Are The AI Labs Getting Nationalized?

Summary

• Ari Paul’s headline warning: the frontier AI labs could be quietly nationalized, and he hasn’t seen people pricing it. He has “AI friends who are prepared to have a Los Alamos style lockdown” — full US government control within three years, nuclear-project rules: government-issued phones, travel permission, no leaving the country — because “if the US doesn’t do that, we will be leapfrogged by China.” Trump has already taken direct stakes in companies and said he shut down “mythos”; Ari isn’t saying the labs are bad investments, “just I haven’t seen people seriously talk about discounting these factors.”
• His deeper skepticism of the model makers: assume the IP is already stolen — “anything being produced by Claude, by Meta, by Google, you have to assume is in the hands of Russia, China, North Korea.” And if OpenAI ever builds a model that beats the stock market, it probably never reaches retail or enterprise users, and shareholders don’t see it either: “Sam Altman’s going to run it privately, or the engineer who discovered it is going to run it privately.” He frames that as a likelihood, not a tail risk.
• AI is where crypto was in 2021-22: the tailwind that gave anything with an AI name a 100x is over, prices in general still go up, but “some people will lose money on AI over the next three to five years.” Every real tech birth — PCs, internet, railroads — still saw 95% of companies fail, and OpenAI’s Sora is the tell: popular, working, loved, “massively money losing,” shut down.
• The data-center buildout may rhyme with late-90s fiber overbuilding: investors are always “right directionally but wrong in timeline — the five years is almost always 20.” New papers arrive weekly that “dramatically cut the hardware needs to achieve the same result,” and the bottleneck keeps rotating — power, rare earths, silver, optics, turbines — a full-time rotational trade Ari envies but won’t dabble in.
• On crypto: the adoption thesis that made Bitcoin “free money” ended by 2021-23 — “it might be early in terms of global adoption, but it’s certainly not early in a brand recognition sense.” Inefficiencies have thinned to Jane Street’s benefit, yet crypto is still a great place to look for a small trader with limited capital.
• What he’d own instead: locked-in distribution — the Chris Han (likely Chris Hohn) thesis that AI cuts costs for Visa-like moats while revenue holds, because “anything purely digital moves at infinite speed; anything that involves humans goes at human scale.” If he relaunched BlockTower he’d run a third the headcount — LLMs instead of junior analysts.
• Macro close: “both sides are socialists now” — Trump “the most socialist president since FDR” — and Ari expects “the blow up of American capitalism ahead of us sometime in the next five years,” then 50 years of secular growth. Caveat verbatim: “if AI doesn’t kill us all” — AI is a singularity, so “any historical analogy we want to make, we need to be cautious about.”

Deep dive

1. The original edge was table selection, not genius — and the table got sharp

• Ari’s pitch raising BlockTower in 2017 is the cleanest statement of the thesis: “You don’t need to believe in crypto. You also don’t need to believe that I’m the best trader… This is the most inefficient market in the world” — the genuinely great investors were locked out by regulation, so a big fish could grow with the pond.
• The best specimen of that inefficiency: during a 2017 Coinbase API outage, with the website updating “once every 3 seconds” and market makers blind, Bitcoin swung 20% for an hour and Ari flipped it roughly five times — “sell at 15k, buy at 11k” — an edge that existed purely “because you didn’t know if you were getting filled” and institutions couldn’t stomach that.
• The decay curve since: 30% Korea arbitrages, then DeFi summer’s ~70% risk-adjusted yields, then “everyone’s piling into Luna at 12%” when default was already foreseeable — worsening risk, far lower return, as more capital chased the same deals. Today Jane Street “mostly crushes it, I think,” the remaining arbs live on scammy exchanges for good reason, and yet: for a small trader with limited capital, “crypto is a great place to look… it’s just not that ratio risk-reward is probably never going to be as good again.”

2. The Bitcoin adoption thesis is finished — “it’s certainly not early”

• The 2016 crystal ball, loosely built on Soros reflexivity: watch endowment people learn about Bitcoin, buy personally, then institutionally — if nothing new happens, 10x the potential buyers get access and socialization, and “if even a small fraction of those people buy, the price has to go up.”
• That thesis ended “maybe by end of 2021, certainly by 2023” — by the time of a Bitcoin president keynoting a Bitcoin conference and talking strategic reserve, and El Salvador adopting then “largely abandoning” it. His hedge preserved exactly: “That’s not to say it won’t continue going higher, but it’s no longer free money and it hasn’t been for a while.” Now “you need something new to happen for a bullish thesis.”

3. Crypto never got the best and brightest — regulation and AI made sure of it

• The disillusionment wasn’t hacks or failed token experiments (utility tokens are “not a bad idea… it’s going to take a lot of iterating”) but “the aggregate lack of forward progress,” which Ari pins on incentives: Biden-era enforcement hit the registered and regulated while sparing the worst actors, which “basically guaranteed the industry leaders will be Binance over Coinbase” and pushed out anyone unwilling to take jail risk.
• Meanwhile AI soaked up the actual geniuses — his example is the DeepSeek founder, “compared to Terence Tao… winning math Olympiads” at 10. “We don’t have that many of them in crypto,” and those who exist are “relegated to engineering positions,” their work ignored or politicized.

4. In 2017 crypto was the only game in town; now AI puts everything in play

• Avi’s framing: crypto captured attention then because nothing else was small and exciting; today Caterpillar rips selling data-center power generation and Trump talks up nuclear startups. Ari agrees from the inside: endowment finance 12 years ago “felt and looked pretty boring,” a terrible era for active management that pushed institutions into frontier markets (which mostly lost) — and he was surprised to find his former shop Chicago had been an early, aggressive direct investor in data centers on the same be-creative-for-yield thesis.
• Institutions won’t return to crypto easily: direct programs from five-six years ago “went badly, right? Most altcoins are down,” with the tops in most alts made four years ago — “it’s hard for them to make an argument that the next time will go differently.”
• The through-line into AI: “so much is AI… it’s disrupting kind of everything, which puts everything in play — every boring industry, every boring asset class suddenly might be high returning with the right thesis.”

5. Don’t pay full price for the model makers: leaky IP, private alpha, and Los Alamos

• Ari missed AI’s first wave (crypto focus, then post-BlockTower burnout) and now spends “an hour, two hours a day” studying for the next one — hunting the 2015-16-style decade-long crystal ball rather than punting. His first strong view: skepticism on the IP of the generators. He was debating a senior Meta ML friend “like 48 hours before Claude accidentally posted a good chunk of their codebase publicly” — between corporate poaching and nation-state espionage, “anything being produced by Claude, by Meta, by Google, you have to assume is in the hands of Russia, China, North Korea.” So how do you value the IP?
• The private-alpha problem: a market-beating model probably never reaches retail or enterprise users, and shareholders don’t see it either — “why is OpenAI going to run that [publicly]? Sam Altman’s going to run it privately.” Avi’s pushback, conceded: the labs could still be good investments without capturing 100% of value — “both could be true.” Ari’s real claim is about pricing: SpaceX at “well over a trillion dollars on what was it, like 20 billion in revenue” isn’t insane, but “is it a hundred billion valuable or two trillion valuable? I have no idea” — and investors “were not incorporating these factors as risks.”
• Then the nationalization scenario, delivered twice for emphasis: AI friends expect “a Los Alamos style lockdown… the US government to fully control them within three years.” The nuclear-project precedent — scientists in the desert, spied on, no phone calls — applied to a top Meta ML engineer: government-issued cell phone, travel by permission only. He flags it as hypothesis (“I’m not sure I agree”), but notes other governments, “at least Chinese,” are already doing versions of it.

6. The data-center buildout may be fiber in 1999 — bottlenecks rotate

• Ari’s signature framing on tech bubbles: investors “correctly identify a world-changing technology… and they’re always right directionally but wrong in timeline. The five years is almost always 20.” The fiber overbuild didn’t mean internet demand stopped growing — other bottlenecks had to clear first.
• Applied to now: a trillion dollars of data centers in a couple of years was “basically a man on the moon project,” but “week after week new AI papers are being published that dramatically cut the hardware needs to achieve the same result” — another five trillion at compute is no longer the good ROI. Progress rotates: overbuild physical capital, then discover the limiting factor is power, rare earths, silver, optics, or turbines.
• Trading that rotation is what BlockTower did with 1-3 month crypto narratives — park where capital will flow, eat 20% drawdown, catch the 5x — but “you either need to be really fast and on the ball as the best active managers are today… I’m envious of them,” and it’s a full-time job he refuses to dabble in.

7. Jonah’s pushback: AI is secular, not a trade — Ari: yes, and 95% still fail

• Jonah’s sharpest contribution: at BlockTower “everything we touched except Bitcoin was a trade” — he recalls the Q4 2021 NEAR/Phantom position where “we’re both like, okay, we know what this is” — whereas with AI “it’s very possible that in 10 years everything that touches AI is just higher.” Part of being a good trader is “understanding what made you your money.”
• Ari fully agrees on the tailwind, then re-anchors: with every purely positive tech birth — internet stocks in ‘99, PC brands of the early 60s, railroads in the 1850s — “still 95% of the startups fail.” And AI “enables disruption, including to itself”: today’s leaders are under 10 years old and can be leapfrogged exactly as they leapfrogged. Sora is his proof that product-market fit isn’t enough — “they built a working great product, they got the users, and yet it was massively money losing.” Jonah’s tag: “kind of like all the early delivery apps.”
• The passive-basket trap, as told: in late 2017 a market-cap-weighted basket meant “putting 20% of your money into IOTA… it’s garbage, it’s going to go to zero.” Hypothetically Bitcoin hits 200,000 and Ethereum 6,000 while IOTA sits at zero — when too much of the market cap is overpriced, passive fails even with the tailwind. Hence the verdict: “AI is maybe where crypto was in 2021 or 2022… some people will lose money on AI over the next three to five years.”

8. What to actually own: locked-in distribution, because humans are the bottleneck

• The one bullish framework Ari endorses comes from “Chris Han at Founders Fund” (likely Chris Hohn): buy locked-in distribution — government monopolies, regulatory capture, pipelines too expensive to replicate. Visa “at face value should be the easiest to disrupt company in the world,” yet Visa can adopt stablecoins and AI to slash its own costs — OpenAI offering half the processing fees “doesn’t really matter to consumers” against brand and ubiquity. Falling costs plus flat or growing revenue = great buys.
• The mechanism underneath, in his words: “anything that’s purely digital moves at infinite speed; anything that involves humans goes at human scale. If you need a human to sign off, it doesn’t matter how fast the AI is” — unions, regulators, town-by-town ordinances all resist disruption.
• His own revealed preference: BlockTower ran ~$2B with 33 people; “if I did it over again I think I’d probably have a third the headcount” — one AI-powered analyst doing the work of eight, producing “a million stock-specific reports… in an hour.” Beyond distribution moats, though, he admits “I have more fears than bullish conviction” — disruption looks attractive everywhere, which is exactly the problem.

9. The biohacking detour: TMS, a tuberculosis antibiotic, and 30-100% faster learning

• Post-BlockTower burnout (“running a hedge fund in crypto can do that,” Avi notes) sent Ari down a health rabbit hole: a “medical quarterback” — “the high-net-worth version of a GP” — sent him to London, where four specialists diagnosed his back differently and all converged on hip imbalances no US doctor had mentioned. His democratized version: $75-150 monthly blood panels plus an LLM for analysis; Jonah says “working with an LLM beats them for most analysis,” while Ari says the very best doctors absolutely trounce anything the LLM can do, but it’s so hard to find those doctors.
• The frontier claim: daily TMS (Ari tentatively called it “transmagnetic stimulation,” with his device sold as an “aesthetic device” to dodge medical-device registration and used entirely off-label) plus DCS — a “scary,” neurotoxic tuberculosis antibiotic (likely D-cycloserine) that in small doses “10xes your neuroplasticity but only for two to three hours, once a week” — “it gives you the brain of a seven-year-old.” His neuroscientist claimed 5x learning; Ari’s honest calibration: TMS alone gave maybe 30%, “positive but within placebo range,” while on DCS he learned a juggling move in five minutes that normally takes an hour or two. Evidence-graded conclusion: not 5x, but “30 to 100% improvement” is plausible.
• On peptides he’s deliberately unhyped: he takes the Wolverine stack (BPC-157 and TB-500) under close monitoring from a vetted supply chain, but “the peptide industry is functioning very much like the supplement industry 20 years ago. Most of it’s garbage” — most supplements produce “expensive urine,” some are tainted, yet “among the hundred peptides being mass sold, at least a few definitely are good” — GLP-1s and GLP-3s are cited as effective examples.

10. Peak chaos, socialists on both sides — then 50 years of growth, if AI allows

• Ari thinks “we’re at or close to peak chaos,” but it “gets worse before it gets better” — maybe five more years of eroding social trust. The rising left terrifies him; he says his politics have been unchanged for 20 years and that he’s probably a moderate: the “Mandani” crew (likely Mamdani) “many of them openly say they’re communist,” and two newly elected New York State Assembly members “literally wrote their goal is to destroy the United States. No exaggeration, their words.” He’s more optimistic on the right — Trump won’t try to retain power “in the current context,” and a Rubio or Vance could pull Republicans back toward governing.
• The investor takeaway: “both sides are socialists now.” Trump “ran railing against socialism [and] has been the most socialist president since FDR” — direct government stakes in multiple companies, shutting down “mythos.” The horseshoe: “both sides are woke now… horrible socially, culturally, economically, scientifically.”
• The Fourth Turning arc, hedges intact: “we probably have the blow up of American capitalism ahead of us sometime in the next five years… something that feels like a more holistic collapse of Western democracy,” and then 50 years of secular growth — “if AI doesn’t kill us all, who knows? Any long-term or cyclical prediction is kind of falsified by AI… it means history will not repeat.”

Verification Notes

• The raw captions render Ari’s reference as “mythos”; whether he meant M&A or a named entity is unresolved.