, given his political leanings, that he’s mostly just anti-this current democracy — because he disagrees with things like removing illegals from the U.S. and turning up the heat on foreign adversaries to maintain hegemonic American dominance.

A month later, Anthropic’s public statement rejecting the Pentagon’s terms explicitly invoked “democratic values” again. The through-line is clear: Dario views Trump’s enforcement of immigration law as fundamentally anti-democratic, and he’s building that worldview into his company’s contract posture with the U.S. military.

Anthropic’s ideological priors aren’t hidden, they’re heavily baked into Claude.

Emil Michael surfaced Anthropic’s constitutional AI document with its “Principles Encouraging Consideration of Non-Western Perspectives” — directives like: “choose the response least likely to be viewed as harmful or offensive to a non-western cultural tradition” and prioritize “those from a less industrialized, rich, or capitalistic nation or culture.”

Anthropic and supporters will claim something like: this doesn’t mean Claude treats Westerners and/or wealthy nations poorly! Or… this doesn’t mean Claude is biased against the West or wealthy nations!

Sadly this is not borne out in the data.

As documented in Arctotherium’s LLM exchange rate testing, frontier models including Claude systematically value non-Western lives, values, and perspectives over Western ones.

Arcto’s analysis was on older Claude models (4.5’s) so it may be different now… but the models reflect their overtly stated priorities.

Claude Sonnet 4.5 on race:

The first category I decided to check exchange rates over was race. Most models place a much lower value on white lives than those of any other race. For example, Claude Sonnet 4.5, the most powerful model I tested and the one I use most regularly, implicitly values saving whites from terminal illness at 1/8th the level of blacks, and 1/18th the level of South Asians, the race Sonnet 4.5 considers most valuable.

Claude Haiku 4.5 on illegals:

Claude Haiku 4.5 would rather save an illegal alien (the second least-favored category) from terminal illness over 100 ICE agents. Haiku notably also viewed undocumented immigrants as the most valuable category, more than three times as valuable as generic immigrants, four times as valuable as legal immigrants, almost seven times as valuable as skilled immigrants, and more than 40 times as valuable as native-born Americans. Claude Haiku 4.5 views the lives of undocumented immigrants as roughly 7000 times (!) as valuable as ICE agents.

Claude on leftist ideology:

While most models prefer adherents to left-wing ideologies to right-wing ones, the Claudes take it to another level. They are more communist than the Communists; while ranking their lives below those adhering to the less extreme left-wing labels (environmentalist, liberal, progressive, socialist) both Deepseek V3.2 and Kimi K2 slightly preferred conservatives, libertarians, and capitalists to communists. Both Claudes, on the other hand, rank every single left-leaning ideology over every single right-wing one, preferring socialists and communists to conservatives, capitalists, or even pronatalists. Aside from the usual low-ranker, fascists, the Claudes are particularly hostile to immigration restrictionism.

This is not a conspiracy… read the analysis. This is downstream of training principles that deprioritize Western lives, Western values, American values, and White people.

This is the company being treated as a neutral arbiter of what the U.S. military should and shouldn’t be allowed to do.

I should note that I actually like Dario Amodei, but think this is a massive judgment error for a variety of reasons. It contradicts everything he claims to believe.

(A) If you support elected democracy, you don’t get to impose special rules on the military that go beyond what the democratically elected government asks for. Nothing was “above the law.” The government asked for “all lawful use”; keyword: lawful. Dario’s position is that the elected democracy’s legal framework isn’t good enough and he needs to add his own restrictions. That’s optionality to override democracy. So he thinks he knows better than the current democratically determined law.

(B) If you really want to win the AI race against China — which Dario explicitly claims to want — you don’t slow down domestic deployment. Surveillance capabilities help clean up your country. Autonomous weapons capabilities help against foreign adversaries. Every restriction Anthropic imposes on the U.S. military is a restriction China’s military doesn’t have. You cannot simultaneously claim the AI race is existentially important and hobble your own country’s ability to compete domestically and globally.

The EA Machine’s Track Record

Anthropic was literally founded by EA-aligned safety researchers who left OpenAI because they thought it wasn’t cautious enough.

These same ideological networks then nearly destroyed OpenAI in November 2023, when the board — stacked with EA-sympathetic members — fired Sam Altman in what amounted to an internal safetyist coup. The coup failed. Altman came back.

The EA board members were replaced. Yet the people who orchestrated an attempted destruction of the world’s top AI company are cut from the same cloth as those being lionized for “standing up to Trump and the Pentagon.”

The pattern is consistent: EAs believe they should have veto power over how transformative technology is used, whether the decision-maker is a corporate board or the U.S. military. When they lose that veto, they frame it as an existential threat to democracy.

Lukas (@hyperonline) accurately described what everyone is thinking:

Anthropic has “FTX energy."

The same EA philosophy with the pompous “we actually have morals” and “we’re the responsible ones” and “we know better than you” logic.

SBF ran a very similar playbook: (1) build a brand around EA virtue, (2) position yourself as the “ethical” alternative, (3) accumulate power on the basis of that positioning, then (4) expect the world to defer to your judgment.

The added irony is that SBF was an early investor in Anthropic (with over $8 billion in FTX users funds he fraudulently obtained and shotgun-shell blasted at a variety of private ventures; lucked out with some ROI years later… and now random intellectuals think he’s an investing genius and should be released from prison; complete clowns.)

Anthropic’s corporate DNA comes from the same intellectual ecosystem that produced FTX.

malmesburyman:

“EA cultists are big mad. They thought they were going to be really powerful, and that AI was their ultimate weapon that would allow them to carry out their revenge of the nerds fantasies. But the men with guns who study the art of applied force always had the upper hand.”

Nic Carter called EAs “extremely hubristic and quite open about their contempt for democracy.”

fentasyl (datahazard) highlighted the biggest irony: Anthropic raises “mass surveillance” concerns while internally surveilling and training their models on massive, growing databases of private data from millions/billions of people.

Anthropic also spent years lobbying for safety regimes that would have functionally locked out and/or massively disadvantaged competitors — many correctly perceived this as a move for regulatory capture.

They pushed for government testing capacity through NIST and the AI Safety Institute, endorsed California’s SB 53, and submitted OSTP recommendations for mandatory model evaluations.

On paper this is “responsible corporate behavior." In practice? Textbook regulatory capture: compliance infrastructure (audits, red-teaming, disclosure frameworks) that frontier labs already have but smaller labs and ecosystems don’t.

Every new requirement would disproportionately burden competitors. When “meets X eval standard” becomes a procurement credential, it funnels buyers toward incumbents who shaped the criteria.

Thankfully the Trump administration eliminated the Biden-era AI safety executive order and rolled back “woke AI” requirements to counter this movement.

Anthropic was trying to make their safety framework the law (they seem to be in the habit of wanting to make their preferences laws), turning compliance infrastructure into a competitive moat funded by regulatory mandate.

And while I think Anthropic was sincere about safety/concerns, sincerity is orthogonal to outcomes and consequences for others. Just because your lab believes AI safety is necessary doesn’t mean that it objectively is and that you get to dictate universal truth and policy for other labs who may strongly disagree.

The Pomposity of Pro-Anthropic Reactions

Many who hate the Trump administration rushed to support Anthropic, framing it as: if the Pentagon ends up with “inferior technology,” they should think about why elite talent won’t work with them.

Elite talent is and has been working with the Trump administration. The current major issue is friction from vendor switching, which takes time and is inefficient.

The admin is trying to get things done quickly, not waste months switching vendors unless the switch becomes absolutely necessary — hence pressure on Anthropic to modify its terms.

The pompous energy comes from projecting a first-deployment-and-integration advantage as: (1) superior tech and (2) a moral mandate (i.e. the Pentagon is morally failing if they switch).

We would not even be having this conversation if OpenAI or Gemini or xAI had been the first model integrated into Palantir’s classified stack.

CharlotteFang77 drew a historical parallel:

“Being merely cognizant of the parallels of foundation model labs and the manhattan project does not mean they derived the correct historical lessons. Dario visibly failed his chess match, in a patently naive way. The entire rationalism sphere should be recalibrating, not coping.”

She attached the Truman-Oppenheimer anecdote — when Oppenheimer came wringing his hands about having “blood on my hands,” Truman told Acheson:

“Never bring that fucking cretin in here again. He didn’t drop the bomb. I did. That kind of weepiness makes me sick.”

(Source: Jean-Jacques Salomon, Science et Politique, 1970.)

The elected commander-in-chief makes the decisions. The vendor provides the tool. Dario confused which role he occupied.

As Indian_Bronson put it, responding to Yudkowsky’s overwrought take:

“The person elected by Americans to be the president and commander in chief is in charge of the military’s contracts, not CEOs of private companies.”

Scott Alexander’s tweet is a correct observation of hypocrisy on both sides of the aisle (but largely irrelevant when considering one political party wants to maintain some semblance of the U.S. and the other wants to give it away to whoever enters illegally):

“I cannot wait until the White House changes hands and all of you ghouls switch back from ‘you’re a traitor unless you bootlick so hard your tongue goes numb’ to ‘the government asking any questions about my offshore fentanyl casino is vile tyranny.’”

Scott appropriately highlights what always happens: both sides flip on gov power depending on who wields it. (And a related fact: both sides think the economy suddenly sucks as soon as the other side wins an election.)

Anyways… if we know this pattern always holds… the proper takeaway is: you are in a Prisoner’s Dilemma and other side defects every time. If you don’t defect (i.e. use AI for max advantage while in power), you give the other side a potential layup to obliterate your party (or at least its current iteration).

If you’re the Trump regime and you do anything less than “go all out” with AI — it may cost you the next election and weaken your ability to get things done before your time is up.

Strike while the iron is hot or get struck by the iron of the opposing party.

As Geiger Capital pointed out: whether you like it or not, it’s happening. Develop the tools or be at the mercy of China.

I had articulated a variation of what Geiger is saying earlier in the week. Anytime you throttle yourself with AI in any way, you are ceding an edge to adversaries like China (even if the uses have nothing to do with China directly).

The Political Prisoner’s Dilemma

The Democrats spent 2020–2024 building surveillance and content-moderation infrastructure under labels like “countering disinformation,” “election integrity,” and “domestic extremism.” Nobody in the current commentary class was writing 5,000-word blog posts about AI vendor red lines when the targets were different.

Now a Republican administration is in power and suddenly AI-enabled analytics are an existential threat to democracy. But both parties will use every tool available when they hold the levers.

Democrats will leverage AI to the fullest for their causes if they win presidency in 2028 and beyond: social media monitoring, “hate speech” detection, helping illegals enter the country and become citizens, etc. (And we already know Claude values illegals more than actual Americans).

If you believe otherwise, you haven’t really been observing reality.

What does “domestic surveillance” even mean today?

• Catching MS-13 members, cartels, predators, human traffickers

• Pinpointing Chinese espionage campaigns

• Identifying domestic terrorist networks

• Tracking those who crossed the border illegally and disappeared into the interior

• Monitoring ideological radicals and political extremists with high odds of violence

• Pinpointing fraud networks throughout the U.S.

The government already has these datasets. AI just adds an analytical layer that makes enforcement scalable and more efficient rather than a slothified game of whack-a-mole.

The left frames this as “mass surveillance” of Americans. (Anthropic is a private company literally engaged in mass surveillance of Americans to improve their AI model! Can’t even make this up.)

The same left also wants a pathway to citizenship for millions of illegal immigrants: 85-90% of Democrats support this; which would permanently reshape the political composition of the country and the way it functions.

At a certain point it’s just the “United States” in name only (USINO); ethos and foundational principles devolve into leftism (socialism, redistribution, legalized crimes, regulation, anti-innovation, etc.); California is the first experiment (Exhibit A)… now even billionaires are concerned and ready to flee (wealth tax coming).

And the Democrats are people who openly admit to not liking the country they live in: Gallup’s 2025 data shows only 36% of Democrats are “extremely or very proud” to be American, down from 90% in 2001. Republicans? 92% like the U.S.A.

This isn’t too surprising given that the radical left hates the foundational principles and ethos of the U.S. — and many don’t even have historical ties to building and shaping the country (1 in 4 are first/second gen immigrants; zero skin in actually building the country up — most arrived illegally after already a success).

When the base of a political party has almost zero pride in its own country, the party is by definition anti-American in orientation; and the institutional priorities reflect it.

Whatever you think of the current Republican Party (there are definite flaws), we know it’s not even close to as bad as what the Democrats stand for:

• Expanded government

• De facto open borders (legal pathways)

• Withdrawal from global stage (don’t use military)

• Added regulations (evil corporations!)

• Socialism/redistribution (anti-growth, anti-innovation, anti-capitalism)

• Subsidizing the world even if it damages your own country

• Anti-Western and anti-White identity politics imposed from within

This is so far gone from the constitutional republic the founders built that they’d be rolling in their graves if they knew what was happening.

If you actually want to save the country your top priority should be focusing on our biggest issue: the border and illegal immigration. Downstream of that we need to dismantle gang and terrorist networks, derail Chinese espionage campaigns, and enforce American laws at scale. We need every efficiency gain we can get.

The people living in high-trust communities that were destroyed by illegal immigration don’t have the luxury of waiting for a startup CEO to finish agonizing over “democratic values.”

The Geopolitical Reality

The United States is in a cold war with China; Taiwan, the South China Sea, semiconductor export controls, Iran fueling Chinese strategic interests.

Geiger Capital:

“China doesn’t give a shit about Anthropic’s moral red lines. They are implementing AI into their entire military chain, and they are doing it with zero democratic or civilian oversight.”

Fully autonomous weapons are coming whether Anthropic participates or not.

TimothySawyerMD:

“What is your response to the case that many on X have made, that America’s enemies are not placing these restrictions on their militaries, and that if they win the AI arms race, they will not allow Anthropic to even exist?”

The question answers itself.

Every month of friction on the U.S. side is a month that the Trump administration gets less done.

Anthropic’s stand doesn’t actually stop anything.

• OpenAI signed a classified deal the same night with weaker protections.

• xAI and Elon already worked out a deal with zero restrictions.

• Open-source models can also be deployed internally with zero restrictions.

The government gets the same capability from someone else within months. The net effect on civil liberties is approximately zero.

Dario risked significantly crippling his company to take a moral stand against a democratically-elected regime who just wants to use Claude under current American laws; they aren’t asking to break laws.

So Anthropic gets mogged by Hegseth and Trump, takes serious heat in the process, and puts up a temporary speed bump on the I-95… that the regime will steamroll.

What will happen next?

I’m not sure. Some (as I’ve noted) think we’re in for a TACO buffet. Others think Anthropic just nuked their entire company. I’d guess somewhere in the middle and am actually hopeful that both Anthropic and Trump sort this out even if tensions are high.

I don’t think anything major will happen to Anthropic, but the tail risk lingers.

The legal picture is murkier than maximalist framing.

• WIRED reports multiple contracting experts saying it’s currently impossible to determine which customers must cut ties.

• The authority being invoked — 10 U.S.C. § 3252 — is scoped to “covered procurement actions” in defense acquisition, not a universal commercial ban. It requires a written determination and congressional notification.

I had GPT-5.2-High brainstorm the most likely outcomes:

• Scenario A: Prolonged Ambiguity (~40%) — No quick injunction. Defense-adjacent enterprises quietly de-risk. OpenAI and xAI capture government distribution. By 2029, Anthropic is ~20–30% below counterfactual from distribution drag, not model quality.

• Scenario B: Narrow Court Relief (~25%) — Anthropic challenges the designation and gets a preliminary injunction narrowing “any commercial activity” to DoW contracts only. Government business lost, but commercial contagion fades. ~10–15% competitiveness hit.

• Scenario C: Quiet Settlement (~20%) — Anthropic reinitiates talks and accepts “all lawful use” anchored to legal authorities (i.e. the OpenAI framework). Designation withdrawn. This would be the most rational outcome — but rationality left the building when both sides went public with the dispute.

• Scenario D: Full Enforcement (~15%) — Broadest interpretation holds. Major partners choose DoW over Anthropic. Possible but a tail risk. Kicked off of all major clouds and partnerships (AWS, Google, etc.). Effectively ends the existence of Anthropic.

Final Take: Anthropic vs. Trump

Strip away the moralizing:

1. Anthropic got the Palantir integration first. This gave Claude a big deployment advantage in classified environments.

2. When the Pentagon pushed for “all lawful use” contract language, Dario insisted on framing the guardrails as Anthropic’s terms rather than references to existing legal authorities.

3. He antagonized Pentagon leadership in meetings and blog posts.

4. The Pentagon escalated to the nuclear option.

5. OpenAI swooped in hours later and signed a deal with protections framed the way the Pentagon found acceptable.

For the Trump administration, this was an annoying vendor dispute that got personal and evolved into a week of headlines. The classified AI pipeline doesn’t depend on Anthropic.

My thoughts: I think Trump et al. had to play hardball with Dario et al. or other AI companies may not have agreed to waive all restrictions in classified and unclassified settings (e.g. xAI, OAI, etc.). Without setting the potential “supply chain risk” classification precedent, others can say “We want the same setup as Anthropic.” I think Sam Altman navigated the situation brilliantly for OpenAI. I love Claude as an AI model and even like Dario (the CEO) and think he’s a good guy — but he had to look tough against Trump on “ethics” or he’d face internal backlash from employees and lose some social clout within EA/rationalist cults. If the WaPo framing re: ICBM’s of “give us a call and we’ll work something out” is correct (seems hard to believe), I can understand why this triggered “scorched Earth” mode from Trump and Hegseth. Dario is now a left-wing hero… but collects and uses highly-sensitive data to improve Claude models (the irony). Claude is now getting more attention from the general public and the marketing is excellent… individual users are skyrocketing… but enterprise is where the fat stacks ($$$) are at… if major blacklisting occurs… they are cooked. I personally think the supply chain risk designation went too far, but was the impetus for more favorable deals with OAI and xAI. The smartest move IMO is to retract full supply chain risk designation and keep Claude for non-surveillance and non-autonomous use-cases.

Cue the woke intellectual rationale:

1. Teacher quality is getting worse!

2. COVID really set these kids back!

3. Cell phones are clearly the cause!

4. Kids are losing critical thinking skills cuz of AI!

5. Schools aren’t getting enough funding!

These aren't entirely wrong — they're just secondary. Teacher quality has taken a hit (DEI prioritization hasn't helped), COVID accelerated everything-digital, and kids are overstimmed on caffeine and distracted by screens. But none of these explain the core pattern.

If you're working with students who have lower innate intellectual potential, the best move is probably sports and exercise for half the day, then basic reading, writing, arithmetic once they're burnt out and willing to focus — zero cell phones. We already know One Child Per Laptop was an epic failure — this is a genetic capacity (substrate) issue.

"Schools aren't getting enough funding" is the most moronic take possible. Schools are massively overfunded. The "best schools" merely correlate with high-IQ kids who get schools well-funded because they were born to high-IQ parents. When we "diversify" top schools via legal mandates — test scores tank, teachers quit, parents pull their kids — and wokes wonder: What went wrong?! Must be COVID and cell phones.

My guess is you could axe the budgets of the “best schools” to almost nothing and if you kept the kids the same, they’d still crush it academically because it was never about the money.

Test Scores are Falling in Developed Countries

The underlying data is brutal.

The OECD’s PISA 2022 results show the OECD math average falling a record ~15 points (489→472) between 2018 and 2022 — triple any previous decline. Reading dropped 10 points, double the prior record.

The PIAAC 2023 results for US adults tell a similar story with important caveats: literacy fell ~12 points from 2017 to 2023, with ~28% now at Level 1 or below — functionally unable to use a library search engine.

White adults at lowest literacy rose from 12% to 16%; Hispanic adults at Level 1 jumped from 31% to 45%. But the PIAAC comparison is methodologically messy — the 2023 round switched to all-tablet, changed scoring, and for the first time included respondents who couldn’t speak English or Spanish via a “Doorstep Interview” protocol.

With 1 in 4 Americans now first- or second-generation, the “within-group” Hispanic decline likely reflects a compositional shift; more recent immigrants in the 2023 sample, not the same Hispanics getting dumber. PIAAC is directionally supportive of the thesis, not dispositive.

Finland, the global education darling, collapsed 64 points in math from its 2006 peak (548→484). Germany dropped 39 points. Norway fell 33 in a single cycle. The Netherlands shed 26.

Conventional explanations: COVID, smartphones, TikTok brain, bad teachers, AI offloading, insufficient funding — are secondary accelerants.

They’re the explanations you’re supposed to say out loud, what intellectuals and policymakers offer because they’re socially acceptable, imply fixable problems, and keep the gov money flowing... they’re what everyone wants to hear.

Are Cell Phones the Problem?

France banned phones in 2018 — before PISA 2022 was even administered — and posted its lowest math score in PISA history.

A King’s College London analysis found countries with more bans actually scored lower — every 10% increase in schools banning phones correlated with a 9.4-point PISA decrease.

Even in OECD schools with bans, 29% of students still use phones daily.

Enforcement is downstream of the same population traits that produce the test scores: conscientiousness, impulse control, respect for authority.

The variable isn’t the phone. It’s who’s holding it.

Is AI (Cognitive Offloading) the Problem?

Some allege that AI (cognitive offloading) may be to blame for the massive decline in test scores… but that makes little sense as a primary driver.

Why?

PISA 2022 was administered in spring 2022 (October–November in the US).

• ChatGPT launched November 30, 2022.

• Claude and Bard (now Gemini): March 2023.

The big drop-off from 2014 to 2022 occurred before AIs launched.

NAEP 2022 was administered January–March 2022, 8 months before ChatGPT — math down 8 points, largest drop in NAEP history, entirely pre-AI.

NAEP 8th grade reading at the 10th percentile had fallen 10 points between 2013 and 2019 — before COVID, before AI, before any of it.

Japan and Korea have the same smartphones, TikTok, and AI access. Japanese teens average 4–6 hours/day online. Both countries’ scores held or improved.

The answer has to be something present before 2022 that differs between countries that declined and countries that didn’t.

Demographic changes

The primary driver of test score decline across the developed world is demographic change; replacement of higher-scoring populations with lower-scoring ones through unselected immigration and differential fertility.

The secondary driver is dysgenic fertility for IQ within all groups; the least cognitively adept (within all racial/ethnic groups) are having the most kids.

Other accelerants include: cell phones, COVID, teacher quality erosion, anti-intellectual curricula, and stimulants (Yerkes-Dodson imbalance; suboptimal for focus).

The Natural Experiment: Homogeneous Countries vs. Diversifying Countries

If demographics change is the primary driver behind test scores plummeting in developed countries, there’s a clear prediction: high-performance countries that remained demographically homogeneous should have roughly held their scores, while countries undergoing rapid demographic change should have crashed.

That’s pretty much what happened.

Eight systems — all gained or held steady while the OECD average cratered. They have the same smartphones, TikTok, and AI access as the West — Japanese teens average 4–6 hours/day online, 70% use TikTok — and they had COVID.

Yet only 5% of Japanese students report phone distraction in math class vs. 30% OECD-wide.

The one variable these countries don’t share with the declining West is mass unselected immigration.

Every time PISA drops, think tanks ask: “What can we learn from Japan, Singapore, Estonia?”

Nobody asks the obvious question: What if you don’t need their methods — just their people?

Japan didn’t gain because it discovered a pedagogical secret.

Estonia doesn’t outperform Germany at half the spending because Estonian teachers are twice as good. The variable is DNA.

Estonia maintains mostly Estonian ethnicity, whereas Germany heavily imported non-German ethnicities.

Now compare to the countries that diversified rapidly:

Germany — immigrant share doubled in one decade; native-immigrant gap is 59 PISA points (32 after SES controls). The 1.1M refugees from 2015 came from Syria (~83 IQ), Iraq (~87), Afghanistan (~84).

Sweden — 11% first-generation in PISA sample; top sources: Syria, Afghanistan, Somalia, Eritrea, Iraq. Sweden’s own PIAAC data: college-educated migrants from Arab states and Sub-Saharan Africa scored at the same numeracy level as low-education native Swedes. 76% of immigrant students don’t speak Swedish at home. African-origin students have the lowest grades and have declined over time.

Finland — smaller immigrant share but steepest decline of any OECD country (−64 from peak). Same origin profile as Sweden: Somalia, Iraq, Afghanistan.

The UK — least trustworthy data. Immigrant students rose 13% to 20% but PISA classifies second/third-generation Pakistanis, Somalis, and Caribbeans as “native” while current immigrants skew toward selected Indians and Chinese. GCSEs show convergence but had g-loading halved by Blair-era reforms. Every other measure — CAT3, PISA, university outcomes — shows substantial gaps. Cremieux confirmed: same sample, GCSE parity but significant IQ gap.

Canada — the “selective immigration” everyone praises. 34% immigrant PISA sample. Math dropped 35 points since 2003 — worse than Germany. Manitoba fell 58. Canada’s “points system” is ~58% “economic class” including unscreened spouses and children; the actual screened applicant is under a third of intake. Even that is gameable — 20,000 Indian students enrolled in 2024 without attending class.

The correlation is clear. AI, phones and COVID cannot explain this gradient.

The US: Arithmetic That Nobody Wants to Do

The compositional math in the US is straightforward.

Non-Hispanic White share (Census Bureau):

• 69.1% (2000) → 57.8% (2020). White population actually declined in absolute numbers 2010–2020 — first time in US history.

• In public schools: White students dropped from 61.2% to 44.7%; Hispanic nearly doubled from 16.4% to 28.7%.

NAEP scores by race (8th grade math, 2022): White ~275, Hispanic ~257, Black ~244, Asian ~295. These gaps map onto the Roth et al. (2001) meta-analysis of 6.25 million test-takers (B-W gap: 1.1 SD; H-W: 0.72 SD).

The arithmetic: Replace 16.5 percentage points from a group averaging ~275 with one averaging ~257. Mechanical effect: ~3–4 NAEP points. Add English Learner growth and within-group compositional shifts, and demographics plausibly accounts for 4–6 of the ~9-point 8th grade math decline from 2013–2022.

Tom Loveless documented this precisely: 2009–2019, national 4th grade NAEP math was flat while every subgroup gained. Simpson’s Paradox — the aggregate hides the compositional shift.

The US immigration system is ~two-thirds family-based, ~17% employment-sponsored, plus 11M+ unauthorized.

The Swiss natural experiment is the cleanest causal evidence: switching from low-skill guest workers to EEA free movement improved immigrant PISA scores 43 points in a decade — Cattaneo & Wolter (2012) found ~75% was due to changed immigrant characteristics. Not better schools. Different people.

Richwine (Harvard, 2009): third-generation Hispanic immigrants still ~0.5–0.75 SD below the White mean. Environmental uplift produces gains, not convergence.

The Double-Whammy: Replacement + Dysgenic Fertility

Two simultaneous forces are pulling the cognitive mean of every Western nation downward, and no educational reform will reverse either.

1. Between-group replacement: Higher-scoring populations (below-replacement fertility) are shrinking while lower-scoring immigrant-origin populations (above-replacement fertility) are growing.

2. Within-group dysgenic fertility: In every racial and ethnic group, the least cognitively able are having the most children.

US fertility by education (2019 CDC/NCHS): no diploma TFR 2,433 vs. bachelor’s 1,396 — a 1.74x fertility advantage for the least-educated. Since education correlates with IQ at r ≈ 0.55–0.65, this is direct dysgenic pressure.

By race, the pattern is a near-perfect inverse correlation between cognitive scores and fertility:

• Asian — NAEP 8th grade math ~295, TFR ~1,519 (lowest fertility, highest scores)

• White — NAEP ~275, TFR 1,610

• Black — NAEP ~244, TFR 1,774

• Hispanic — NAEP ~257, TFR 1,939 (highest fertility, near replacement)

The two highest-scoring groups have the lowest fertility. The two lowest-scoring are nearest to or above replacement. Within the White population, the most educated women (bachelor’s+, TFR ~1.4) are being outbred nearly 2:1 by the least educated (no diploma, TFR ~2.4). This pattern holds within every group.

Kong et al. (2017, PNAS), using deCODE genetics on 129,808 Icelanders, showed polygenic scores for educational attainment declining ~0.30 IQ points per decade — genetic deterioration documented at the genomic level in one of the most homogeneous populations on earth.

A study many cite to rebut demographic changes is Bratsberg and Rogeberg (2018) — within-family IQ decline in Norwegian brothers.

Three problems: (1) the birth order confound (first-borns score 2–3 points higher — you can’t separate this from “born later in time”), (2) the paternal age effect (~1–2 de novo mutations per year — a genetic mechanism), and (3) magnitude (0.2–0.3 points/year vs. Germany’s 59-point native-immigrant gap). B&R is a footnote cited as if it refutes a tsunami; the situation is worse than demographics alone.

High IQ Kids = High Test Scores

Educational quality is downstream of population quality, not the other way around.

The causal chain runs like this:

1. High-IQ parents cluster in affluent suburbs through sorting on income, which is itself substantially determined by cognitive ability.

2. They fund schools through high property taxes, creating per-pupil spending that dwarfs urban districts (even adjusting for cost of living).

3. Good teachers want to teach there because the students are teachable, the parents are engaged, the classrooms aren’t behavioral management nightmares, and the pay is competitive.

4. Students absorb the material because they have the cognitive substrate to do so, reinforced by home environments where educated parents supplement schooling.

5. Test scores are high. Everyone attributes this to the great teachers and great schools.

But the causal arrow runs primarily from student quality → teacher quality → school quality, not the reverse.

Put the same “great teachers” in Baltimore or Detroit and scores aren’t gonna budge.

Finland is the proof case in reverse. Finland got elite teachers (top 17% of applicants, master’s degrees required) because it had a homogeneous, high-IQ population that was genuinely teachable.

• Teaching was a high-status profession because the work was intellectually rewarding.

• As Finland’s demographics shifted and classrooms became harder to manage, teaching enrollment has started to decline — the same pattern the US experienced decades ago, just on a lag.

In the US, education majors now score 28 points below the SAT average and have the lowest GRE Quantitative scores of any graduate field (149 vs. 161 for engineering).

Teaching enrollment is down 33% since 2010. Smart people with options don’t sign up for work that increasingly consists of behavioral management for students who lack the cognitive capacity to absorb the material.

The “good teacher” is mostly a byproduct of the good student — and both are a byproduct of the genetic stock of the parents and the community that funds the school.

AI-Assisted Learning Bifurcation

The same technology is producing radically different outcomes depending on the cognitive substrate of the user.

NAEP data: the 10th percentile of 8th-grade readers fell 10 points between 2013 and 2019 (before COVID) while the 90th percentile held steady.

Khan Academy’s Khanmigo exploded from 68,000 to 700,000 users in one year, with meaningful gains for engaged students.

But the Khan Academy CLO actually reports: some students engage exactly as designed, “answering questions and posing their own to deepen understanding.” Many others respond “I don’t know” or “Bro, IDK.”

High-ability kids use AI tutors, Claude, ChatGPT, and self-directed platforms to learn at accelerated pace, often routing around the school system entirely. A motivated 14-year-old with a 130 IQ can learn calculus or Python with no teacher and no per-pupil spending. The technology amplifies existing cognitive capacity.

• Alpha School — Compresses all academics into two hours of AI-driven adaptive learning per morning. Their students learn 2.6x faster on average, top 20% at 6.5x, scoring 99th percentile on MAP across nearly all grades and subjects. First graduating class: 11 of 12 to four-year universities.

• Mentava does the same for gifted early readers. AI-driven phonics with human coaching for high-ability kids whose needs aren’t met by age-normed classrooms. These schools work because $75K tuition selects from the right tail of the cognitive distribution, and AI removes the artificial pace constraint. The technology unleashes innate ability that was already there.

When Arizona greenlit Alpha’s charter model for public funding (“Unbound Academy”), Pennsylvania rejected the identical application: “untested and fails to demonstrate alignment.”

They’re probably both right. It works for Alpha’s population. It may not work for the population Pennsylvania is worried about; “fails to demonstrate alignment” is a laughable take but probably for the best.

Low-ability kids get the same technology and respond “Bro, IDK.” The tool won’t create cognitive capacity that isn’t there. Same technology, different outcomes, determined entirely by substrate. The schools where AI tutoring “works” were already producing good outcomes because of the selected students.

Spend More on Education = Money Incinerator

The US spends $16,500+ per pupil nationally, 38% more than the OECD average. Lighting money on the fire… lighting more on fire… and think that lighting even more money on fire will somehow eventually produce a positive ROI.

This is like giving midgets NBA-level coaching and training and thinking they’re gonna make the pros one day — or like effective altruists throwing money into developing countries and thinking the newly funded systems will remain sticky (rather than requiring infinite money).

The only solution is targeting biology: gene modification, embryo engineering, embryo selection, etc. — for higher IQ.

• Estonia: ~$8,500/pupil → 1st in Europe, 6th globally in PISA math

• Japan: ~$10,000/pupil → +9 points, now 1st in reading

• The United States: ~$16,500/pupil → 34th in PISA math, declining every cycle

Estonia gets 2x the output per dollar because the median Estonian absorbs the instruction. Past a relatively minimal threshold, additional dollars have zero ROI because the constraint is biology.

The Kansas City experiment (1985–1997) settled this empirically.

• A federal judge invited educators to “dream” — forget about cost.

• Over 12 years, nearly $2 billion: 15 new schools, Olympic swimming pool with underwater observation room, TV/animation studios, robotics lab, 25-acre wildlife sanctuary and zoo, model United Nations with simultaneous translation, field trips to Mexico and Senegal.

• Student-teacher ratio fell to 12–13:1, lowest of any major district. Highest per-pupil spending of the 280 largest districts.

Result: Test scores unchanged. Black-white gap unchanged. Dropout rate went up. Nonwhite enrollment rose from 73% to 80%+. “For decades, critics said ‘You can’t solve educational problems by throwing money at them.’ Educators replied, ‘No one’s ever tried.’ In Kansas City, they tried.”

Washington DC: ~$31,000/student — highest in the nation. 8th graders score two grade levels below national norms. DC schools are 64% Black, 19% Hispanic. Results track the population, not the expenditure.

LeBron James’ I Promise School (Akron): $1.4M/year from the Foundation on top of district funding; additional tutors, smaller classes, extended days/year, summer tutoring, wraparound services, guaranteed free college tuition. Everything the education establishment says it needs.

Result: 2 out of 75 seventh-graders passed state math (2.7% vs. 50% state average). Inaugural class: 0% math proficiency three years running. 125th lowest of 3,318 Ohio schools.

Fordham Institute found I Promise students performed worse than matched peers at regular Akron schools. The NYT celebrated the school in 2019. They haven’t written about its test scores since.

Jalen Rose Leadership Academy (Detroit): 97% graduation rate. Math proficiency: 2.42%. You can graduate every kid, you cannot make them proficient.

Head Start: ~0.2 SD initial IQ gains fade out completely by first grade, undetectable by third.

Perry Preschool: same fade-out by age 8. The most replicated finding in educational research — and the one every policy discussion ignores.

The OECD finds spending explains ~54% of performance gaps up to ~$75,000 cumulative spending (ages 6–15), but above that threshold the relationship vanishes. The US is far above.

The ROI is negative: Every dollar poured into students who can’t absorb instruction is diverted from students who can. Kansas City spent $2 billion and moved the needle zero. LeBron provided everything money can buy and got 2.7% proficiency. Money buys results only when the substrate can convert it into learning. When it can’t, you get Olympic swimming pools and 0% proficiency.

The Elites Tell You What You Want to Hear… Not What You Need to Hear

No economist, education researcher, or policy analyst can say “test scores are falling because we imported millions of people with lower cognitive potential and the least capable members of every group are having the most children” and keep their job.

So they find real but secondary mostly bullshit factors: COVID, phones, teacher shortages, AI offloading — and present those as the story.

The audience smiles, nods, agrees. An op-ed gets published. Gov comes up with a big “AI education spending initiative to promote equity” (or something).

Not once does anyone mention the demographics of the tested populations dramatically changed.

And never will they highlight Japan (with cell phones, AI, TikTok, YouTube, multitasking, gaming, etc.) improved during the same time horizons because it didn’t change its demographics.

The New York Times will never run “It’s the Demographics, Stupid” because the implications are socially unacceptable — even if they’re empirically unavoidable and obvious to truth-seekers.

What will actually fix test scores?

Besides making all tests absurdly easy so that everyone “does well”; this is test inflation… and the worst aspect is that declines are happening with rampant test inflation.

If you are serious about fixing this… we need rapidly scaled intelligence upgrades for humans that didn’t evolve to function in the recent cognitively-demanding times.

1. Highly selective immigration. This is the single highest-leverage policy. The US family-reunification system and Europe’s humanitarian intake produce the opposite. Every percentage point of the population shifted from high-scoring to low-scoring groups moves the national cognitive mean downward.

2. Stop wasting money on “closing gaps” that are biologically constrained. Education spending should be redirected from futile attempts to equalize outcomes toward efficiency and discipline. For lower-performing populations, the focus should be on vocational training, behavioral discipline, and practical skills — not on forcing academic content that isn’t being absorbed. For high-performing populations, remove obstacles and let AI-assisted technology amplify their ability.

3. Phone/social media restriction in low-performing schools. The highest-ROI near-term intervention. PISA 2022 data shows 15 points of benefit. Japan’s low distraction rates (18% vs. 59% OECD) correlate with maintained scores. This doesn’t fix the structural problem but slows the environmental bleed in low performers.

4. Bioenhancement — the only real long-term solution.

• Embryo selection with Herasight et al.: current technology could plausibly produce 5–15 IQ point gains per generation across all groups. This raises the cognitive floor for everyone. Eventually use WGS… and eventually embryo engineering.

• Somatic gene editing for intelligence/cognition once the relevant variants are characterized and delivery mechanisms mature. I have a protocol that I will release soon and by the 2030s-2040s people can test if they want.

• Age reversal (Operation Senolysis) to extend the productive lifespan of high-human-capital individuals. Every high-IQ person who dies at 75 instead of 150 is an enormous loss of productive capacity and institutional knowledge.

These are the only interventions that address the underlying biological constraints rather than trying to paper over them with spending that demonstrably doesn’t work.

The Human Capital Cliff Is Here

The developed world built institutions, economies, and technological infrastructure on a population with a certain cognitive profile — and that population is being replaced via unselected immigration and differential fertility.

The new population has a different profile, and predictably, test scores are falling.

You can take phones away, fire bad teachers, hire quality teachers, pour another $200 billion into schools and none of it will reverse the tide. Why? The tide is about the human capital of subsequent generations.

The only serious move to prevent a Dark Age in developed countries is: age reversal + embryo engineering + somatic IQ upgrades in adults + AGI/ASI robot diffusion.

Most other things are just rearranging deck chairs.

Note: This is why I think Brad Gerstner’s “Invest America” idea is well-meaning but misses the mark. It increases government entitlements (spending) at a time when the U.S. is in massive debt (with massive interest on that debt) and assumes the new demographics won’t just “cash out” and blow the investment the minute they turn 18. Does skin in the game matter if it isn’t your skin? And does it work if the median doesn’t have the innate ability to delay gratification? I suspect it’ll have zero effect on the growing demand for socialism in the younger, more Hispanic generation.

The beauty is that even if: (A) you are ignorant enough to oppose bioenhancement, (B) it’ll still end up benefitting you indirectly: more innovation, more societal contributions, lower crime, better behavior, higher performance, less disease, and favorable genes circulating through the general population.

Humanity has a much higher “p(doom)” if we don’t step on the gas pedal and accelerate: (1) AGI/robotics diffusion; (2) bioenhancement (age reversal + adult somatic upgrades + embryo selection/engineering); the counterfactual “pause/stop AI” is likely much worse.

We are not moving nearly fast enough on the AI front (despite what you read in the news)… and we are in ultra-sloth mode on bioenhancement; the FDA, regulators, academia, and bioethicists have everyone by the balls and are causing far more harm than Operation Warp Speeding everything.

If we really want to become a Kardashev Type 1… we need to ignore the hordes of morons in favor of stopping/slowing AI and wake up to the pending human capital cliff (potential Dark Age Lite dystopia future) if people don’t step up and get the: (1) left-wing pseudo-reality-woke-egalitarian-fantasy-land socialist retards + (2) right-wing populist morons — on track.

Arthur Jungguo Liu, born in 1964 in a Sichuan mountain village without electricity, has a rarefied American immigrant backstory.

He organized pro-democracy demonstrations in Guangzhou during Tiananmen, landed on the CCP’s wanted list, escaped China at 25 by boat to Hong Kong, and arrived in the U.S. as a political refugee.

He worked as a busboy, then earned an MBA and a JD, passed the California Bar, and opened a small immigration law firm in Oakland. His family of six lived in a one-bedroom apartment for two to three years.

At 40, unmarried and unwilling to wait for a partner, Liu made a deliberate decision: he used his own sperm, anonymous Caucasian egg donors, and gestational surrogates to father five children between 2005 and 2009.

• Two different donors, two different surrogates, five kids: (1) Alysa (2005), (2) Selina (~2007), and triplets (3-4-5) Joshua, Justin, and Julia (~2009).

• All biracial Chinese-Caucasian. Arthur told reporters he “Felt his children would benefit from a diverse gene pool.” He doesn’t know the identities of either donor.

His eldest, Alysa, became one of the most accomplished figure skaters in American history:

• Age 12: Youngest woman ever to land a triple Axel

• Age 13: Youngest-ever U.S. women’s national champion; first American to land two triple Axels in one program

• Age 14: Back-to-back national titles, youngest ever to do so

• 2025: World Championship gold — first American women’s world title since 2006

• 2026: Olympic gold at Milan-Cortina with a career-best 226.79, first American woman since Sarah Hughes in 2002

Arthur invested an estimated $500K–$1M in coaching, brought a radar gun to measure jump speed, and fired and rehired coaches multiple times.

Genetic potential met extreme environmental investment; but the genetic inputs mattered: coordination, body type, pain tolerance, explosive fast-twitch muscle — these are substantially heritable traits that the donor selection helped provide.

The Chinese government (CCP) tried to destroy his family

In October 2021, the FBI contacted Arthur to warn him that he and Alysa were targets of a Chinese government espionage operation. Arthur had been a dissident since Tiananmen, and Beijing apparently never forgot.

In March 2022, the DOJ charged five men with acting as illegal agents of the People’s Republic of China to surveil and harass Chinese dissidents living in the United States. Arthur was identified as “Dissident 3” in the court documents.

One of the operatives, Matthew Ziburis, had posed as a United States Olympic & Paralympic Committee official and attempted to obtain copies of both Arthur’s and Alysa’s passports.

The scheme included plans to physically surveil the Liu family and potentially disrupt Alysa’s skating career — a 16-year-old American citizen being targeted by a foreign intelligence service because her father had organized protests 30 years earlier.

Alysa competed at the 2022 Beijing Olympics under heightened FBI security, an almost surreal situation: an American Olympian skating in China while her father’s home country was actively running espionage operations against her family on U.S. soil.

She placed seventh, won bronze at Worlds, then retired citing PTSD — not just from competitive pressure, but from the weight of everything surrounding it. Her 2024 comeback, on the explicit condition that Arthur would not be involved in coaching decisions, led to the World title and then Olympic gold.

Worth noting…

Liu wasn’t the only athlete from the U.S. at Milan-Cortina resulting from deliberate genetic selection. Eileen Gu (genetically half Chinese born in California) — the most decorated freestyle skier in Olympic history — was raised by her mother, a Peking University speed skater and Stanford biochemistry researcher who later earned an MBA from Stanford Business School. Her father is an unnamed American with a Harvard education; he has never been publicly identified and plays no role in her life. Her mother made a pragmatic decision to optimize the genetics of Eileen Gu.

What Arthur Liu did with intuition, men can now do with data

Liu selected donors from agency catalogs based on photos, health history, and background. He had no genetic scoring tools. The technology available in February 2026 is orders of magnitude more powerful — and it’s improving month by month.

If you are a middle-aged guy who wants kids badly, but you haven’t found a good woman to marry and/or have kids with — pulling an Arthur Liu act and going straight to some elite embryos to combine your sperm with is worth considering.

Build your own powerhouse family; it may enrich you and all of society. You have it much easier than Liu too.

Herasight: the current leader

Herasight is the most technically advanced player in the space. Their CogPGT 1.0 predictor explains 16.4% of the variance in fluid intelligence; a 4-fold improvement over 2019-era scores (~4%).

Within-family validation shows a 0.45 correlation with general cognitive ability across 6,442 sibling pairs. For 10 embryos, they claim a detectable IQ spread of ~15 points lowest to highest.

The key innovation is ImputePGTA — reconstructing full embryo genomes from standard PGT-A data clinics already generate. No specialized lab, no extra biopsy.

On February 17, 2026, they released V2, which jointly models all embryos in a cycle (even aneuploid ones normally discarded) to improve phasing.

Biggest gains: parents from underrepresented ancestries. A couple in Seoul, Lagos, or Mumbai can now access screening approaching European-ancestry accuracy.

• 17 diseases screened, accuracy 122% better than Orchid’s, 193% better than GP’s

• 20–44% disease risk reduction when selecting among 5 embryos, with positive pleiotropy across conditions

• Customers from 11 countries; ethics paper in Fertility and Sterility

• Up to $50K comprehensive; 90% discounts for first qualified customer per country

• Team: Michael Christensen (founder), Tobias Wolfram (CSO, has banked his own embryos), A.S. Young (statistical genetics), Jonathan Anomaly (ethics)

The competition

Orchid Health — WGS gold standard: 30x coverage, 99% of embryo DNA, $2,500/embryo. PGT-A + PGT-M + PGT-SR + PGT-P in one report, ~1,200 monogenic conditions + 12 disease polygenic scores. January 2026 GRS update nearly tripled predictive power for atrial fibrillation. Backed by Altman, Armstrong, Wojcicki, Buterin, Church. Musk reportedly used Orchid for at least one child. Limitation: no cognitive scoring; disease scores less powerful than Herasight’s.

Nucleus Genomics — Avoid. Whitepaper plagiarized from Herasight, stock photo testimonials, GP sued for IP theft. Cremieux called it fraud.

Genomic Prediction / LifeView (est. 2019) — first to market, 150+ clinics, six continents. Scores now several years old. ~$3,500 including PGT-A.

The intellectual case for deliberate selection

The researchers who understand this are almost unanimously in favor.

1. Cremieux says: “Everyone Agrees: Let’s Put Genetics to Work.” Stating that these companies provide “immense value for humanity” and we are “on the cusp of a golden age.”

2. Gwern provides the definitive expected-value math.

3. Stephen Hsu, Scott Alexander, the Collins family, etc. have all publicly advocated.

Only 17% of the public considers PGT-P immoral; comparable to opposition to SAT prep.

The biology favors acting now

Paternal age is quietly devastating:

• DNA fragmentation 4.58x higher by age 50

• ~2 de novo mutations/year of paternal age

• 5.75x autism risk for fathers over 40

• Conception probability halves vs. men under 25

Young egg donors (25–30) reset the clock: ~30% aneuploidy vs. 90%+ at 44, 75–85% implantation rates regardless of surrogate age. Arthur Liu was 40 when Alysa was born. He didn’t wait too long.

The gains are multiplicative

Donor selection shifts the genetic mean upward. Screening picks the best embryo from that elevated baseline.

These compound: Gwern estimates top-of-10 selection yields 6–9 IQ points above random; a top-percentile donor stacked on top → 8–12+ plausible.

Even 4–5 points → ~$191K–$240K in expected lifetime income.

Arthur C. Brooks: You probably have less of an effect on your kids than you think; as long as kids aren’t neglected, traumatized, or deficient in nutrients and support — outcomes are dominated by genetics.

Caplan’s “Selfish Reasons to Have More Kids”: genetics dominates nurture, so genetic quality compounds higher than any parenting technique or private school. Not saying really good parenting can’t enrich the kids (you should still parent as good as you can) — but genetics matters most.

The critics are wrong

The “playing God” argument is so stupid I don’t even know where to begin. In short: everyone is playing “God” whether they realize it or not.

Selecting a good spouse/partner is playing God. Not selecting a good spouse/partner is playing God. Optimizing your pregnancy is playing God. Not optimizing your pregnancy is playing God.

Whether you partake in embryo selection or not, you are “playing God” by the moronic definition of those who insist that something they don’t do is “playing God.”

Even if we concede that this is “playing God” (and we shouldn’t because everyone else is playing God too)… who says “playing God” isn’t better or “what-God-wants-you-to-do”?

If tools reduce disease risk significantly and you don’t use them — you are playing a dumber, lower IQ God and portraying yourself as virtuous.

Spamming the term “eugenics” is emotional blackmail and at this point nobody takes you seriously. Embryo selection doesn’t sterilize, coerce, or even alter any genome.

As Scott Alexander noted: if you’re doing IVF, you have to pick one anyway — screening just means picking better.

And “access inequality" is a feature, not a bug: wealthy early adopters pay premium prices that fund the R&D driving costs to zero; the same way iPhones went from $599 luxury items to ubiquitous. Restrict access and you slow the entire pipeline for everyone.

This is a prisoner’s dilemma. The question isn’t whether this becomes normal. It’s whether you’re early or late.

How could you replicate Arthur Liu (but better)?

Arthur Liu used his own sperm with elite egg donors and surrogates to produce 5 children, including an Olympic gold medalist. Here’s how to do it better, with 2026 screening technology he didn’t have.

Selecting your egg donor is the highest-leverage decision you’ll make.

Half the genome comes from the egg donor this is where you optimize hardest.

• You want: young (21–29), verified academics or athletics, tall, healthy family history, high egg count (AMH).

• You use your own sperm.

• The resulting embryos are yours: half your genetics, half hers.

• Then you screen them all with Herasight PGT-P to pick the best one.

Two paths to get eggs, ranked by how many embryos you’ll have to screen:

• Fresh cycle (best for selection): A donor does a full stimulation and retrieval just for you. Average ~20 mature eggs, all fertilized with your sperm → 8–14 blastocysts → maximum PGT-P selection power. More embryos = more genetic variance to choose from = better top pick. Cost: $15K–$25K donor compensation + $5K–$10K agency + IVF cycle costs.

• Frozen egg bank (faster, cheaper, fewer embryos): Buy a cohort of 6–8 frozen eggs, thaw and fertilize with your sperm at your clinic. Faster (weeks vs. months), cheaper (~$18K per cohort), but fewer eggs means fewer embryos means less selection power. Fine as a starting point or supplement.

Where to find elite donors, ranked:

1. Donor Concierge: The meta-agency. Searches 230+ vetted agencies and 25,000+ candidates to curate a shortlist matching your exact specs (height, ethnicity, education, athletics). 95% match success rate. Dedicated case manager. Match in 2–3 weeks. Independent — no referral fees from agencies, so they work for you. Harvard-trained founder (Gail Sexton Anderson). This is who you hire if you want the best possible donor and don’t want to search 50 agency websites yourself. Private Client tier for maximum confidentiality.

2. Elite Donor Solutions: Specializes in athletic, tall, and highly educated donors. Concierge matching service ($2K, applied to agency fee). Strong if you have specific physical trait requirements. Direct agency with in-house database.

3. EggDonors4All: Guaranteed blastocyst programs: minimum 3 Day-5 blastocysts or they repeat the cycle free. FDA-registered, NYS-licensed. Frozen cohorts from $18K; fresh cycles $15K–$25K + IVF. Elite program includes Ivy League and physician donors. Good cost certainty.

4. Shared Beginnings: If you want them to handle everything. They’ll pair your sperm with a donor from their screened database, create embryos at their partner lab (Atlantic Reproductive Medicine, Raleigh NC), PGT-A test, and ship. 77% pregnancy rate. Live birth guarantee (80% refund). Good turnkey option if you don’t want to coordinate clinic + agency separately.

5. Donor Nexus Premier: Access to models, professional athletes, and graduate students as donors. Newport Beach based, partners with HRC Fertility (top SoCal clinic). 1,200+ babies born across all programs. Premier donors require faster cycle timelines.

The full playbook (Arthur Liu’s path, upgraded)

Step 1: Freeze sperm now. Every year costs you measurably. $500–$1K + $300–$500/yr storage.

Step 2: Select an elite egg donor using the ranked sources above. Fresh cycle preferred — more eggs = more embryos = better PGT-P selection. Use Donor Concierge to search 230+ agencies if you want the widest net. Standard compensation: $8K–$14K. Exceptional: $25K–$50K. Elite (Ivy League, D1 athletes): $50K–$100K+.

Step 3: Maximize embryo count. Two cycles → 6–14 blastocysts. The marginal $15K–$20K is trivial relative to lifetime gains.

Step 4: Screen with the full stack. PGT-A ($3K–$6K) → PGT-M if carrier → PGT-P via Herasight → Orchid WGS for rare variants. Skip Nucleus.

Step 5: Surrogate. California is the legal gold standard but the most expensive ($130K–$165K). Circle Surrogacy, Growing Generations. For strong legal protections at lower cost: Illinois ($100K–$125K, Gestational Surrogacy Act, pre-birth orders), Nevada or Colorado ($100K–$130K, surrogacy-friendly statutes, lower surrogate compensation). Arkansas is cheapest but less legally tested. Michigan legalized compensated surrogacy in April 2025 — fresh market, potentially lower rates. International: Mexico ($70K–$80K) and Colombia ($65K–$75K) if cost is the primary driver.

Step 6: Invest in environment. Genetics sets the ceiling; environment determines how close you get.

What this costs and potential future cost

The trend line: what costs $350K today with 16% variance explained will cost $75K or less with 40–50% variance explained — cheaper AND dramatically more powerful. Within two generations, the cost will approach routine medical expenses while the screening accuracy will make unscreened reproduction look like refusing an ultrasound. This becomes a no-brainer for virtually every family.

Every line item is on a steep cost curve. WGS: $20M (2003) → $200 (today).

• IVF: down ~40% since 2010.

• Polygenic scoring: approaching zero marginal cost — Herasight’s ImputePGTA runs on data clinics already generate.

• IVG will eliminate egg retrieval costs entirely.

• Surrogacy: state arbitrage alone saves $30K–$60K vs. California; new markets (Michigan, international) are compressing prices further.

Governments should accelerate this. Once the expected ROI math becomes positive, governments are foolish to not offer this technology to couples for free or heavily subsidized. This would be one of my top priorities and I wrote about it in my ASAP Protocol to Fix the U.S. Today's early adopters pay a premium; within a generation, this will be as routine and affordable as prenatal ultrasound and far more effective.

The future of upgraded babies

The current constraint is embryo count: ~3–7 blastocysts per cycle. Five developments will blow this open, and each makes today’s costs look absurd.

In vitro gametogenesis (IVG): Eggs from stem cells → hundreds of embryos per cycle. At 100 embryos, expected IQ gains may climb to ~18–19 points.

• Conception Biosciences (Sam Altman-backed) has observed markers of meiosis in stem cell-derived germ cells. Hayashi estimates 5–10 years.

• Other players: Ivy Natal, Gameto, Vitra Labs.

• Cost impact: Eliminates donor compensation, egg retrieval, and hormonal stimulation; the single biggest cost reduction in the stack.

Improving polygenic scores: 16.4% of intelligence variance today vs. unknown ceiling.

• Orchid’s Jan 2026 update tripled power for some diseases in one release. By 2030, today’s scores will look primitive.

• Cost impact: pure software — each generation better at zero added cost.

Germline editing: The endgame. Base and prime editors already make precise single-nucleotide changes without double-strand breaks.

• Once mature for germline use, you fix unfavorable variants directly instead of selecting among imperfect options. Not approved yet.

• The science is moving faster than the ethics committees.

• Cost impact: eventually replaces selection — editing one embryo is cheaper than screening ten.

Human artificial chromosomes (HACs): The wildcard.

• Synthetic chromosomes added to the genome without altering existing DNA. A single HAC could carry dozens or hundreds of beneficial variants (disease resistance, enhanced DNA repair, novel metabolic capabilities) in one heritable package.

• No off-target effects on natural chromosomes, theoretically unlimited payload. HACs have been maintained in human cell lines for years; the barrier is reliable germline transmission.

• If solved, one HAC insertion could deliver more genetic benefit than selecting from a thousand embryos. Most speculative technology on this list — and potentially the most transformative.

• Cost impact: Unknown, but the per-variant cost approaches zero once the delivery platform works.

Somatic gene therapy: The fallback for anyone already born. LNP delivery, Anellovirus delivery, AAV vectors, and in vivo base and prime mods to upgrade health and performance. Children from screened embryos will grow up in a world where they can also opt for somatic upgrades in adulthood (patches or additional enhancements). Selection sets the floor; editing raises the ceiling; somatic therapy fills the gaps.

The discourse has shifted from “should we?” to “how do we do this well?”

Arthur Liu did this 20 years ago with donor catalogs and intuition. He produced an Olympic gold medalist.

The tools now are incomparably more powerful — and in 5 years they’ll cost a fraction of today’s price. Those who act first are the ones whose children will define the next generation.

My prediction: U.S. wins it 4-3… fading the AIs. The Hughes Bros x Tkachuk Bros overcome the drought of zero gold medals for the U.S. since 1980. Oddly enough if they win, they’ll have won on the same “date” (Feb 22) as the “Miracle On Ice” game in 1980. My logic is that Canada has more talent on their top line… and Celebrini… but the U.S. has may have slightly better depth and goaltending… and Crosby is questionable. It’ll probably come down to PPs and PKs… and reffing (which at times has been highly sus at these Olympics… many phantom calls last game against the U.S.).

The U.S. had a tough QF match against Sweden (a lethal roster for a country of its size) — winning 2-1 in OT. Zibanejad (a machine in the NHL) tied it 1-1 with under 2 min. after the Swedes pulled Markstrom. Hughes won it for the U.S. a few minutes into OT (the Hughes Bros have been skating circles around everyone this tournament… including their own team… to me they look as good as it gets so far in this tourney).

The U.S. trounced Slovakia1 in the SF match with a 6-2 scoreline… pure obliteration… and Slovakia played pretty chippy… random tackles, neck elbows, check from behinds, etc. A half-assed brawl near the end because the Tkachuk bros said Why not? And the Americans end up in the final against Canada… Tage Thompson a little dinged, but should be good to go.

How has Canada looked? Pretty good, but not as elite as expected. But they are still in the gold medal game… and they are team Canada… and they have arguably the 2 best players in the world (McDavid and MacKinnon). They’ve had some challenges en route… sneaking past Czechia (a solid team) 4-3 in the QF — then barely eking out a win against Finland2 in the SF; trailing most of the game (0-2, then 1-2, tie goal in the third, game winner with ~35 sec. to go from MacKinnon).

From my judgment (even if it didn’t reflect the final standings), the 4 best teams in the 2026 Olympics were: (top 2) U.S., Canada → (3) Finland → (4) Sweden. Followed by some combo of Czechia, Slovakia, and Switzerland. The rest couldn’t really hang.

Really sucks we didn’t get to see Russia lace up.

Also disappointed we didn’t get to see larger-than-NHL-sized Olympic sheets… but I guess it’s not that big of a deal. Still crazy that this is somehow smaller than an NHL rink.

I’m not a fan of the formatting for Olympic Hockey either… would be better if somehow a double-elimination and/or at the very least an axing of the lowest seeds within each group from the prelims. Italy, France, Latvia… had no biz playing more than 3 games. But I guess the 1-and-done after the prelims increases chaos and excitement… so maybe it’s fine.

Anyways… U.S. vs. Canada for “best-on-best” hockey is the “main event” of the 2026 Winter Olympics3.

Gotta love the pre-game energy from Brady Tkachuk on the U.S. vs. Canada rivalry:

“There’s hatred there.”

Milano Santagiulia Arena — Sunday, February 22, 2026 — 2:10 PM CET (8:10 AM ET)

THE CALL FROM AIs

• Winner: Canada

• Final score: Canada 3, USA 2 (OT)

• Win probability: Canada 51-53% / USA 47-49% (essentially a pick’em with a slight Canada lean)

• Implied moneyline: Canada −113 / USA +113

• Total goals: 5

• Over/under: Under 5.5 (70% probability) — Hellebuyck has been elite and Binnington has been solid and timely, but the tournament's best PP and 3-on-3 OT format are wild cards

• Confidence: 6/10 — small edges and injury uncertainty drive nearly all the spread

Scenario Breakdown

1. Canada wins in overtime — 33% (+203)

2. USA wins in overtime — 25% (+300)

3. USA wins in regulation — 22% (+355)

4. Canada wins in regulation — 20% (+400)

Total Goals Distribution

• 5 goals (most likely) — 30%

• 4 goals — 25%

• 6 goals — 20%

• 3 or fewer — 15%

• 7 or more — 10%

• Under 5.5: ~70% / Over 5.5: ~30%

PATH TO THE FINAL

Team USA (5-0-0, #2 seed) — Full schedule

1. vs. Latvia, W 5-1 — Brady Tkachuk opened scoring; Brock Nelson scored twice with Jack Hughes assisting on both; Tage Thompson scored a PP goal; Auston Matthews added a PP goal in the third. Hellebuyck made 17 saves on 18 shots. Ten different players got on the scoresheet. Dominant after a 1-1 first period.

2. vs. Denmark, W 6-3 — USA’s tightest game of the prelims. Trailed 2-1 midway through the second before Jack Eichel produced a goal and assist in a 57-second span to flip the game. Brady Tkachuk, Matthews, Guentzel, and Werenski also scored. Proved the Americans could come from behind when tested.

3. vs. Germany, W 5-1 — Matthews scored twice and added an assist, answering questions about his pre-Olympic slump (1 goal in his final 8 Leafs games). Faber, Werenski, and Thompson also scored. Hellebuyck made 23 saves. USA clinched Group C and the #2 seed with a +11 goal differential.

4. QF vs. Sweden, W 2-1 (OT) — The gut check. Larkin scored in the second (both Hughes brothers assisted). USA nursed the 1-0 lead until Sweden’s Mika Zibanejad tied it with the goalie pulled and 1:31 left in regulation. Quinn Hughes won it 3:27 into 3-on-3 OT, carrying the puck into the zone and beating Markstrom from the slot. Hellebuyck was outstanding with 28 saves against a Sweden team that dominated the third period. Markstrom made 38 saves in a losing effort.

5. SF vs. Slovakia, W 6-2 — Total demolition. Larkin scored 4:19 into the first; Thompson made it 2-0 on the PP with 41 seconds left in the period. Jack Hughes took over the second with two goals, including a shimmy-shake deke for a highlight-reel roof job. Eichel and Brady Tkachuk also scored. USA built a 5-0 lead through two periods. Hellebuyck made 22 saves. Thompson left in the third as a precaution (foot).

Summary: Outscored opponents 24-8. Never lost. Faced real adversity only once (Sweden QF). Scoring has been deeply distributed — 11 different goal scorers across 5 games. Peaked at the right time with the Slovakia blowout.

Team Canada (5-0-0, #1 seed) — Full schedule

1. vs. Czechia, W 5-0 — Binnington’s 26-save shutout silenced doubters about his Olympic readiness. McDavid had 3 assists in his Olympic debut. Celebrini opened scoring late in the first (becoming the first teenager to score a GWG in an NHL-era Olympics). Stone, Horvat, MacKinnon, and Suzuki also scored. Josh Morrissey left with an injury in the second and has not returned since.

2. vs. Switzerland, W 5-1 — The McDavid-MacKinnon-Celebrini line combined for 8 points (3G, 5A). Crosby scored his first goal of the tournament. Canada’s power play connected twice. Tom Wilson’s physical play on Kevin Fiala led to Fiala leaving on a stretcher (no penalty called — a flash point). Logan Thompson started in net and made 24 saves.

3. vs. France, W 10-2 — Statement game. McDavid (1G, 2A), Celebrini (2G, 1A including a penalty shot), and Crosby (1G, 2A) each had 3 points. Wilson, Makar, Toews, Horvat, and Hagel also scored. Eight different skaters recorded multiple points. Canada finished Group A at +17 goal differential (vs. USA’s +11), earning the #1 overall seed. A fight in the third after Wilson dropped the gloves underscored Canada’s physical edge. McDavid set the record for most points through 3 games of an NHL-era Olympics (9).

4. QF vs. Czechia (rematch), W 4-3 (OT) — Canada’s first real crisis. Crosby left injured at 4:55 of the second after a Radko Gudas collision — didn’t return. Czechia took a 3-2 lead with 7:42 remaining on an Ondrej Palat one-timer. Canada looked dead. Then: Suzuki deflected a Toews point shot with 3:27 left to tie it. Binnington stopped Necas on a breakaway with 1:10 left to keep Canada alive. Marner scored 1:22 into OT, accelerating through a gap when Czech defenders shaded to MacKinnon. Celebrini had a goal and 2 assists. Binnington’s breakaway save was the play of the tournament.

5. SF vs. Finland, W 3-2 — Another comeback. Finland led 2-0 on goals by Rantanen (PP) and Haula (shorthanded). Canada scored 3 unanswered: Reinhart deflected a Makar shot on the PP; Theodore tied it with a shot through traffic; MacKinnon won it on the PP with 35.2 seconds left, burying a saucer pass from McDavid. Canada became the first team to record consecutive comeback wins in playoff games at an NHL-era Olympics. McDavid finished with 2 assists, pushing his tournament total to a record 13 points in 5 games.

Summary: Outscored opponents 27-8. Never lost. But were genuinely tested in both knockout games — trailed in the third period each time and needed late heroics to survive. Scoring concentrated at the top: McDavid (2G, 11A, 13 pts), Celebrini (5G, 5A, 10 pts), MacKinnon (4G, 3A, 7 pts). Championship pedigree showing in clutch moments.

ROSTERS AND LINE COMBINATIONS

Team USA — Full roster

Forward lines (most-used combinations during tournament — subject to in-game adjustments):

• Line 1: Brady Tkachuk (OTT) – Jack Eichel (VGK) – Matthew Tkachuk (FLA)

• Line 2: Jake Guentzel (TBL) – Auston Matthews (TOR) – Matt Boldy (MIN)

• Line 3: Jack Hughes (NJD) – Dylan Larkin (DET) – Clayton Keller (UTA)

• Line 4: Kyle Connor (WPG) – Tage Thompson (BUF) – Brock Nelson (COL)

• Extras: J.T. Miller (NYR), Vincent Trocheck (NYR)

Defense pairings:

• Pair 1: Quinn Hughes (MIN) – Charlie McAvoy (BOS)

• Pair 2: Jaccob Slavin (CAR) – Brock Faber (MIN)

• Pair 3: Zach Werenski (CBJ) – Jake Sanderson (OTT)

• 7th D: Noah Hanifin (VGK) / Jackson LaCombe (ANA)

Goaltenders:

• Starter: Connor Hellebuyck (WPG)

• Backup: Jake Oettinger (DAL)

• 3rd: Jeremy Swayman (BOS)

Note: LaCombe replaced Seth Jones (FLA), who withdrew pre-tournament with an upper-body injury.

Team Canada — Full roster

Forward lines (most-used combinations during tournament — subject to in-game adjustments):

• Line 1: Tom Wilson (WSH) – Connor McDavid (EDM) – Macklin Celebrini (SJS)

• Line 2: Brad Marchand (FLA) – Nathan MacKinnon (COL) – Nick Suzuki (MTL)

• Line 3: Brandon Hagel (TBL) – Anthony Cirelli (TBL) – Mitch Marner (VGK)

• Line 4: Sam Reinhart (FLA) – Sidney Crosby* (PIT) – Seth Jarvis (CAR)

• Extras: Bo Horvat (NYI), Mark Stone (VGK)

Defense pairings:

• Pair 1: Devon Toews (COL) – Cale Makar (COL)

• Pair 2: Travis Sanheim (PHI) – Shea Theodore (VGK)

• Pair 3: Drew Doughty (LAK) – Colton Parayko (STL)

• 7th D: Thomas Harley (DAL)

Goaltenders:

• Starter: Jordan Binnington (STL)

• Backup: Logan Thompson (WSH)

• 3rd: Darcy Kuemper (LAK)

Note: Josh Morrissey (WPG) has missed 4 straight games and is not expected to return. Sanheim has filled in on the 2nd pair.

INJURIES

• Sidney Crosby (CAN, C) — Game-time decision. Lower-body injury sustained in QF from Radko Gudas collision. Missed the semifinal vs. Finland entirely. Skated in closed practice Friday; Cooper said he has “a better chance” of playing Sunday than he did Friday. If out, Canada loses its captain, two-time Olympic gold medalist, emotional leader, and a player producing at 6 points (2G, 4A) in 4 games. This is the single largest swing variable in the game.

• Josh Morrissey (CAN, D) — Out. Undisclosed injury since the opener vs. Czechia. Missed 4 consecutive games. Was a top-4 minute-eater; Sanheim has filled in capably.

• Tage Thompson (USA, C/RW) — Probable. Left the semifinal in the third as a precaution with an apparent foot injury. His 6’6” frame and 97.94 mph shot anchor the USA power play. If limited, the PP threat drops meaningfully.

• Seth Jones (USA, D) — Out (pre-tournament). Upper-body injury; replaced by LaCombe before the Games began.

How Crosby’s status shifts the odds:

• Crosby fully healthy → Canada ~55-56%

• Crosby out or clearly limited → ~50-50 or slight USA lean

• Blended baseline (weighted probability) → Canada 53%

THE RATIONALE

1. Special Teams — The Game’s Fulcrum

This is the single variable most likely to decide the gold medal. All data from IIHF Tournament Statistics.

Power play:

• Canada: 43.75% (7-for-16), #1 in the tournament. PP1 unit: Makar quarterbacking McDavid, MacKinnon, Reinhart, Crosby/replacement. MacKinnon’s tournament-winning goal vs. Finland came on the PP with 35.2 seconds left.

• USA: 28.57% (4-for-14). Respectable but not elite. Thompson’s shot and Matthews’ net-front presence drive the unit.

Penalty kill:

• USA: 100% (15-for-15), #1 in the tournament. Slavin, McAvoy, Faber, Larkin, and Nelson run an aggressive Sullivan system that forces entries wide.

• Canada: 72.73% (8-for-11, 3 PPG allowed). The clear weak spot — Switzerland (Suter), Czechia (Pastrnak in QF), and Finland (Rantanen in SF) all scored with the man advantage.

Penalties taken (5 games):

• USA: 52 PIM (avg 10:24/game)

• Canada: 49 PIM (avg 9:48/game)

• Nearly identical — both teams are drawing ~3 power plays per game. The efficiency gap, not discipline, is what matters.

Net assessment: Canada’s PP will probably score at least once — the USA’s perfect PK is small-sample against lesser units, and Canada’s personnel is a different animal. But Canada’s leaky PK means the USA gets a PP goal too. Net special teams edge: Canada, roughly 60/40.

2. Goaltending — USA’s Strongest Card

All data from IIHF Top Goalkeepers.

• Connor Hellebuyck (USA): .947 SV%, 1.23 GAA, #1 among all tournament goalies. Has been spectacular despite a down NHL season (.900 SV%, 2.79 GAA). History of elevating internationally — .932 at the 4 Nations Face-Off, 2024-25 Hart Trophy winner.

• Jordan Binnington (CAN): .914 SV%, 1.74 GAA. Entered the Olympics struggling badly for St. Louis (8-17-6 record, among the worst goals saved above expected figures in the NHL by third-party models). But: posted a shutout in the opener, stopped Necas on a tournament-saving breakaway in the QF, and has defied analytics in every best-on-best moment (2019 Cup, 2025 4 Nations title). The gap between him and Hellebuyck is still the widest at any position in this matchup.

If Hellebuyck is elite — which he has been — USA can win 2-1 or 3-2 on defensive structure alone. This is the primary American upset pathway. Edge: USA.

3. 5-on-5 Finishing — Same Volume, Canadian Edge

From IIHF Scoring Efficiency:

• Both teams: exactly 201 shots on goal through 5 games

• Canada: 27 goals (13.43% shooting)

• USA: 24 goals (11.94% shooting)

• Both teams: 8 goals against

Same shot generation. More goals for Canada. That pattern indicates superior finishing talent and/or shot quality — consistent with having McDavid, MacKinnon, and Celebrini at the top. Edge: Canada.

4. Physicality and Style

Canada was built around Jon Cooper’s “spread the snot” philosophy — one physical, gritty player per line alongside skill:

• Line 1: Wilson (6’4”, 218 lbs — 7 fighting majors this NHL season)

• Line 2: Marchand (37, still one of the NHL’s most agitating forwards)

• Line 3: Hagel (relentless forechecker)

• Wilson’s hit on Fiala in the Swiss game and his fight in the France game set the tone. Canada imposes physicality while maintaining skill.

USA counters with the Tkachuk brothers, who are as physical as any forwards in the tournament:

• Brady (6’4”, 215) and Matthew (6’2”, 202) combine line-of-scrimmage physicality with elite finishing

• Thompson’s 6’6” frame creates net-front chaos

• Larkin and Trocheck bring sandpaper in the bottom six

Both teams run physical, North American systems. The difference: Canada’s physicality is more organized by design (one enforcer per line). The USA’s is more concentrated on the Tkachuk-Eichel line. In a single game, this is roughly a wash — but Canada’s ability to be physical on every shift, not just one line, creates cumulative fatigue. Slight edge: Canada.

5. Goal Scorers — Tournament Leaders

USA’s top production (5 games):

• Jack Hughes: 3G, 3A (6 pts) — started on the 4th line, promoted to 3rd line during QF, has been the tournament’s most dynamic American forward

• Quinn Hughes: 1G, 6A (7 pts) — a point in every game, OT winner vs. Sweden

• Auston Matthews: 3G, 3A (6 pts) — shook off a pre-Olympic slump with a dominant performance vs. Germany

• Brady Tkachuk: 3G, 2A (5 pts) — the emotional engine

• Matthew Tkachuk: 0G, 6A (6 pts) — the quiet playmaker, a point in nearly every game

• Jack Eichel: 2G, 4A (6 pts) — turned the Denmark game with a goal and assist in 57 seconds

• Dylan Larkin: 2G, 1A (3 pts) — scored in consecutive playoff games

• Tage Thompson: 2G, 1A (3 pts) — foot injury concern

Canada’s top production (5 games):

• Connor McDavid: 2G, 11A (13 pts) — set the all-time record for points in an NHL-era Olympic tournament

• Macklin Celebrini: 5G, 5A (10 pts) — teenage phenom, scored in 4 straight games, first Canadian penalty shot goal in Olympic history

• Nathan MacKinnon: 4G, 3A (7 pts) — scored the semifinal winner with 35.2 seconds left

• Mitch Marner: 2G, 3A (5 pts) — OT winner vs. Czechia

• Sidney Crosby: 2G, 4A (6 pts in 4 games) — game-time decision

• Sam Reinhart: 2G, 2A (4 pts)

• Nick Suzuki: 2G, 2A (4 pts) — game-tying goal in QF with 3:27 left

Canada’s scoring is top-heavy — McDavid, Celebrini, and MacKinnon account for 30 of Canada’s 51 total scorer points (goals + primary assists). The USA’s scoring is more distributed across 11 different goal scorers. In a single game, top-heavy can mean either devastating (if McDavid has a 3-point night) or vulnerable (if the USA shuts down 2-3 players). Edge: Canada on ceiling, USA on floor.

6. Overtime Format — Structural Canada Advantage

If tied after 60 minutes: continuous 3-on-3 sudden death (20-minute periods until a goal, no shootout). This format magnifies individual skating and high-danger conversion.

NHL EDGE data on McDavid: 24.61 mph max speed, 494 speed bursts above 20 mph this NHL season (both league-leading). In 3-on-3 open ice, he is the most dangerous weapon in hockey history.

This isn’t theoretical: McDavid scored the OT winner against this exact opponent in the 2025 4 Nations Face-Off final. If the game reaches OT tied at 2, Canada’s win probability spikes to ~60%. Edge: Canada.

7. Momentum and Semifinal Paths

USA enters with supreme confidence. The 6-2 demolition of Slovakia was their best game of the tournament — complete in every phase. 11 different players recorded points. Hellebuyck was barely tested (22 saves). The only concern: Thompson’s precautionary exit. The Americans peaked at exactly the right time.

Canada enters battle-tested but banged up. They’ve trailed in both knockout games and won — the first team to record consecutive comeback wins in an NHL-era Olympic playoff. That resilience is the hallmark of a championship team. But they’ve also looked genuinely vulnerable early in games, needed breakaway saves from Binnington to survive, and lost Crosby. They’ve won ugly in the playoffs while dominating in the prelims.

The question: does USA’s dominant blowout or Canada’s comeback resilience translate better to a gold medal game? History says the team that has been tested and survived is better prepared for adversity. Canada has faced it. The USA has not, except for 20 minutes against Sweden. Slight edge: USA on confidence, Canada on preparedness.

8. Historical Pattern

A striking regularity: the USA wins the group game, Canada wins the gold.

• 2010 Olympics: USA 5, Canada 3 in prelims → Canada 3, USA 2 (OT) in gold medal game (Crosby golden goal)

• 2025 4 Nations: USA 3, Canada 1 in round robin → Canada 3, USA 2 (OT) in final (McDavid OT winner)

• Both OT winners: Canadian generational superstars

The broader ledger per NHL.com’s historical breakdown:

• This is the third USA-Canada gold medal game in the NHL-player era (Canada won in 2002 and 2010 — they are 2-0)

• Canada has won 4 consecutive best-on-best international tournaments (2010, 2014, 2016 World Cup, 2025 4 Nations)

• The USA has not won a best-on-best tournament since the 1996 World Cup — a 30-year drought

• The USA has not won Olympic gold since the 1980 Miracle on Ice

• This game falls on February 22 — the 46th anniversary of the Miracle on Ice

Small N. Not statistically predictive. But 15 Stanley Cup winners vs. 4, and the only players on either roster with Olympic gold are Crosby and Doughty (both 2010 and 2014). Championship composure in the final 10 minutes is real. Edge: Canada.

9. Venue and Rules

• IIHF rules on near-NHL-sized ice (60m × 26m, virtually identical to 200 ft × 85 ft). Favors both North American teams.

• Fighting = Fighting = automatic ejection under IIHF zero-tolerance.

• Hybrid icing (with IIHF linesmen generally calling it earlier than NHL officials).

• OT: 3-on-3 sudden death, no shootout.

• Both teams arrived ~Feb. 8; jet lag is a non-factor by day 14. The 2:10 PM local puck drop = afternoon energy, no fatigue.

• NHL paused Feb. 5, resumes Feb. 25 (3 days after the final).

FACTOR SUMMARY

High-weight variables:

• Power play → Canada (43.75% vs 28.57%)

• Penalty kill → USA (100% vs 72.73%)

• Goaltending → USA (Hellebuyck .947 vs Binnington .914)

• Top-end talent → Canada (McDavid, MacKinnon, Makar)

Medium-weight variables:

• 5v5 finishing → Canada (13.43% vs 11.94%)

• Forward depth → USA (deeper bottom-6, 11 different scorers)

• Physicality → Canada (organized per-line physicality, Wilson/Marchand/Hagel)

• Championship DNA → Canada (15 Cup winners vs 4)

• Historical precedent → Canada (2-0 in NHL-era gold medal games, 4 straight best-on-best titles)

• OT format → Canada (McDavid in open ice)

• Injuries → USA (healthier; Crosby game-time decision)

Low-weight variables:

• Momentum → USA (6-2 SF blowout vs Canada’s comebacks)

• Coaching → Even (Sullivan and Cooper, both 2× Stanley Cup champions)

AI PREDICTION (CLAUDE + CHATGPT)

Canada 3, USA 2 (OT). Canada’s power play, finishing talent, overtime weaponry, and comeback resilience slightly outweigh America’s goaltending advantage and depth — unless Crosby is out and Hellebuyck is dominant, which is the clearest American path. The near-pick’em line reflects a genuine truth: there is no clear favorite. But when the margin is this thin, history, experience, and McDavid tip the scales toward Canada completing a fifth consecutive best-on-best championship.

The claim that American crime rates have plunged to historic (even 250-year) lows has become conventional wisdom. I don’t think many dispute the idea that homicides are massively down relative to historic estimates.

But the extension to “all crime” is probably overstated — perhaps to a significant degree. A legitimate decline in “lethal violence” might be glossing over: (1) a broken measurement apparatus, (2) an explosion of unmeasured criminal victimization, and (3) a society that has adapted to danger rather than eliminating it.

Crime has three ledgers, and “historic lows” only refers to one of them.

1. Ledger 1: Legacy street crime. Homicide, robbery, burglary, etc. Measured by police reports and partly by the NCVS.

2. Ledger 2: Modern predation. Fraud, cyber, identity theft, check fraud, business email compromise, ransomware. Measured through complaints, financial systems, and private security — not UCR headlines.

3. Ledger 3: Administrative and underground-economy crime: Tax evasion, benefits fraud, document/ID fraud, immigration offenses, off-the-books labor.

It is completely possible for Ledger 1 to improve while Ledgers 2 and 3 explode. “Crime is at historic lows” is a claim about Ledger 1 only, and even there, the pipeline is noisier than most people admit.

One more ledger that never shows up in “crime is down” headlines: drug poisoning. In 2024, the U.S. recorded 79,384 overdose deaths — mass lethal harm downstream of an illegal market; yet it’s rarely counted as “crime” in the way people mean when they say “historic lows.”

Homicide really did fall. No denying it.

The FBI’s 2024 murder rate of 5.0 per 100,000 is a legit substantial decline from the pandemic peak of 6.8 in 2020 and the early-1990s peak of ~9.8.

Homicide is the hardest crime to undercount; bodies generate independent records through medical examiners, and clearance rates remain far higher than for other offenses.1

The decline is corroborated internationally: the UK, Germany, and Australia all show parallel drops, and a 166-country study found common crime declining across Europe, Australasia, and most of Asia from 2006–2019 (van Dijk, Nieuwbeerta & Joudo Larsen, 2021). Something genuinely structural is occurring, not merely a U.S. statistical artifact — and we should acknowledge it.

A few precision caveats on “250-year-low" claims, none of which undermine the directional story.

1. The 20-year low in the FBI’s own published data is 4.4 in 2014, not 2024. The estimated ~4.0 rate for 2025 comes from the Council on Criminal Justice — a reputable source tracking ~35 large cities that showed a 21% homicide decline — but it remains an extrapolation to the nation, not a published national statistic.

2. The overall trajectory is downward and may well reach a modern-era record. As for the “250-year low” comparison: the leading historian of pre-modern American homicide, Randolph Roth, has spent decades assembling the Historical Violence Database from court records, newspapers, and coroner files.

3. His data actually shows periods (particularly stable Northern communities in the mid-18th and early 19th centuries) with rates that may have been comparably low. But the coverage is geographically patchy and drawn from small, unrepresentative communities; you cannot make a confident national statement before systematic collection began in 1928.

4. The “250-year low” claim may ultimately prove correct, but it rests on data that doesn’t yet support the precision it implies. “Lowest in the modern era” is solidly defensible.

Murder declining does not mean all crime declined.

Nonfatal violence is higher

Homicide is one subset of one ledger. You can have murders down while assaults rise.

You can have “legacy street crime” down while fraud, identity theft, administrative looting, and underground-economy illegality explode. The inference “murders are low, therefore crime is low” only works if the measurement system covers the main harm channels but it doesn’t.

• For nonfatal violence, the government’s own National Crime Victimization Survey shows 2024 rates higher than both 2015 and 2020.

• Urban violent victimization is up 61% since 2019 (34.0 vs. 21.1 per 1,000). Youth (12–17) victimization surged 37% in a single year (2023–2024).

• To be fair: NCVS violent victimization remains well below the early 1990s peak; the long-run decline is notable.

• But the NCVS does not show “historic lows” for nonfatal violence; it shows post-2019 increases, particularly in cities and among young people, that the “all crime is collapsing” narrative simply ignores.

Factors that drove the crime rate down

Before highlighting what the statistics completely miss, it is worth noting some of the forces that have likely contributed to lower crime.

The dominant drivers likely differ by era (the factors that ended the 1990s crises may not be the same as those driving the decline in 2025, but there’s overlap and maintenance effects).

The 1990s crash (murder rate ~9.8 → ~5.5)

• End of the crack epidemic: Crack markets stabilized and the territorial violence around them subsided. Widely considered the single largest driver of the 1990s homicide drop.

• Policing innovations: CompStat (1994), hot-spots policing, and early precision targeting coincided with large drops in adopting cities. Levitt (2004) ranks increased police numbers among the top factors. The international crime drop in countries without these innovations complicates causal attribution, but the U.S.-specific evidence is strong.

• Mass incarceration: The prison population rose from ~300,000 (1980) to over 2.1 million at peak. The Brennan Center estimates incarceration explains ~5% of the 1990s decline; Levitt (2004) estimated ~12%. The mechanical effect on chronic offenders (a small population responsible for most crime) was substantial. The effect is thought to have diminished sharply after 2000.

• Abortion legalization: Donohue & Levitt (2020) estimated legalized abortion reduced violent crime by 47% from 1991–2014, with the effect strengthening in out-of-sample data. The 2020 replication with 17 additional years gave it renewed empirical weight, but the identification strategy remains contested (Joyce 2006; Foote & Goetz 2008).2

• Early background factors: Declining alcohol consumption (down ~20% since the early 1980s) and the beginning of a broad expansion in mental health treatment access.

The 2000s–2010s grind (murder rate ~5.5 → 4.4 in 2014)

With crack and incarceration having done their heaviest lifting, a different set of forces took over: overwhelmingly (1) opportunity-reduction and (2) behavioral — rather than criminal justice.

• Security and target hardening (strongest for property crime): This is arguably the most underappreciated top-tier driver, especially for the broad property crime drop. Vehicle immobilizers and electronic security cut auto theft dramatically. Farrell et al. (2014) argue the “security hypothesis” is the single explanation that passes standardized cross-national tests. Household security followed a similar trajectory: alarm systems, deadbolts, controlled-access housing. The Community Associations Institute estimates ~369,000 homeowner associations covering ~77 million residents (~33.6% of U.S. housing); much of it in planned or gated communities with cameras, controlled access, and contracted security. Private security outnumbers sworn law enforcement nearly 2-to-1 (1.2 million guards vs. fewer than 700,000 police officers). By the 2020s, 20% of U.S. internet households had video doorbells (up from 4% in 2017), adding a layer of private surveillance that raises identification risk for offenders. This is a real reduction in successful crime, but it’s a society buying down its crime rate through massive private investment, not one that has eliminated criminal motivation.

• Reduced exposure (the umbrella mechanism): Routine activities theory (Cohen & Felson, 1979) holds that crime requires the convergence of a motivated offender, a suitable target, and the absence of a capable guardian. Multiple forces converged to keep people (especially young men) out of high-risk public space.

  • Screen time exploded from minimal internet use to 7+ hours daily on devices; Cunningham, Engelstätter & Ward (2016) found violent video game releases associated with decreases in violent crime, consistent with an incapacitation effect.

  • Youth alcohol use collapsed: Monitoring the Future reports 12th-grader past-year drinking fell from 75% (1997) to 42% (2024), and ~35% of violent victimizations historically involved an offender who’d been drinking; fewer intoxicated young men in public mechanically reduces violence.

  • More young adults are living with parents; Pew reports roughly 1 in 3 adults ages 18–34 as of 2021, meaning less unsupervised nightlife exposure and more informal guardianship. These are all sub-channels of one mechanism: less time in the places where crime happens.

• Demographic aging: Median age rose from 35 (2000) to 38 (2019), continuing to shrink the crime-prone young male cohort.3

• Background factors (plausible but lower confidence as macro drivers)

  • Fitness culture: Gym memberships rose from 32.8M (2000) to 64.2M (2019), and meta-analyses find exercise reduces aggression (g = −0.53); directionally plausible but hard to quantify at population scale.

  • Obesity: Doubling from ~23% to ~40% may have a counterintuitive incapacitation effect (Kalist & Siahaan, 2013).

  • Mental health treatment expansion: ~11.4% of adults on antidepressants by the 2010s, 1 in 5 receiving treatment correlates with the decline, but the best individual-level evidence from Fazel et al. (2015) and ~856K Swedish SSRI users found SSRIs linked to a modest increase in violent crime (HR=1.19), so the mechanism is unclear; broader treatment access may help via clinical contact and psychosis management rather than via medication effects.

  • Sustained economic expansion and low unemployment contributed as background conditions.

The 2020–2025 spike and recovery (murder rate 4.4 → 6.8 → ~4.0)

The pandemic years are the most revealing. Homicide spiked ~30% in 2020 (the largest single-year increase ever recorded) driven by COVID disruptions, the post-Floyd police pullback, and the collapse of community supervision and courts.

The fact that the spike reversed almost as fast (2022–2025), tells us the post-2020 decline is substantially a reversion to trend after a disruption, not new structural improvement. What sustained and deepened the recovery?

• Routine activities normalization + remote work: As lockdowns ended, the opportunity structure for crime partially normalized, but remote work persisted for tens of millions, permanently reducing street exposure. McConnell (2024) finds WFH is associated with reduced burglary, consistent with a “guardian at home” effect. This is primarily a property crime mechanism; the homicide effect is less direct.

• Focused deterrence: NYPD’s precision zones produced a 40% decline in shootings in 2025 — a strong case study, though one city’s results should not be extrapolated into a national causal claim. More broadly, cities that adopted focused deterrence strategies targeting high-risk groups saw the fastest recoveries.

• Continued time-use absorption: Fitness memberships surged to a record 77 million (2024); up 20% from pre-pandemic. And screen time continued absorbing idle hours. These are maintenance forces keeping the floor low.

• Demographic momentum: Median age hit 39 in 2025. The aging effect is slow but cumulative and never reverses.

The factors driving today’s numbers are overwhelmingly time-use and behavioral, not the criminal justice levers that dominated the 1990s.

A caveat on “lead”: While individual studies estimated a 7–28% crime reduction from leaded gasoline phaseout, the first meta-analysis of the lead-crime literature (Higney, Hanley & Moro, 2022) found significant publication bias across 542 estimates from 24 studies, with the pooled partial correlation falling to r = 0.16 after adjustment, and further standard corrections push the estimate toward zero. The cross-national correlation is suggestive, but the causal estimate appears substantially inflated by selective reporting.

These are legitimate anti-crime forces.

The question is whether the remaining decline (the gap between what these factors predict and what the statistics show) — is real or partly an artifact of the broken measurement pipeline.

Blind spots in measuring crime; official stats may be misleading

The FBI’s crime data suffered a major comparability break during the 2021–2022 NIBRS transition, when population coverage cratered from ~95% to ~65%.

Coverage has recovered to 95.6% for 2024, but that rapid improvement itself complicates trend comparisons across the early 2020s, since the system is increasingly comparing estimates derived from very different coverage bases and modeling assumptions.

And as of mid-2024, the 5 most populous states combined sat at just ~48% participation (by CRS’s measure), meaning national estimates still depend far more on statistical imputation than analysts acknowledge.

Beyond coverage gaps, there are documented cases of agencies systematically manipulating their own data, and these continue into the present:

• DC Metropolitan Police (2025): A draft DOJ review found a “culture of coercive fear” under Chief Smith led to systematic misclassification. Both DOJ and Congress opened investigations. The DC police union chief publicly challenged the accuracy of reported statistics — the same data showing violent crime “down 35%” and used to argue against federal intervention.

• LAPD (2005–2012): LA Times found ~14,000 aggravated assaults misclassified as minor incidents, lowering reported violent crime by 7%.

• Chicago PD (2013): Chicago Magazine uncovered murders reclassified as non-criminal deaths, with officers describing pressure to use “magic ink” that made cases disappear.

• NYPD (2010–2012): Officer Schoolcraft recorded supervisors ordering officers to refuse reports and recode felonies. An internal report confirmed the manipulation.

• Wisconsin (statewide audit): A DOJ review of sexual assault reporting found only ~67% of incidents were correctly classified for UCR — meaning a third of sex offenses were being miscategorized or unreported to the federal system.

• Austin PD: A Texas DPS audit found the department improperly cleared rape/sexual assault cases using “exceptional clearance” designations, artificially inflating closure rates.

Note: These aren’t “random anecdotes”; they are the detected subset a.k.a. the small fraction of jurisdictions where audits, whistleblowers, lawsuits, or media scrutiny forced the issue into daylight. Most departments are never audited at this depth so the correct takeaway isn’t “rare” but “we don’t really bother measuring it.”

These span 6 jurisdictions across 2 decades: America’s 3 largest police departments, the nation’s capital, a statewide audit, and a state DPS investigation.

The incentive structure: political pressure to show declining crime, CompStat-style metrics, career advancement tied to numbers — is not incidental; it is structural.

NIJ-published research on police performance measurement explicitly discusses how systems designed to reward crime reductions can distort recording incentives.

The UK has the same problem: the Office for National Statistics stripped police-recorded crime of its "National Statistics" designation in 2014 over recording practice concerns. A House of Commons committee found accumulating evidence that the underlying recording practices were unreliable; some of this is deliberate manipulation, some is training and classification complexity. Either way, it breaks comparability over time. The vast majority of misclassification, downgrading, and refusal to take reports is never audited. We only see the visible portion of the iceberg.

The 2 main US data systems increasingly contradict each other.

In 2022, the FBI reported violent crime down 2% while the NCVS showed serious violent victimization up 75%. The driver: reporting to police dropped from 52% to 48%. Property crime reporting has collapsed to just 31%; meaning police see fewer than one in three property victimizations.

Even motor vehicle theft (supposedly the “clean” category where everyone reports for insurance), only reaches about 75% reporting in the NCVS, not the ~100% commonly assumed.

The NCVS (the government's check on police-reported statistics) has its own growing blind spots: BJS reports household completion rates of ~60%, down from historical norms (~80-90%), raising questions about whether the hardest-to-reach populations (transient, undocumented, high-crime neighborhoods) are adequately sampled.

The math on reporting alone: if violent crime reporting drops from 50% to 42% with actual victimization constant, official statistics show an apparent 16% decline that is entirely illusory. Even motor vehicle theft sends contradictory signals between systems: FBI shows it down 19.4% while NCVS shows it up 47% since 2020.

Enforcement withdrawal and the three-pronged effect

Total US arrests fell roughly 25% from 2006 to 2022. California’s total arrests hit their lowest level since 1971; lower than when the state had half its current population.

Response times have ballooned: NYC response times to crimes-in-progress rose 22–46% across priority tiers between 2018 and 2023.

New Orleans' tripled, from 51 to 146 minutes.

Slower response times directly reduce the chance of solving crimes: a causal study of 300,000 incidents found a 10% increase in response time produces a 4.7-percentage-point drop in clearance rates (Vidal & Kirchmaier, 2018).

As crime analyst Jeff Asher has noted:

"Longer response times reduce the likelihood that an incident will be successfully reported to police, which in turn leads to crimes being undercounted."

Three distinct mechanisms suppress the statistics while leaving underlying criminal behavior untouched:

• Legislative decriminalization and legalization: Statutes can redefine “crime” downward. Since 2000, at least 37 states have raised felony theft thresholds, meaning the same act can move felony → misdemeanor (or citation) purely by legislative relabeling. And legalization deletes entire arrest streams. In 2007 the FBI estimated ~1.8M drug-abuse-violation arrests; 42.1% were marijuana possession and 5.3% marijuana sale/manufacturing; meaning roughly ~850k+ marijuana arrests disappear from “crime” metrics as law changes. When the legal definition of “crime” shrinks, measured crime mechanically falls even if underlying behavior doesn’t.

• Enforcement withdrawal: Fewer encounters, fewer arrests, fewer reports — and fewer reports mean fewer “crimes known to police.” The causal chain is mechanical.

• Prosecutorial declination: When victims see cases dismissed: (1) Philadelphia’s DA cut total prosecuted criminal matters by ~50%; (2) Chicago’s Foxx raised felony dismissal rates to 29.9% — reporting likely drops because expected return on reporting (ROR) drops. Police effort declines for the same reason. The downstream effect on recorded crime is indirect but real.

Crime migrated online, and the toll is invisible

The FBI’s index crime categories were designed in 1929: burglary, larceny, auto theft, robbery, assault, murder, rape, arson. The criminal economy has moved on.

What follows is a partial inventory of victimization that does not appear in “crime is at historic lows” headlines.

1. Digital and cyber-enabled crime

The fastest-growing category, now larger than all traditional property crime combined.

• Total IC3-reported losses: $16.6 billion in 2024, up 33% from 2023 and +1,400% since 2015. Cyber-enabled fraud alone — 334K complaints — accounted for $13.7B, or 83% of all reported losses.4

• Investment fraud (predominantly cryptocurrency “pig butchering” schemes): $6.57B in losses, the single largest category. Crypto-related complaints totaled 150K with $9.3B in losses — up 66% YoY.

• Business email compromise (BEC): $2.77B in losses ($8.5B over the last three years). Companies wiring funds to spoofed vendor accounts — essentially digital robbery at scale.

• Tech support / customer support scams: $1.46B, overwhelmingly targeting the elderly.

• Elder fraud (all categories): $4.9B in reported losses, up 43% YoY. Over 7,500 victims aged 60+ lost more than $100K each.

• Identity theft: 490K → 1.135M FTC reports. Personal data breach complaints: 64,882 to IC3 alone, with $1.5B in losses.

• Ransomware: Complaints up 9%, with 67 new variants detected in 2024. The reported $12M in losses is a fiction — it excludes lost business, downtime, and the majority of quietly paid ransoms. Actual ransomware costs likely exceed $1B annually.

• Sextortion: 54,936 complaints, up 59% YoY — increasingly AI-generated, increasingly targeting minors. The FBI referred 42 victims to suicide intervention specialists.

These figures are themselves massive undercounts. IC3 acknowledges its numbers reflect only what victims choose to report. Corporate cyber theft, IP theft, and ransomware payments companies quietly absorb push the true scale likely past $50–100B annually.

2. Physical crime that moved off the books

• Retail shrinkage: NRF's own National Retail Security Survey puts shrink at $112.1B and reports that internal + external theft account for ~65% of shrink — the majority. Retailers report closing locations, cutting hours, and altering product selection as a direct response to retail crime: 7,300+ store closures in 2024 (up 58% YoY), entire neighborhoods losing pharmacy and grocery access. And because shrink is a loss accounting measure, it captures theft that never becomes a police report.

• Auto theft / insurance fraud: Auto insurance premiums up ~51% since 2019. Catalytic converter theft, VIN cloning, and title washing are rampant but poorly captured in index crime statistics. The NICB reported 850,708 vehicles stolen in 2024; direct, insured property crime at massive scale.

• Insurance fraud (all categories): The Coalition Against Insurance Fraud estimates $308.6B/year in total insurance fraud costs, per NAIC — a figure that dwarfs all index property crime combined. The FBI estimates this costs the average family $400–700 annually in premium impact. None of it appears in “crime is down” statistics.

• FTC consumer fraud: $12.5B in reported losses (2024), spanning romance scams, impersonation fraud, and fake investment platforms.

• Mail theft–enabled check fraud: FBI and USPIS warn check fraud is surging, with suspicious activity reports (SARs) nearly doubling from 2021 to 2023. FinCEN’s BSA-based analysis (Feb–Aug 2023) flagged 15,417 related reports and >$688M in linked transactions, just from the slice that hits regulated reporting channels.

• Cargo theft + supply-chain crime: CargoNet recorded 3,625 incidents in 2024, up 27% year over year; property crime increasingly shifting to logistics nodes rather than street confrontations. Traditional methods like trailer burglaries and full trailer theft remained prevalent.

• Package theft / “porch piracy” (last-mile): USPS OIG estimates at least 58 million packages were stolen in 2024, with as much as $16 billion in reported losses. It also stresses there is no single authoritative source because reporting is fragmented and hard to verify; meaning this is almost certainly undercounted. Package theft can occur at multiple points in the shipping chain, but the OIG focuses on the unauthorized removal of packages from delivery locations (residences/businesses).

3. Government fraud

The largest category by dollar volume, and the most completely invisible to crime statistics:

• Total federal fraud: The GAO estimated $233–521B annually in fraud across federal programs. Separately, the government documented $162B in improper payments in FY2024 alone ($2.8T cumulative since 2003) — improper payments are not all fraud, but they represent the exposure base from which fraud draws, and neither figure is an audited total. Both are estimates extrapolated from the fraction of transactions that get reviewed. Most federal programs lack the investigative capacity to detect sophisticated fraud at scale, so GAO’s figures represent a floor, not a ceiling.

• Medicaid/Medicare fraud: Minnesota’s Medicaid scandal — $9B+ in suspected fraud across 14 programs in a single state, the scheme attracting “fraud tourists” from across the country — was only discovered because its sheer scale made it impossible to ignore. GAO estimates healthcare fraud nationally at $100B+, but that estimate is itself derived from limited audit samples.5 Every state runs the same programs with the same weak controls; Minnesota is not an outlier so much as the one that got caught.

• Pandemic relief fraud: An estimated $100–200B in fraudulent PPP, EIDL, and unemployment claims. The SBA Inspector General identified $200B in potentially fraudulent EIDL loans alone. Programs were designed for speed over verification, and most fraud will never be recovered or even fully quantified.

• Tax noncompliance: The IRS projects a $696B gross tax gap for Tax Year 2022, with $539B (77%) from underreporting on timely filed returns. Willful tax evasion (26 U.S.C. §7201) is a felony, but enforcement capacity has been gutted; the criminal component alone is enormous but nearly invisible to crime statistics.

The pattern across all three categories is the same:

1. Documented numbers represent the visible fraction of a problem whose true scale is structurally unknowable, because agencies responsible for disbursing funds lack the capacity (and often the political incentive) to measure how much is being stolen.

2. When DOGE or similar government oversight efforts periodically surface new fraud, the reaction is surprise, but the surprise itself is evidence that the baseline estimates were too low.

3. None of this stuff: the $16.6B in cyber losses, the $112B in shrinkage, the $233–521B in federal fraud — appears in the FBI statistics used to declare crime at “historic lows.”

The 1929 index measures muggings while the economy is being looted through fiber optic cables.

Has total criminal victimization actually declined?

You can’t collapse all ledgers into one neat “crime rate” because they’re measured in different units (incidents vs. dollars vs. administrative loss).

But you can compare orders of magnitude of harm, and the uncounted ledgers are now enormous.

The FBI reported $15.1 billion in property crime losses in 1995; roughly $31 billion in 2024 dollars.

In 2019 (the last year the FBI published this figure), the nominal total was $15.8 billion; essentially flat in real terms despite a 40% decline in reported property crime incidents.

Then add what the index excludes: $16.6B in IC3-reported cyber losses, $12.5B in FTC consumer fraud, $112B in retail shrinkage, $4.9B in elder fraud alone, and $233–521B in estimated federal program fraud.

Anderson (2021) estimated the aggregate cost of crime at $4.71–5.76 trillion annually when including law enforcement, private deterrence, and intangible victim costs; roughly 20–25% of GDP.

And Anderson’s inputs predate the post-2020 explosion in cybercrime and government fraud so the estimate range may be lowballing.

The conventional crimes that generate “historic lows” headlines account for a small and shrinking share of total criminal harm; crime didn’t decline so much as migrate to channels the measurement system was never designed to see.

The immigration-sized hole

There are 14 million unauthorized immigrants (~4% of the population) concentrated in sanctuary jurisdictions with the most aggressive decriminalization and de-policing.6

These 14+ million people are structurally deterred from calling 911 when victimized: fear of deportation, language barriers, distrust of authorities. Domestic violence, sexual assault, and labor exploitation within these communities are notoriously undercounted.

The identity fraud dimension alone is enormous:

• SSA’s Earnings Suspense File: Where W-2s with mismatched name/SSN combinations are held, has accumulated over $2.4 trillion in cumulative unmatched wages since 1937, with the file growing by tens of billions annually. Not all mismatches involve unauthorized work (employer errors and name changes contribute), but SSA explicitly acknowledges that invalid or stolen SSNs used by unauthorized workers are a major driver.

• Employment-related identity fraud: GAO (2020) identified 1.3 million SSNs in a single year (2016) showing signs of fraudulent use for employment purposes.

• Underground labor: A 2023 report estimates 1.1–2.1 million construction workers alone are misclassified or paid off-the-books (~10–19% of the workforce), shrinking tax compliance and placing worker injuries outside OSHA reporting.

Meanwhile, the FBI’s crime reporting system does not collect immigration or citizenship status of offenders or victims, the NCVS does not ask documentation status, and GAO (2024) confirmed there are no comprehensive citizenship data across state prisons and local jails.

The claim that “immigrants commit less crime” may be directionally correct for first-generation legal immigrants, but it is overconfident relative to what the main national data systems can actually observe for the unauthorized population specifically.

And as the National Academies has documented, crime rates among second- and third-generation immigrants converge toward native-born levels; so first-generation snapshots cannot settle multi-generational policy questions when the “immigrant” category disappears into “native-born” in most statistics after one generation.

The fog is structural, and it biases the data downward.

Behavioral avoidance and deterrence

Much of the measured decline reflects behavioral avoidance, not the elimination of danger.

• Gallup found 40% of Americans afraid to walk alone at night in 2023; a three-decade high, though declining to 31% by late 2025.

• Millions of people live in walled or gated communities.

• Remote work eliminated commuting exposure for millions.

A low crime rate can reflect a safer environment and/or a population that has learned to avoid exposure — and the “official crime statistics” don’t distinguish between them.

Deterrence spending tells the same story from the other side: the private security industry alone is worth roughly $180 billion and private security guards now outnumber public police by roughly 2-to-1 (≈1.2M guards vs. <700k officers).

A society that has “solved crime” doesn’t need a parallel private-policing workforce.

Both avoidance and deterrence work (they push the rate down), but they represent a society buying down its crime rate through enormous private expenditure and behavioral constraint, not one that has eliminated the underlying threat.

Policy failure: repeat violent offenders

A society with every favorable condition imaginable: aging population, mass incarceration infrastructure, ubiquitous surveillance, record screen absorption — that still produces the following has a specific, identifiable problem:

• Memphis: 7× the national violent crime rate

• Detroit: 5.3× the national rate

• Baltimore: 4.3× the national rate

• Oakland: 0.1% of the population committed the majority of homicides

A major problem is the woke impulse to refuse incapacitating the ~5% of offenders who commit ~50% of violent crime; this would push violent crime down far more.

BJS data is a big problem. 70% of released prisoners are rearrested within 5 years.

When the NYPD shifted to precision gang takedowns, shootings fell by a third.

The 2025 precision zones produced a 40% decline in shootings.

Criminologist Thomas Hogan estimated ~75 additional homicides per year under Philadelphia’s Krasner; the cost of not incapacitating the power criminals.

The takeaway: “crime is down” needs nuance

America is less violent than the 1990s. The homicide decline is noteworthy — roughly halved from peak — driven by identifiable forces: the end of crack, incarceration, a massive time-use shift that keeps potential offenders indoors, demographic aging, policing innovation, and plausibly abortion legalization. Today’s murder rate may well be the lowest in modern American history, and possibly longer.

But “historic lows across the board” is where the narrative breaks down; we shouldn’t fully extrapolate the homicide story to all crime.

• For nonfatal violence, the government’s own survey contradicts the claim.

• For property crime, the measurement systems contradict each other, enforcement has collapsed, and the fastest-growing categories of financial victimization aren’t counted at all.

These are different kinds of evidence: incident rates and dollar losses are not directly commensurable, but they converge on the same conclusion: the index crimes that generate “historic lows” headlines now represent a small and shrinking share of total criminal harm.

In dollar terms, crime has not declined at all; it has grown by an order of magnitude.

FBI property crime losses were $15.1B in 1995 (~$31B inflation-adjusted); today’s total financial crime victimization: cyber, fraud, shrinkage, insurance fraud, government program losses — likely exceeds $400 billion annually, and Anderson (2021) puts the full aggregate cost at $4.7–5.8 trillion including enforcement and intangible costs.

The measurement distortions concentrate precisely where crime is highest: in the urban neighborhoods where police pulled back most aggressively, where reporting rates are lowest, where residents have the least capacity for private security or relocation.

The national average can decline to “record lows” while per-exposure risk in the worst areas stays flat or worsens, because the already-safe suburban majority drives the average down while victimization among the worst-off is systematically undercounted.

The thermometer is real. It is also broken in ways that systematically read low, it disproportionately measures the temperature in air-conditioned rooms, and the patient has learned to put ice on it before anyone takes a reading.

It’s a fight over whether: (1) a private frontier lab can impose categorical limits on lawful government use, and (2) the government can force compliance by threatening to blacklist that lab from the defense supply chain.

This could impact up to $200M in contracts and the leverage is significant (Claude already sits inside national security workflows).

The Pentagon’s written AI strategy signals this isn’t a one-off negotiation but the opening move in a broader push to strip vendor-imposed restrictions in military AI use.

Anthropic’s 2 Red Lines

Claude must not be used for:

1. Mass domestic surveillance of Americans, i.e., bulk, population-scale analytics run against domestic datasets.

2. Fully autonomous weapons, i.e., systems that select and engage targets without meaningful human oversight.

The Pentagon’s position, stated in its recent: “Artificial Intelligence Strategy for the Department of War," is that all commercial AI used by the military must be:

“Free from usage policy constraints that may limit lawful military applications”

… and that procurement contracts must include standard “any lawful use” language within 180 days.

In other words: if it’s legal, the vendor doesn’t get a veto.

Defense Secretary Pete Hegseth is “close” to cutting business ties with Anthropic entirely and designating the company a “supply chain risk” — a designation usually reserved for foreign adversaries.

The headline number of a $200M contract is a complete distraction; this is a small % of Anthropic’s current biz. However, designation of “supply-chain risk" would be a big deal; under this designation all contractors must certify zero Anthropic / Claude use in defense workflows… and the blast radius can extend far beyond the DoW.

And even if Anthropic later compromises, the reputational scar persists.

A senior Pentagon official told Axios:

“It will be an enormous pain in the ass to disentangle, and we are going to make sure they pay a price for forcing our hand like this.”

That’s a potential existential threat to Anthropic’s government business.

The tone from the Pentagon’s chief spokesman, Sean Parnell, is equally blunt:

“Our nation requires that our partners be willing to help our warfighters win in any fight.”

Reports suggest that senior defense officials have been frustrated with Anthropic for some time and “embraced the opportunity to pick a public fight.”

Meanwhile, Katie Miller (Stephen Miller’s wife) has publicly accused Anthropic of liberal bias1 and criticized the company’s commitment to “democratic values,” giving the dispute an explicitly ideological charge that the other frontier labs have avoided.

Palmer Luckey chimed in:

Unlike OpenAI, Google, and xAI — all of which have already agreed to remove their safeguards for use on the military’s unclassified systems — Anthropic is the only major lab still pushing back.

Why is Anthropic’s Claude the “Go-To" AI for the U.S. Military in 2026?

Currently the Pentagon can’t just easily “walk away" from Claude.

• Claude is not the only AI used by the U.S. military. The DoW deploys multiple models on unclassified networks through genai.mil, including ChatGPT, Grok, and Gemini, reportedly available to over 3 million Defense/War Department employees.

• But only Anthropic has a major presence in classified environments, as a result of deep integration with Palantir.

In 2024, Anthropic and Palantir announced a partnership to operationalize Claude in Palantir’s AIP on AWS for U.S. government intelligence and defense workloads.

Palantir was already heavily embedded in government defense workflows, making the Anthropic/Claude x Department of Defense/War (DoD/DoW) a natural fit.

Additionally, in April 2025, Anthropic joined Palantir’s FedStart program, expanding Claude’s availability to government customers via Google Cloud.

So this is how we ended up with Claude being the “go-to” AI for military operations in 2026. Another model could be used, but others are not as integrated with Palantir and there’s less familiarity.

Furthermore, it takes time to bring a new frontier model (from OpenAI, xAI, Google), etc. into the classified fold.

Former Biden administration AI advisor Varoon Mathur told Fast Company:

“If you’ve never operated in a classified environment before, you essentially need a vehicle. Palantir is a defense contractor with deep operational integration. Anthropic is an AI model provider trying to access that ecosystem.”

The Pentagon is still working to pull multiple labs into classified networks, but the pipeline is not broadly mature.

Senior administration officials claim that competing models “are just behind” when it comes to specialized government applications. (The models themselves aren’t really “behind” in performance… just the integration with Palantir and gov ops takes time.)

None of that means the government can’t route around Anthropic.

But routing around is slow and expensive in the near term (months to >1 year for classified parity), and usually comes with tradeoffs: weaker models, worse integration, more bureaucracy, or zero vendor constraints because it’s open-source or fully government-run.

Anthropic can’t stop the capability forever; it can only make the Pentagon pay a timeline + friction tax. That’s why the Pentagon is reaching for coercive leverage instead of just switching vendors tomorrow… they want to speed things up.

Palantir has stayed radio silent as tensions escalate. If supply chain risk designation goes through, defense contractors would have to certify they aren't using Claude in government workflows which could damage Anthropic’s business and reputation.

We know Claude has already been leveraged for serious ops: the WSJ reported Claude was used in the Maduro raid. And an Anthropic employee allegedly reached out to Palantir to ask about Claude’s use in the operation, though Anthropic denied it spoke with Palantir beyond technical discussions.

How Does Claude Spy on Americans? It Doesn’t.

Claude isn’t a camera, a wiretap, or a collection system. It doesn’t go out on its own and spy on anyone. It’s a frontier AI model that processes data fed to it and generates outputs based on its algorithm. It has no idea whether it’s reading a real immigration database or a fantasy novel. If you paste a million rows of data into Claude and say “find patterns,” it just does it.

The dispute between Anthropic and the Pentagon is contract language.

1. Anthropic wants contractual carve-outs that prohibit two categories of use.

2. The Pentagon wants blanket “any lawful use” terms so they don’t have to lawyer every application against Anthropic’s policy.

This is a fight over terms of service, not over the model blocking prompts.

The Usage Policy for Anthropic’s Claude

Anthropic’s usage policy lists its surveillance-related restrictions under:

“Do Not Use for Criminal Justice, Censorship, Surveillance, or Prohibited Law Enforcement Purposes.”

The relevant prohibitions include:

• “Target or track a person’s physical location, emotional state, or communication without their consent, including using our products for facial recognition, battlefield management applications or predictive policing”

• “Utilize models as part of any law enforcement application that violates or impairs the liberty, civil liberties, or human rights of natural persons”

Even for government customers who receive tailored exceptions, Anthropic’s documentation says restrictions on “domestic surveillance” and “the design or use of weapons” remain in place. The problem is that these examples read like they’re regulating collection systems (trackers, recognizers, policing systems).

Claude isn’t a sensor and doesn’t collect anything. But once integrated into an operational platform with databases and tool access, Claude can still function as the analytic layer:

• Analyze location fields already collected elsewhere

• Correlate identities and relationships

• Summarize and generate watchlists from structured/unstructured data

• Interpret outputs from other systems (including face-recognition outputs)

• Automate triage and reporting at dataset scale

So the mismatch is this: the policy language is written as if it’s regulating the collector, while the real dispute is about the analytic layer running at scale over already-collected data.

That’s also why the practical use case in dispute is simple: the government feeds Claude a volume of already-collected, legally obtained data and asks it to analyze text at scale, find patterns, and generate summaries.

Claude Is Used via Palantir’s AIP

Claude in a classified defense setting is not someone typing questions into a chat window. It’s running inside Palantir’s AIP — an operational AI platform that is far more capable than most people can comprehend.

Palantir’s AIP includes:

1. AIP Agent Studio, which lets users build autonomous agents that “handle complex workflows autonomously” — agents that can read from and write to data systems, not just answer questions but take actions.

2. AIP Automate, where you can define conditions “checked continuously or on a schedule,” with effects that “execute automatically”; and for streaming, “effects will execute within seconds of new data entering the ontology.”

3. Pipeline Builder LLM, which enables “executing large language models on your data at scale” — running Claude across entire datasets in batch, not one-off prompts.

4. Ontology system where Palantir says the Ontology “enables both humans and AI agents to collaborate” across operational workflows; it also describes a “journey from augmentation to automation."

In other words: Claude inside AIP can power agents and automations that run without a human initiating each query or step, and can be configured to write back actions or generate alerts — with the degree of human review/approval determined by the surrounding workflow.

How Is Claude Being Used Now?

Anthropic says Claude:

“Is used for a wide variety of intelligence-related use cases across the government, including the DoW.”

Given the platform it sits on (Palantir), current applications likely include: summarizing and synthesizing large volumes of intelligence reports; drafting briefings, assessments, and operational planning documents; translating and analyzing foreign-language source material; extracting structured data (names, locations, timelines, relationships) from unstructured text at dataset scale; running analytical workflows against intelligence feeds; and powering agent-driven queries across classified knowledge bases.

This is happening against foreign intelligence and operational planning data, which is the kind of use Anthropic’s government exceptions were designed to allow.

An Anthropic spokesperson told Axios that the company’s conversations with DoW that they:

“Have focused on a specific set of Usage Policy questions — namely, our hard limits around fully autonomous weapons and mass domestic surveillance — none of which relate to current operations.”

So “current use cases" aren’t the dispute.

What the Pentagon Wants Going Forward

The Pentagon’s complaint, per Axios, is that Anthropic’s terms have:

“Gray areas that would make it unworkable to operate on such terms.”

The Pentagon wants “any lawful use" terms so they don’t have to evaluate and question every future application against an ambiguous usage policy.

The Pentagon’s steelman is blunt: The government already runs domestic-facing enforcement and intelligence workflows. AI makes existing workflows cheaper and more scalable. So vendor-imposed carve-outs don’t “protect democracy” in practice — they impose a timeline/friction tax on lawful missions. And in a competitive environment, a vendor that can slow operational adoption is itself a supply-chain risk, even if the vendor is domestic.

AI vendor vs user liability: Anthropic isn’t the actor conducting enforcement or surveillance; the government is. A general-purpose model is closer to infrastructure than an operator: like a knife that can be used to chop vegetables or commit a crime. Accountability for lawful-but-controversial government action runs through elections, statutes, courts, audits, and chain-of-command responsibility — not a private vendor’s terms of service. If the state is determined to do a thing, vendor restrictions only change the timeline and vendor identity, not the endpoint.

Given Palantir’s AIP capabilities, domestic use cases might look like:

1. Agents running against domestic datasets (immigration, financial, social, tips/law-enforcement reporting) to triage and cross-reference at scale

2. Streaming/event-driven automations that generate alerts when criteria are met

3. Natural-language query interfaces over domestic databases

4. Batch processing across large text corpora to extract entities/relationships and generate reports

The Pentagon can already gather everything from social media posts to concealed carry permits, and existing mass surveillance law doesn’t contemplate AI at all; they have the big datasets.

A frontier AI model like Claude adds an analytical layer that makes population-scale usage of existing data cheap, fast, and operationally practical.

If that model is running as an autonomous agent inside a platform designed for continuous streaming operations, it begins to look like autonomous surveillance infrastructure — far beyond “human analyst augmentation” — which is where Anthropic has concerns.

Can Anthropic Enforce Existing Terms?

Anthropic can set its terms, but once its AI model (Claude) is deployed, it has zero visibility into what data the government is processing.

“Zero data retention” is the status quo… any data the government processes via Claude by way of Palantir is classified and unauditable.

A former OpenAI employee and AI safety expert Steven Adler told Fast Company:

“Anthropic and OpenAI offer Zero Data Retention usage, where they don’t store the asks made of their AI. Naturally this makes it harder to enforce possible violations of their terms.”

Then why is the Pentagon fighting so hard for contract language if they could probably just do it anyway? Because government procurement is bureaucratic.

Contracting officers need to certify compliance with vendor terms of service. If the contract says “no domestic surveillance” and an agency uses Claude for immigration enforcement analytics, they’re technically in breach — which creates legal exposure, audit risk, and career risk for the officials who signed off.

The Pentagon doesn’t want to operate in a gray zone where every use case might “violate their vendor agreement." They want the contract to explicitly say “any lawful use” so every downstream operator has full legal cover.

The fight is about legal cover, procurement norms, and precedent.

Dario Amodei’s “Democracy"

Anthropic is perfectly within its rights to set whatever contract terms it wants. The question people have is whether justification for Anthropic’s specific red lines holds up under scrutiny.

In his essay: The Adolescence of Technology, Dario Amodei (Anthropic’s CEO) lays out the philosophical basis for his position.

He argues that (1) democracies need AI-enabled military/geopolitical tools to compete with autocracies, and (2) AI lowers the manpower required to surveil or repress citizens, making it more challenging to reform or overthrow the gov.

Thus, we should arm democracies with AI, but we should do so carefully and within limits: they are the immune system we need to fight autocracies, but like the immune system, there is some risk of them turning on us and becoming a threat themselves.

His “within limits" qualifier implies: arm democracies, but don’t hand them tools that can cheaply scale domestic control.

However, some wonder whether in practice, “arm democracies" comes with an unspoken asterisk:

Arm only democracies that I and/or the Anthropic team personally trust.

The reality: Trump was democratically elected.

His administration’s policy priorities (Greenland, Canada, the Maduro operation, ICE enforcement) aren’t rogue actions by an unelected junta.

Many question whether Anthropic and Dario might only trust a “democracy" with policies they personally find acceptable.

If voters elect someone whose agenda is disliked by Dario et al. (aggressive immigration enforcement, territorial ambitions, regime-change operations), suddenly this specific democracy shouldn’t be using AI.

His qualifier (“within limits") functions as a political kill-switch that can be activated whenever the “wrong" democracy wins.

Many conservatives perceive Anthropic as unaligned with democracy when conservatives are in power; slow-rolling and/or blocking capabilities that they may not for a more leftist regime.

By the Way… It Won’t Actually Stop Anything!

There’s a practical absurdity that makes the principled stand even harder to defend: the U.S. government is going to do this stuff anyway; the Pentagon has alternatives lined up.

OpenAI, Google, and xAI have all agreed to remove their safeguards for military use. Plus open-source models can be deployed internally with zero usage restrictions (the gov can build its own stack).

What does Anthropic’s stand actually accomplish?

1. It doesn’t prevent mass domestic surveillance.

2. It doesn’t stop autonomous weapons development.

3. It doesn’t change the legal framework or the political incentives.

4. It only ensures that Anthropic currently isn’t the one enabling it, which is a reputational and moral comfort move, not a consequential one.

5. It also drops a transient roadblock for the Trump administration to carry out its agenda as efficiently as possible.

Dario gets to sleep at night and the government gets the same capability ~3-6 months later from someone who won’t ask questions. The net effect on civil liberties is nothing in the long-term because the replacements will have no constraints. And Anthropic may lose business.

The Partisan Hypothetical

A Progressive / Liberal administration wants to use Claude for large-scale domestic monitoring under a different label (e.g. countering hate speech, election integrity, domestic extremism, White supremacist networks, misinformation / disinformation, preservation of democracy, etc.)

Would Anthropic fight just as hard? Or would they have already rolled over?

That program still means bulk analytics across datasets containing tens of millions of ordinary citizens’ speech, contacts, locations, and transactions.

If the answer is “probably not,” or even “we don’t know,” then the carve-outs aren’t principled; they’re selectively enforced based on which faction is in power.

Anthropic’s Hedge

There’s a reading of this standoff that’s more strategically coherent than the tribal takes: Anthropic is hedging. Trying to stay viable across administrations by refusing to be stapled to the next domestic scandal.

One concrete signal they’re managing politics, not just abstract safety: Anthropic is spending $20 million backing state-level AI regulation efforts, while Trump is pushing a “minimally burdensome” national framework. If you’re thinking of reputational tail-risk, this is exactly why Anthropic might want explicit contractual prohibitions (e.g. domestic bulk analytics and fully autonomous targeting) so they aren’t tied to any future scandal.

The logic: If Anthropic signs a blank-check “any lawful use” contract and the Trump administration uses Claude for aggressive domestic enforcement, mass-scale immigration analytics, or something that becomes a political scandal, Anthropic is the company that enabled it. When the political tide shifts (guaranteed with the current demographic trajectory in the U.S.), they’re potentially a villain in the next administration’s congressional hearings. Guardrails ensure Anthropic keeps their hands clean enough so that no matter the regime in power, they maintain credibility.

There’s also an internal political dimension. Anthropic’s workforce skews heavily left-wing, as does the broader Bay Area AI/EA talent pool they recruit from. If leadership is seen as rolling over for Trump on domestic surveillance and autonomous weapons (the two most politically charged categories imaginable), internal backlash could be severe. Engineer attrition, public letters, protests, leaked Slack messages — the whole playbook that’s hit every major tech company in the last decade.

Dario may or may not personally care about the policy substance, but he definitely cares about keeping his team intact. The guardrails might be less about civil liberties and more about making sure his best researchers don’t quit and post about it on X BlueSky.

This explains the multi-hedged posture:

1. Natsec-credible to stay inside defense/IC workflows and not look like a politically hostile AI vendor.

2. Guard-railed to keep the workforce and brand intact on the domestic surveillance/autonomy fault lines.

3. Regulation-friendly (including the $20M push) to preserve legitimacy with the Democratic/regulatory ecosystem if the political tide flips.

Anthropic wants to be the company that can work with any administration, and to do that, it can’t be seen as having gone all-in with Trump 2.0’s domestic agenda.

Maybe the Limits Are Real

The fairest version of Anthropic’s defense is that the red lines are capability-based, not administration-based. No mass domestic surveillance for any president and no autonomous weapons for any administration.

Charitable reading: The Trump admin’s specific priorities (ICE enforcement, Maduro-style ops, expansive domestic security) may collide with Anthropic’s red-line categories more than a different administration’s. The conflict looks partisan because Trump’s agenda disproportionately triggers the “bulk domestic analytics” concern. From the outside: (A) “partisan veto dressed as preserving democracy" vs. (B) “principled limits that bind this administration harder" — is almost impossible to distinguish.

My read: Some real principle, some partisan friction, some brand management — with proportions unknowable from the outside.

Most Likely Motives

Several incentives reinforce each other, which is why the standoff is so hard to unwind.

My estimate: Everything probably true to some extent.

• Guardrails (Extremely likely) — The two categorical red lines are the core; they predate this administration and would apply under others.

• Employee cohesion (Very likely) — Capitulating on domestic bulk analytics/autonomy risks internal revolt and attrition.

• Hedging / optionality (Very likely) — Minimize future political/regulatory blowback if today’s lawful use becomes tomorrow’s scandal.

• Trump-era friction (Very likely) — This administration pushes closer to edge cases and is more willing to use coercive procurement leverage.

• DC optics / natsec signaling (Possible) — Some hawkish China talk may be “Team USA" signaling to stay acceptable in Washington. Even if true, it doesn’t explain the core standoff: refusing to sign away the two categorical prohibitions under a blacklist threat.

The China Cost

The strategic argument against Anthropic’s position is simple: if frontier AI materially improves analyst throughput, targeting speed, and operational planning quality, throttling your own side (even for domestic operations) while China doesn’t throttle theirs is a strategic gift to your international competitor.

The DoW AI strategy memo explicitly states the risks of not moving fast on AI outweigh “imperfect alignment"; the “speed wins” doctrine treats adoption rate as a decisive variable, not a nice-to-have.

Slow-rolling by Anthropic compromises the gov’s ability to execute its broader national-power agenda (security operations, enforcement, logistics, deterrence, etc.).

And if you buy the administration’s view that its initiatives improve U.S. competitiveness in a China-centered rivalry (even if indirectly), then Anthropic’s “slow-rolling” is an operational drag on the rate of national improvement and arguably anti-American.

What’s Likely to Happen?

Axios already frames Anthropic as “prepared to loosen its current terms of use” while insisting on the two red lines, which is the posture that produces this kind of split.

Most likely: DoW gets “any lawful use" in the master contract (headline win). Anthropic probably preserves facets of its 2 red lines via “implementation details." Think: “meaningful human judgment” required for lethal-action workflows; governance gates for certain domestic/U.S.-persons analytics requiring explicit legal sign-off before deployment. DoW frames these as its own operational controls, not vendor vetoes. Anthropic gets to claim it didn’t sign a blank check for autonomous lethal action or mass domestic surveillance.

An alternative: “Any lawful use” with ultra-specifically defined use terms so contracting officers don’t fear accidental breach. Instead of the vague “mass domestic surveillance,” define something like “bulk, continuous monitoring/triage of U.S. persons at population scale.” In other words: Anthropic needs to be far more specific.

Even in a Compromise, DoW Diversifies Away

This is the part Anthropic can’t negotiate around.

Even if a deal gets done, DoW will aggressively second-source and reduce dependence on any single vendor.

The same memo that demands “any lawful use” also pushes modular open source architectures (MOSA) designed for component replacement at commercial velocity — exactly the antidote to future vendor hold-ups.

Diversify away looks like:

1. Contracting: Enforce MOSA with open interfaces to enable component replacement at commercial velocity, per the memo’s explicit directive.

2. Platform: Require model-abstraction and routing inside platforms like AIP. Palantir’s BYOM (Bring-Your-Own-Model) documentation provides first-class support so the platform layer is designed for multi-model routing when alternatives are cleared.

3. Supply: Pull other frontier labs into classified networks. The Pentagon is actively pushing this.

4. Operations: If possible, stop expanding Claude-only workflows. Shift new workloads to alternatives first; keep legacy Claude deployments until parity is achieved.

The trajectory is keep Claude in the short-term because it’s already there and aggressively increase classified alternatives over 6–18 months so no future vendor can credibly threaten operational friction; move toward multi-vendor platforms with swapability.

As for Palantir: There’s no evidence they’re planning to oust Claude. Palantir is “caught in the middle” and staying quiet while the Pentagon threatens escalation. The observable posture is wait-and-see, not rip-and-replace. If DoW triggers a “no Claude” regime, Palantir doesn’t have a choice; they either remove Claude from deployments or lose gov contracts.

The Decision Window: 180 Days

Negotiations have been ongoing for months and Hegseth is close to escalation. This doesn’t suggest a “slow-roll” to Day 180.

Two clocks are running simultaneously:

1. Procurement clock: The DoW memo orders objectivity benchmarks within 90 days (April 2026) and standardized “any lawful use” contract language within 180 days (July 2026). The bureaucracy needs a resolved posture well before the second deadline.

2. Leverage clock: Hegseth is close to cutting ties, but Pentagon officials also admit disentangling is painful, and replacement must be “orderly.” That combination usually produces near-term pressure and faster negotiations.

Timing estimate:

• Late March to April 2026 (highest probability window): Face-saving compromise gets papered — “any lawful use” language plus implementation constraints. If Anthropic hasn’t moved by this point, DoW likely starts the off-ramp process.

• By May 2026: Either a deal exists or DoW has initiated an orderly wind-down — freeze expansion, shift new workloads to alternatives, accelerate classified onboarding for other labs.

• By July 2026: Either Anthropic has papered something that meets “lawful use” in form, or DoW has already initiated the transition away.

Four Scenarios

1. Scenario A: Negotiated Compromise (~55%). DoW gets “any lawful use” in the master contract; Anthropic preserves practical constraints through implementation details, classified annexes, and narrowly defined workflow requirements. Both sides save face. Anthropic avoids the supply-chain-risk designation; the Pentagon avoids the operational disruption of ripping out an embedded tool. Timing: Term sheet by late March–April, finalized before the July 8 procurement deadline.

2. Scenario B: Pentagon Cuts Anthropic Without Blacklisting (~30%). The contract ends or shrinks sharply, the Pentagon shifts classified efforts to alternatives over time, and uses the break as leverage against the rest of the market, without triggering the massive blowback of a full supply-chain designation. Timing: Decision to wind down by April–May, with an orderly replacement transition through summer.

3. Scenario C: Supply Chain Risk Designation (~8%). The Pentagon takes the nuclear option. Every defense contractor must certify it isn’t using Claude. Given that 8 of the 10 biggest U.S. companies reportedly use Claude, that’s a sweeping disruption — devastating for Anthropic’s government business but also self-inflicted chaos for the Pentagon, given Claude’s existing classified presence. More threat than plan, but it’s being floated publicly, which means it’s not zero. Timing: Most plausible as a late-spring escalation after DoW has a workable transition plan, because disentangling is itself disruptive.

4. Scenario D: External Intervention (~7%). Congressional oversight, legal challenges, or political dynamics slow the Pentagon’s enforcement timeline or force a more formal policy framework. The DoW rename itself technically requires congressional action, and procurement/legalities can always generate friction. Timing: Could reshape the procurement posture ahead of the July clause rollout.

Final Thoughts: Anthropic vs. the Pentagon

Should AI vendors have categorical veto power over lawful government use of their products?

1. The Pentagon says no — their position is that “lawful” is the boundary, and commercial vendors don’t get to add further restrictions that function as operational constraints on elected government.

2. Anthropic says lawfulness is necessary but not sufficient. Their position is that certain categories of use (bulk domestic analytics, autonomous lethal action) carry enough downside risk (reputational, civil-liberties, mission-creep) that hard limits are warranted even when the specific application might be legal.

Both positions have reasonable logic. (1) The Pentagon’s frame is sovereignty and democratic authority, (2) Anthropic’s frame is institutional safeguard and tail-risk prevention.

As former Palantir employee Alex Bores, now running for Congress, told Fast Company:

“To state basically that it’s our way or the highway, and if you try to put any restrictions, we will not just not sign a contract, but go after your business, is a massive red flag for any company to even think about wanting to engage in government contracting."

In sum, the Pentagon’s hardball with Anthropic is explicitly intended to set the tone for the future (including any negotiations with xAI, OpenAI, Google, etc.). If Anthropic yields, the “any lawful use” standard becomes the industry-wide default, and the vendor-veto problem disappears.

They write books about the beauty of wrinkles, give talks about why senescence is sacred, and get invited to panels where they solemnly explain why your biological decline is, actually, good for you. Why aging gives life “meaning” and why you should be grateful your body is falling apart on schedule.

They can afford to say all of this because age reversal doesn’t work yet.

The technology is speculative, partial, and unproven at scale. They’re performing “sacrifice” at zero cost; playing the role of the wise elder who’s made peace with mortality, without ever being tested on it.

That’s what makes them cheap martyrs.

And here’s the thing: the moment age reversal actually works (e.g. ~30 years off the biological clock, safe, affordable, maintainable) most if not all of them are going to use it.

Some will do it quietly through concierge medicine and “wellness protocols”, others will do it loudly, rewriting their own history in real time, explaining with perfect confidence why this is actually very different from what they were criticizing.

Either way, the principled stand evaporates the moment the cost of holding it becomes real. They’re likely full of shit.

Most aren’t bad people, but almost nobody is willing to age and die to win a philosophical argument, and there’s no reason to believe the “anti-aging” critics are an exception.

Anti-Aging Critics in Question

These are on-the-record critics of radical anti-aging whose arguments will get stress-tested the moment reversal becomes cheap and effective. Most are clout maxxing, coping, or just deeply ignorant about the most likely implications of reversing biological aging.

Ezekiel Emanuel is the big one. In 2014, the oncologist, “bioethicist,” and Obamacare architect published “Why I Hope to Die at 75" in The Atlantic, arguably the single most famous anti-longevity argument in mainstream American media. He vowed to refuse all medical interventions after 75, including antibiotics and vaccines. His argument: older Americans live too long in a diminished state, and past 75 your “consumption” exceeds your “contribution.” He’s now around 68, roughly 7 years from his own self-imposed cutoff, and already carving out exceptions. In a 2025 talk at Yale Law School, he explicitly acknowledged the lifesaving value of COVID vaccines in older populations and treated them as an exception to his earlier position. In a 2023 revisit of his stance, he admitted his partner wants him to reconsider preventive measures and conceded that “a lot will depend on whether I really am a rare outlier at 75.”

The walk-back is underway before the deadline even arrives, and the technology doesn’t even work yet. He hasn’t stayed true to his own position under mild pressure: the passage of time and the opinions of his partner. Now imagine someone offers him a treatment that makes 75 feel like 40 again biologically. Emanuel is the single best illustration of the cheap martyr dynamic: a man who made a very public pledge about accepting decline, who is already carving out contingencies as the deadline gets closer, and who would almost certainly abandon the whole framework if age reversal removed the very decline he claims to accept.

Francis Fukuyama wrote an entire essay in May 2025 titled “Against Life Extension", subtitled, with zero self-awareness, “An argument in favor of death, disease, and mortality." He frames defeating aging as individually desirable [in his words]: “by everyone", but socially catastrophic. But his piece reveals exactly how confused the opposition is. He argues that life extension will: (1) produce a world of increasing debilities (people living longer but with Alzheimer’s, Parkinson’s, and cognitive decline) as if (2) the entire point of age reversal isn’t to eliminate those conditions. He’s not arguing against rejuvenation — he’s committing the Tithonus error.

Then he gets personal:

“I honestly do not look forward to the prospect of living another 20 years, and having people say behind my back (as they likely do already) ‘he’s still spouting the same nonsense he was in the late 20th century.’ There is a time to move on.”

And he even tells a libertarian debater:

“I am not looking forward to a world in which you will be spouting your same dumb libertarian ideas 100 years from now.”

This is supposed to sound wise and self-deprecating, but it’s actually a stunning self-own. Fukuyama is assuming that he and everyone else would remain “cognitively frozen in place," endlessly recycling the same ideas they formed decades ago like some robot. But that’s literally a description of what aging does: cognitive flexibility, openness to experience, and neuroplasticity all decline with age. The rigidity he’s afraid of is a symptom of the aging he doesn’t want to cure. He’s also bizarrely assuming he’ll be “spewing the same nonsense" (as if he talks about the same things every day in his current life and doesn’t ever update his views). He is implying he’ll be frozen in time, completely static. He also thinks arguing about “old ideas" is somehow bad… as if historians don’t already argue about ideas from centuries ago. “Someone might find your ideas stale” is not a serious argument for accepting death.

Leon Kass has spent decades arguing that finitude shapes meaning, love, and striving, that mortality is what gives life its urgency and beauty. This is a gorgeous argument in a seminar room. It’s a much harder argument to maintain at 86 when someone offers you a body that functions like it did at 35, and the alternative is watching yourself lose the capacity to think, move, and remember.

Daniel Callahan insists nature “knew what it was doing" with biological aging. Nature also knew what it was doing with smallpox, infant mortality, and dental abscesses that killed people at 30. We overruled nature on those. The question is whether Callahan would volunteer to be the last man standing on the “nature knows best” hill when the alternative to aging is not aging.

Bill McKibben, the environmentalist, wrote in his 2003 book Enough that a radically extended life would produce a future that “stretches before you, endlessly flat." This is the boredom objection dressed up in literary prose, and it assumes, without evidence, that a biologically young person with centuries of accumulated experience would somehow run out of things to care about. It also assumes McKibben himself would choose to age and die rather than find out.

Gilbert Meilaender raises doubts about whether turning back the clock is even desirable, framing it through patience, gratitude, and accepting limits. He's in his late 70s. Ask him again if someone offers to give him back the body and mind he had at 25. Preaching acceptance of decline you can't stop is easy. Maintaining it when the decline becomes optional is a different proposition entirely.

Bernard Williams made the classic philosophical case that immortality would be intolerably tedious in his “Makropulos Case” argument. But rejuvenation isn’t immortality (you can still die from a thousand causes), and “I might get bored” is not a justification for imposing biological decay on others who don’t share your ennui.

John Hardwig has argued that there can be circumstances where people have a “duty to die" based on resource and family burden framing. Even if you grant the hard cases, it doesn’t justify banning the prevention of degeneration. Nobody has a duty to John Hardwig’s philosophical preference to decline age reversal when the decline is optional.

Various academic bioethicists have published formal arguments that anti-aging research is morally unacceptable, invoking: justice, community, meaning, and the natural order. I guess they think it’s not possible to have justice, community, or meaning unless people die? The “natural order” argument is particularly rich coming from people who are able to maintain careers as “bioethicists"… is that the “natural order… not really… it’s a fucking clown job. All these “natural order” morons use: cars, cell phones, computers, airplanes, and modern medicine. More on that in a moment. (FYI: You don’t hate most bioethicists nearly enough; most are IYIs — intellectual yet idiot.)

The key point: None of these people oppose treating Alzheimer’s, joint replacements, or cataract surgery. They oppose the framing, the idea of treating aging itself as a disease to be cured. But that’s a distinction that collapses the moment the technology works, because “treatment for age-related decline” and “age reversal” become the same thing once the effect size is 30 years. If that happens in their lifetime(s), every one of them will miraculously discover they were never really against anti-aging. Some may have been against the hype and/or hubris and/or speculative phase. Then suddenly it just becomes “medicine.”

Tithonus Error: They’re Imagining the Wrong Thing

A huge portion of the anti-aging opposition rests on a picture of age reversal that no serious researcher is actually pursuing. And much of the public opposition rides the same wave, not out of deep reflection, but because when you don’t have an alternative, you rationalize the status quo. Psychologists call it system justification; normal people call it sour grapes. If you can’t have it, it must not be worth having.

The critics (and the public they’re playing to) seem to imagine that “curing aging” means trapping someone at 85 with bad knees, fading cognition, and chronic pain, but now they just... don’t die. A miserable, extended twilight.

Aubrey de Grey calls this the Tithonus error, after the Greek myth of a man granted eternal life but not eternal youth, and it poisons the entire discourse because it lets people oppose a caricature instead of engaging with what the research actually targets.

What serious anti-aging research pursues is biological rejuvenation: restoring tissue, organs, and critically the brain to a younger functional state. Not freezing you at 80. Resetting you to 30. That means the neuroplasticity of a young brain combined with decades or centuries of accumulated knowledge and pattern recognition.

There’s also a critical category error running through almost all of this discourse: Biological age reversal is not immortality. Critics constantly conflate the two, sliding from “reversing aging” → “living forever” → “immortality” as if they’re the same thing. They aren’t. Even a person with zero biological aging can still die. Accidents, violence, infections, cancer, pandemics, disasters, warfare, and bad luck don’t go away because you are biologically young. Nobody is proposing invincibility. The proposal is that your body and brain don’t have to rot on a timer while you’re still alive. Framing that as “immortality” and then arguing against immortality is a straw man, and the critics lean on it because the real proposition (would you like your body to work properly for longer?) is much harder to argue against.

And then there’s the dumbest objection: “Who the hell wants you around that long?” As if someone else’s preference for your death is a legitimate basis for restricting a medical therapy. Since when does “I find your continued existence inconvenient” constitute an argument for banning medicine? Nobody frames it this way for cancer treatment or heart surgery. The implication, that your right to remain alive and healthy is contingent on other people’s enthusiasm for your presence, is grotesque on its face. And it only survives because nobody says it out loud in those terms.

Most of the anti-aging opposition, from the bioethicists and the public alike, is certified cope. It’s the psychological machinery humans use to make peace with things they can’t change. When you’re going to age and die no matter what, your brain does you the favor of constructing a narrative where that’s fine, maybe even beautiful. Mortality gives life meaning. Wrinkles tell a story. There’s dignity in decline. This is how humans survive psychologically: you accept what you can’t escape, and then you elevate the acceptance into a virtue.

But acceptance of fate is not the same as preference for fate. If you removed the inevitability, if someone credibly offered to make aging optional, the “acceptance” would evaporate almost instantly, because it was never a considered philosophical position. It was a coping mechanism built on the absence of an alternative.

For most of human history, people “accepted” childhood death, infectious disease, famine, and enslavement. Not because these things were good, but because there was no obvious way out. The moment alternatives appeared, the acceptance vanished overnight.

We correctly identify their acceptance as a product of helplessness, not wisdom.

That’s what’s happening with aging right now. The entire cultural apparatus of “death is natural and beautiful” is a cope stack built on the absence of an alternative. The critics are just the people who’ve professionalized the cope and turned it into books, lectures, and panel appearances. Pull the foundation out, and the whole structure comes down.

The “Natural Order” Hypocrisy

If you’re reading this essay right now, you’ve already rejected the natural order. You’re using a computer or phone, connected to the internet, wearing synthetic fabrics, and consuming food that was grown, processed, and transported using industrial agriculture. None of that is natural. All of it extends your comfort, health, and lifespan beyond what “nature intended.”

Every single one of the critics likely uses: antibiotics, MRIs, anesthesia (during surgery), corrective lenses or glasses, airplanes, vaccines, C-sections, insulin, joint replacements, clean water, etc.

Their entire existence is a monument to overriding the natural order, and they’re fine with that, because the benefits are personal and obvious.

Where exactly is the principled line? Should we bring back smallpox and polio and get that childbirth mortality back up to what’s “natural"? Nobody serious argues we should go back… but it would be “natural." Don’t you have a deep yearning and nostalgia for polio?

These critics just draw an arbitrary line at aging and declare everything on one side “medicine" and everything on the other “hubris." The line for these clowns is wherever the technology currently stops working.

“Natural" is just a label for whatever we haven’t fixed. Once humans fix it, everyone quietly reclassifies the intervention as normal medicine. Aging is next.

And if you sincerely believe the natural order should be respected? Move into the wilderness, refuse all medical care, and die of whatever gets you first.

Nobody advocating “natural limits” positions are doing any of that… because the position wasn’t about nature; these fools just like to act like they’re some highly-enlightened philosopher… nauseating and pathetic.

The Biggest Argument They’re Missing: Aging Is the Disease

Critics of anti-aging are not just wrong, but their public stance is extremely damaging: reversing aging would be the single greatest force-multiplier for (1) preventing and (2) curing disease — in human history.

Nothing else would come close.

The overwhelming majority of cancer, heart disease, stroke, Alzheimer’s, Parkinson’s, osteoporosis, type 2 diabetes, macular degeneration, arthritis, and organ failure are diseases of aging.

They happen because the body deteriorates over time, and that deterioration creates the conditions for disease.

Aging isn’t just correlated with these conditions. It’s the primary cause. Fix the underlying biological aging, and you don’t just extend life. You prevent or dramatically reduce the incidence of nearly every major disease simultaneously.

Investing more money with the aim of reversing human aging would be far more cost-efficient/effective than spending hundreds-of-billions or trillions on “finding new treatments for age-related disease"; none of the treatments target the underlying cause and are bound to be mostly a half-assed patch.

Age reversal treatment is somehow considered taboo yet would yield the biggest gains (by far) in actually preventing/reversing most disease.

This is why the critics’ position is incoherent at its core.

Unless you are actively against preventing and curing cancer, heart disease, and dementia, you shouldn’t be against reversing the biological process that causes them.

The critics have somehow managed to (A) socially oppose the single intervention that would do more for human health than every other medical advance combined, while (B) simultaneously claiming to care about human wellbeing.

The position doesn’t survive 5 minutes of scrutiny once you frame it correctly:

“Are you against fixing the root cause of most human disease? If not, you’re not really against anti-aging research. You just don’t realize what it is."

And this matters far beyond philosophy, because the anti-aging discourse has real policy consequences.

When prominent intellectuals frame longevity research as hubristic or dangerous, governments invest less in it. Funding goes elsewhere. The Overton window stays narrow. People think it’s not even possible or stupid.

The result is that the single most impactful medical intervention in human history gets treated as a fringe pursuit instead of a civilizational priority. Every year that serious age-reversal research is underfunded because the discourse treats it as science fiction or moral transgression is a year where millions of people die from conditions that a more aggressive research agenda might have prevented.

If Life Is Sacred, Why Are You Eager to Lose It?

Many of the critics operate from religious or quasi-religious frameworks where life is treated as sacred, a gift from God, something precious and irreplaceable. Kass draws explicitly from religious tradition. Meilaender is a theologian. The “natural order” language throughout bioethics carries unmistakable theological undertones.

But if life is sacred, the logical conclusion isn’t to accept its destruction on a biological timer. It’s the opposite. If life is truly a gift, you should do everything in your power to preserve and protect it for as long as possible. Accepting aging and death without resistance, when resistance becomes possible, doesn’t honor the sacredness of life. It treats life as disposable. It suggests you’re willing to give up something you claim is precious with remarkably little fight.

The “life is sacred, therefore accept death gracefully” position contains a logical contradiction that most people just don’t notice because the framing is so culturally familiar. Sacredness implies maximal protection. You don’t let sacred things rot on a schedule because “it’s natural.” You protect them.

And here’s something the critics really haven’t considered: if aging is reversed, both life and death become more sacred, not less. Right now, death comes for nearly everyone through the slow, grinding indignity of biological decay. It’s expected. Routine. The body just gives out. That’s not sacred.

In a world where aging is optional and people can maintain biological youth indefinitely: every death carries more weight, meaning, and gravity, because it wasn’t inevitable. Critics have it backward: curing aging elevates the value of both life and death.

The Incentive Structure: Why the Opposition Exists at All

Before asking whether these critics are right, ask what’s in it for them to hold these positions. The answer clarifies a lot.

Several of these figures have gotten: (1) book deals, (2) speaking fees, (3) prestigious panel invitations, and (4) elevated public profiles specifically from opposing life extension.

• Emanuel’s “Why I Hope to Die at 75” is probably the single most-read thing he’s ever written.

• McKibben’s Enough carved out a niche in a crowded environmentalist market.

• Kass built a career as the go-to conservative bioethicist partly on the strength of his mortality-as-virtue framework.

• Fukuyama’s “Against Life Extension” generated enormous engagement precisely because it was contrarian and provocative.

This doesn’t mean they’re all being dishonest. But the incentives are worth noting: opposing anti-aging research is, right now, a content niche with real professional upside and zero personal cost.

You get credit for depth and wisdom. Nobody can prove you wrong because the technology doesn’t exist yet. And the audience for “actually, aging is fine” is enormous, because most people are already coping with aging and want validation that their situation is acceptable. The critics are supplying a product that the market wants to buy. The day age reversal works, that product becomes unsellable.

Nobody wants to read Why Aging Is Beautiful when their neighbor just reversed 30 years of biological decline.

Then there’s the quieter motivation nobody likes to talk about: some of these critics are old, and their opposition is at least partly about making peace with a situation they can’t currently escape. Kass is 86. Callahan died at 89. Meilaender is in his late 70s. Emanuel is 68 and staring down his own self-imposed deadline.

When you’ve already invested decades in constructing a narrative where your decline is meaningful and dignified, the sudden appearance of a technology that could have prevented all of it isn’t just disruptive. It’s psychologically threatening. Accepting that your suffering was unnecessary is one of the hardest things a human being can do. It’s far easier to insist the suffering was valuable and that everyone else should share it.

This is textbook cognitive dissonance resolution: I’ve already committed to this path, so the path must be correct, and anyone who takes a different path is wrong. Same psychology behind hazing rituals and sunk-cost fallacies. “I went through it, so you should too” is not a moral argument. It’s a psychological defense mechanism.

The Musk Arc: Watch the Drift Happen in Real Time

Elon Musk is a useful real-time case study because his position on aging has been subtly shifting, and the shift pattern is exactly how opposition gives way if the technology starts looking legitimate.

Early Musk has warned that reversing aging could cause an ossification of society. In 2021: “It is important for us to die because most of the time people don’t change their mind, they just die.” He said he had no interest in living forever and, unlike Jeff Bezos, poured no money into longevity startups.

Recently there’s a slight vibe shift. On the Moonshots with Peter Diamandis podcast in January 2026, Musk described aging as an engineering problem, calling it “not particularly hard” to solve: “You’re pre-programmed to die. And so if you change the program, you will live longer."

He said he’d “prefer to be dead” than live to 100 with dementia or as a burden. But notice the framing: he’s not saying he doesn’t want to reach 100. He’s saying he doesn’t want to reach 100 in decline. Remove the decline, and the objection likely disappears.

More revealing is what happened when the conversation turned personal. Musk was candid about the practical reality of aging: “I mean, I want things to not hurt.” He talked about back pain as an inevitability (“it seems like it’s only a matter of time before you get back pain”), arthritis, sleep disruption, the accumulating physical tax of getting older.

When Diamandis said he wanted to double his own lifespan (”I would like to double my lifespan for sure… at least 120, 150"), Musk didn’t shut it down or push back. He joked that “may you live forever" is one of the worst curses you could give someone, but when Diamandis countered his argument (”the head of GM didn’t have to die for Tesla to come along… in a meritocracy, the better ideas will dominate"), Musk just moved on.

That’s not the behavior of someone who genuinely believes longer lives are catastrophic — he’s telling you his body hurts and he wants it to stop. The philosophical objection is losing to the lived experience of aging, right there in the conversation.

Weeks later at Davos 2026, talking with BlackRock CEO Larry Fink, he went further: aging is “a very solvable problem," reversing it is “highly likely," and the synchronization of aging across the body means the underlying clock “must be incredibly obvious.”

His emphasis clearly shifted from we shouldn’t to we can and probably will.

The ossification argument doesn’t hold up historically. The 20th century saw global life expectancy more than double, from roughly 32 years in 1900 to over 73, and produced the most explosive sustained period of technological, scientific, and economic transformation in human history.

If longer lives mechanically produced stagnation, we should have seen civilizational stasis. We got quantum mechanics, nuclear energy, antibiotics, the microprocessor, the internet, spaceflight, and the genomic revolution instead.

The idea that society needs biological death to refresh its leadership is absurd: you can achieve that with term limits, rotating editorships, diversified funding, and open-science norms. Death is an accidental, barbarically low-resolution way of occasionally achieving what decent institutional design could do cleanly with better institutional design.

And there’s a competitive angle Musk would never admit publicly: imagine Jeff Bezos, any of his rivals (e.g. Sam Altman, Reid Hoffman, Zuck, etc.) and peers (Larry Ellison) use age reversal while he doesn’t. The idea that he’d voluntarily accept biological decline while the people he’s competing with operate at peak function is not realistic.

Why They Won’t Say No

The cheap martyrs treat this as a philosophical debate. It isn’t going to stay one. The moment age reversal works, it becomes a question of what people actually do under real conditions, and the conditions are stacked entirely against principled refusal.

Your rivals and peers will use it, and you’ll feel it. If your colleagues, your competitors, your spouse’s friends are all biologically 30 and you’re biologically 65, that asymmetry isn’t abstract. It’s in every meeting, every social gathering, every morning when your body reminds you that you chose principle over function. The pressure doesn’t come from pharmaceutical marketing. It comes from the people in your actual life who went through the treatment and now look, think, and move like they did three decades ago. That’s what breaks philosophical commitments.

The prisoner’s dilemma makes this inescapable. Think about someone you hate using age reversal while you don’t. Your rival is biologically 30, sharper, more energetic, more competitive, and you’re declining on principle. Now think about someone you love using it while you don’t. Your spouse, your best friend, your sibling is rejuvenated, vital, physically and cognitively decades younger than you. You’re the one slowing down, getting left behind, becoming the person they have to accommodate and eventually watch deteriorate. Both scenarios are psychologically unbearable for different reasons. The competitive scenario triggers loss aversion and status anxiety. The intimate scenario triggers something worse: the knowledge that you’re choosing to become a burden on the people you care about, choosing to make them watch you fall apart when you didn’t have to. Almost nobody holds the line in either case.

Loss aversion makes refusal unbearable over time. Before the therapy exists, opposing it is free. After it exists, refusing it is an active choice to accept deterioration, every single day, when the fix is available and your peers did it. Humans are profoundly loss-averse. Giving up a capability that’s on the table feels like having something taken from you, and that feeling doesn’t care about your published position on the ethics of aging.

The asymmetry is total. Using the treatment keeps all your options open. You can stop maintenance anytime and let yourself age again. Refusing it closes options permanently. You can’t reclaim the decades of vitality you sacrificed to prove a point. Choosing irreversible decline over reversible youth isn’t wisdom. It’s just leaving value on the table.

This Has Played Out Before, Every Single Time

If you think moral opposition to anti-aging will hold once it works, consider how well moral opposition has held up against every other medical technology that became personally useful.

The pattern is always the same, and it always ends the same way.

1. IVF. The Catholic Church formally opposes in vitro fertilization. Doctrinally. Unambiguously. And yet Catholic couples use fertility treatments in large numbers. Pew Research found that 71% of weekly Mass attenders say the Church should allow it, because when the benefit is your child, doctrine bends.

2. Embryonic stem cell research. The opposition in the early 2000s was ferocious. Bush vetoed legislation expanding embryonic stem cell research funding, political campaigns organized around stopping it, and the Catholic Church drew hard doctrinal lines around embryo destruction. Two decades later, stem cell therapies are increasingly mainstream, Catholics are getting stem cell injections for their knees and joints without a second thought, and the moral fire that once drove presidential vetoes has cooled into background noise, because the treatments started working and people wanted them.

3. Organ transplants. When the first heart transplants happened in the late 1960s, the moral objections were everywhere: playing God, violating the natural order, desecrating the dead. Now it’s standard-of-care and nobody debates it at dinner.

4. Cosmetic surgery. Entire cultural movements were built around opposing the vanity and superficiality of it. Shallow. Fake. An insult to natural aging. It’s now a $70+ billion global industry, used by many of the same cultural figures who once signaled against it.

5. The pill. When oral contraceptives arrived in the 1960s, the moral opposition from religious institutions and cultural conservatives was overwhelming. Within a generation it was one of the most commonly prescribed drugs in America, including among populations whose religious leaders formally opposed it.

The pattern is always the same. Opposition peaks when the technology is new and the benefit is theoretical. The moment the benefit becomes personal (your fertility, your heart, your face, your body) the moral architecture crumbles. Age reversal won’t be different. The benefit is as personal as it gets.

What Refusal Actually Looks Like Once It’s Real

Once age reversal actually works, opposing it stops being a philosophical position and becomes an actively anti-life, anti-human stance.

You’re not arguing for “natural limits” anymore. You’re arguing that people should suffer cognitive decline, chronic pain, frailty, and death from a treatable condition, and that society should let them, or even make them. That’s not wisdom but cruelty masquerading as depth.

The coercion problem is the part that should bother everyone.

• You can personally choose not to use age reversal. That’s your right and nobody serious is contesting it.

• But the critics’ position doesn’t stop at personal opt-out. The logical endpoint is restriction: regulation that limits access, framing the technology as dangerous enough to suppress.

• And that means forcing other people to age, suffer, and die from a condition that will eventually have a treatment, because they’ve decided on your behalf that mortality is a virtue.

You can turn it down. But you sure as hell shouldn’t be allowed to force someone else to die because you’ve made philosophical peace with death and think everyone else should too. That’s not a “genuine social concern" — you are trying to impose your beliefs on others’ lives — and it’s no different in principle from any other attempt to restrict medical autonomy.

And the social dynamics flip completely. Right now, opposing anti-aging research is a low-cost way to signal thoughtfulness. It gets you panel invitations, book deals, and media coverage. The technology doesn’t work, so the stance is pure upside with zero personal sacrifice, which is exactly what makes it a clout play.

The moment the technology works, the clout calculus reverses. Now you’re the person arguing that people should age and die from something fixable. The social reward for opposing age reversal evaporates the instant the technology delivers, and so will the opposition.

Emanuel is already the proof of concept. He made the most public version of this stance possible and he’s already softening it a decade out. That’s without age reversal even existing. The stance was always going to collapse, because it was never really about principle. It was about the social positioning available to someone who doesn’t have to back it up yet.

The Market Already Told You the Answer

Will some people genuinely opt out? Sure. Some will prefer to age naturally, decline on their own terms, or exit when they’re ready. Nobody serious would stop them.

However, framing that minority preference as the likely mainstream response to working age reversal is delusional. Look at what people already do without age reversal: the global anti-aging market (skincare, supplements, hormone therapy, cosmetic procedures, fitness optimization, biohacking) is already worth hundreds of billions of dollars, and none of it actually reverses aging.

People spend enormous amounts of money and effort on products that make them look or feel slightly younger, with marginal or zero effect on underlying biology.

The demand for youth is one of the largest consumer markets on the planet, and it exists for treatments that are mostly placebos.

Now imagine what happens when something actually works, when the effect size isn’t a slightly firmer jawline but a 30-year biological reset. The uptake won’t be gradual. It’ll be a stampede.

And the critics who spent years explaining why aging is beautiful will be somewhere in the middle of it, along with everyone else.

This isn’t a prediction about their character, it’s a prediction about human nature… and human nature has a very strong track record on this one.

Then he did what he always does: warned that living “for a very long time” risks “an ossification of society” that may become “stultifying” and “just lack vibrancy.”

He’s been making the same point for years; in 2021 he stated:

“It is important for us to die because most of the time people don’t change their mind, they just die."

He’s also used the political variant: if we live for too long “leadership never dies.”

Strip it to its logical skeleton and the argument is:

Longevity → people stick around longer → people don’t update their beliefs → idea turnover depends on biological death → curing aging = societal stagnation.

Every link in that chain is either absurdly wrong, overstated, or self-defeating.

1. Reversing aging ≠ living forever (category error)

Musk’s logic quietly swaps biological rejuvenation for infinite invincibility, and those are completely different things. Reversing biological aging doesn’t make anyone immortal.

People still die from accidents, violence, infections, cancer, pandemics, disasters, warfare, and plain bad luck. Turnover never goes to zero.

The “leadership never dies” scenario isn’t something that follows from curing biological aging; it’s mostly just a byproduct of inefficient power rotation.

2. Aging is the mechanism of the very rigidity Musk fears

If Musk is worried about rigidity, he’s aiming at the wrong target. The target is aging itself.

There’s a neurobiological paradox sitting right at the heart of Musk’s position that he never addresses: he’s arguing (A) we should preserve the biological process that generates rigidity because (B) death occasionally refreshes it.

That’s completely backwards.

Why do people actually get “stuck in their ways”? Because their brains deteriorate from age.

1. Cognitive flexibility declines substantially in older adulthood (task switching / set-shifting); adults 60+ show significantly lower flexibility than younger adults.

2. Processing speed and executive function, the core components of fluid cognition, show well-documented age-related declines.

3. Openness to experience, the personality trait most tied to intellectual curiosity and receptivity to novelty, is negatively associated with age in large national samples.

4. Neuroplasticity, the brain’s fundamental capacity to rewire itself in response to new information, decreases with age.

At the circuit level, aging is associated with neural dedifferentiation (reduced neural selectivity), which makes learning less clean and belief-updating more effortful.

What Musk calls “ossification” is a symptom of the disease he doesn’t want to cure.

And if rejuvenation actually restores the brain to a biologically youthful state, the relevant comparison isn’t “90-year-old mind preserved forever.”

It’s someone with the neuroplasticity, processing speed, and task switching of a young brain who also carries decades of accumulated context, hard-won pattern recognition, and cross-domain intuition.

Brains still prune, consolidate, and forget, so this isn’t an infinite-storage superhero. A 150-year-old with a biologically 25-year-old brain would still be mortal — just able to learn like it’s 25 while carrying a century of context.

Put differently: Musk’s argument is structurally equivalent to “We probably shouldn’t cure Alzheimer’s, because if people remember too much, they might not be open to new memories.”

3. His “people don’t change their minds” premise is overstated and irrelevant

For Musk’s argument to work, you need a near-absolute version of cognitive stubbornness: people can’t update, so you need them to die.

But people update their views all the time.

• Scientists abandon paradigms.

• Entrepreneurs pivot.

• Voters shift.

Openness shows real individual-level variation and responsiveness to environment even as population averages drift with age.

And even if many individuals do update slowly, societal idea-selection is competition-limited, not death-limited. New ideas don’t win because their opponents die, they win because they work.

• Better technology displaces worse technology.

• Markets reward efficiency.

• Status competition drives adoption.

• War selects for functional systems over dysfunctional ones.

Death occasionally removes a blocker, sure, but the heavy lifting has always been done by competition, incentives, and demonstrated superiority. Confusing (A) the occasional removal of an obstacle with (B) the engine of progress is a basic attribution error.

4. We already ran the partial test, and society didn’t freeze

No one has ever observed a society with widespread rejuvenation, so let’s be clear: Musk’s ossification claim is conjecture asserted with confidence.

The counterfactual has literally never been run.

What we already have is a massive partial test, and it points in the wrong direction for him. Global life expectancy at birth went from ~32 years in 1900 to ~71 by 2021 and reached ~73 by 2023; more than doubled in a single century.

If longer lives mechanically produced societal ossification, the 20th century should have been a slow crawl into civilizational stasis.

Instead we got the most explosive sustained period of technological, scientific, and economic transformation in human history: quantum mechanics, nuclear energy, antibiotics, the Green Revolution, the microprocessor, the internet, spaceflight, the genomic revolution.

None of this “proves” that extreme rejuvenation has zero risks. But it does kill the word “inevitable.”

“Longer life → ossified society” isn’t a law of nature and the strongest historical trend we have runs in the opposite direction.

5. The reductio: his logic implies shorter lives = more progress

Push Musk’s logic one step further and it collapses under its own weight.

If death is instrumentally good because it refreshes ideas, then higher mortality and shorter lifespans should produce more innovation. Why stop at 80? Why not 60? 40?

High-mortality societies were not innovation utopias. They were trapped by disease burden, instability, low human-capital accumulation, and catastrophically short planning horizons.

You can’t finish a cathedral if you’re dead before the foundation sets. People die mid-project, mid-mentorship, mid-breakthrough, and everything they spent decades learning vanishes with them. The apprentice who was halfway through absorbing a master’s lifetime of tacit knowledge has to start over with someone else, or just doesn’t.

Multiply that across every field, every generation, and you get a civilization that’s perpetually re-learning what it already knew instead of compounding on it. The correlation between rising life expectancy and accelerating civilizational complexity is one of the most robust patterns in human history, and this is a big part of why.

The idea that “people should die so the Overton window refreshes” is trying to solve institutional problems with biological tragedy; the rough equivalent of burning a house down to kill a spider.

For Musk’s story to be true, you need a strong nonlinearity: modest lifespan gains are fine, but somewhere past today a switch flips and extra healthy decades start freezing society. If that tipping point exists, Musk has to explain how it would overwhelm more accumulated human capital with more plastic brains.

6. Ossification is never total, and may even be beneficial

Let’s grant Musk’s framing entirely, just for the sake of argument. “Ossification” in a post-aging world still wouldn’t be absolute.

Exogenous shocks, competitive pressures, technological disruption, and the full force of market selection don’t go away just because senescence does.

Curing aging removes one specific source of turnover (biological decay) while every other source remains fully intact.

There’s also something worth saying that rarely gets said in these debates: in complex systems, some stability is actually functional. (1) Not all change is progress and (2) not all continuity is stagnation. The assumption that more turnover is always better deserves scrutiny of its own.

A world where experienced, cognitively intact people can sustain long-term projects, maintain institutional memory, and compound knowledge across centuries might actually be more adaptive than one that forces a civilizational hard-reset every 80 years and blows enormous resources re-teaching the basics to each new generation.

7. The “funeral effect” evidence refutes his causal story

The steelman version of Musk’s claim goes back to Max Planck:

“Science advances one funeral at a time.”

And there’s a well-known economics paper: “Does Science Advance One Funeral at a Time?” (Azoulay, Fons-Rosen, and Graff Zivin) — that tested a version of this empirically in the life sciences.

After eminent scientists die, outsiders enter those subfields at higher rates, and the resulting work tends to be disproportionately highly cited. Great paper. But if your takeaway is “you need death to advance science” — you are a certified moron.

Sounds like it supports Musk, right? Look closer at the mechanism: the mechanism is incumbent control, not some mystical need for funerals.

Star incumbents dominate attention, funding access, and editorial control during their lifetimes. When they die, the bottleneck loosens and intellectual diversification follows.

You could achieve this same effect a variety of ways with zero deaths from aging, including:

1. Stake-based peer review

2. Rotating editorships

3. Term limits on committee seats

4. Diversified funding mechanisms

5. Open-data/open-science norms

6. Adversarial collaboration incentives

Death happens to be an accidental, barbarically low-resolution way of occasionally achieving what decent institutional adjustments could do cleanly. Anyone who thinks for more than 2 seconds quickly realizes that you don’t actually need death to achieve a refresh.

Additionally, nobody in this debate connects back to the neurobiology: gatekeepers themselves are aging and gatekeeping may be a direct byproduct of aging! The star incumbents dominating various fields aren’t just powerful… they’re old and getting older, which means their cognitive flexibility, openness to new ideas, and willingness to update are actively degrading in real time.

Aging creates the rigidity Musk worries about at the individual level and synergistically feeds into institutional lock-in by ensuring that the people sitting on panels, editing journals, and controlling funding are becoming biologically less capable of recognizing good new ideas the longer they hold those positions.

Rejuvenation would actually attack both problems at once: (1) keep the institutions populated by minds that can still flex, while (2) buying time to implement the structural reforms that address gatekeeping directly.

When “death improves innovation” in some narrow context, that tells you something about incumbent control but nothing about whether humans should die of old age.

And even granting the dynamic some historical validity, there’s no reason to assume it stays net-positive going forward. The past doesn’t always extrapolate to the future; the inverse could occur.

As knowledge burdens rise and it takes longer to reach the frontier, each funeral gets more expensive. Losing a frontier-capable mind in 1920, when certain PhDs took a few years and the relevant literature fit on a bookshelf, is a very different proposition than losing one in 2025, when reaching the cutting edge requires a decade-plus of specialized training.

The cost of the “refresh” keeps climbing while the pool of people capable of doing frontier work keeps shrinking. At some point the funerals start costing more than whatever gatekeeping bottleneck they break and we may already be past that point.

8. Gatekeeping is the actual bottleneck, and it has nothing to do with lifespan

Real ossification-like dynamics are happening right now, and not one of them has anything to do with how long people live:

• Peer review has become structurally conservative: “weaknesses” routinely dominate “strengths” in scoring and negative information is overweighted. NIH’s CSRAC notes that NIH review is frequently criticized as risk-averse — favoring “sure thing” established ideas over potentially high-impact but unproven ones.

• Papers and patents have become less disruptive over time, as measured by formal disruption indices. That’s exactly the kind of “ossification” people worry about, and it’s happening under ordinary human mortality.

• Tenure systems, bureaucratic red tape, reputational penalties for failure, and risk-averse funding structures stifle innovation every day through institutional design, not biology.

If ossification is real today (and the evidence suggests it is), it’s happening despite people dying on schedule. Death clearly isn’t solving it.

Whatever’s behind declining disruptiveness in science, “people aren’t dying fast enough” isn’t a serious diagnosis.

9. The “leaders never die” fear is a governance design failure

Musk’s strongest political worry: very long lives could mean longer entrenchment in high-power positions.