Are The AI Labs Getting Nationalized?

Some AI researchers expect the US government to fully control AI labs within three years under a Los Alamos style lockdown, according to Ari Paul. The Manhattan Project analogy involves placing engineers in restricted facilities, issuing government phones, limiting travel, and treating AI as a military program. Paul argues that without military-style control, China will leapfrog the US through corporate espionage. He frames this as an investment risk rather than a firm prediction and notes he has not seen serious discussion of discounting nationalization risk when valuing AI companies. Anthropic and OpenAI are already described as pseudo-nationalized, and some AI insiders are reportedly prepared for a full government takeover.

Corporate espionage between AI firms is described as constant, with Google and Meta actively poaching each other's top machine learning staff. Claude accidentally posted significant portions of its codebase publicly roughly three months before the recording. Nation-state actors including Russia, China, and North Korea are assumed to already possess IP from major AI labs. Given this environment, the speaker is skeptical about the durable IP value created by model developers like Claude, Meta, and Google. The speaker also raises the scenario that if OpenAI or Anthropic built a model capable of beating the stock market, insiders such as Sam Altman would run it privately rather than commercializing it through the company.

The current data center buildout is compared to the overbuilding of fiber optic infrastructure during the late 1990s internet bubble. New AI papers published weekly are dramatically cutting hardware requirements to achieve equivalent results, and the bottleneck has shifted such that deploying another five trillion dollars at compute is no longer considered good ROI. Investing in AI infrastructure plays is characterized as a rotational trade requiring full-time active management. Caterpillar is cited as generating meaningful revenue selling power generators for data centers, and Trump discussing funding for new nuclear startups has generated excitement around uranium investments.

The speakers place AI at roughly where crypto was in 2021 or 2022, meaning the easiest money is gone and the era of 100x returns on any AI name is over. Historical technology waves including personal computers, the internet, and railroads saw roughly 95 percent of startups fail and most leading public companies go bankrupt within five years. AI also enables disruption of itself, meaning current leaders could be leapfrogged just as they leapfrogged prior incumbents. OpenAI shutting down Sora despite its popularity is cited as evidence that even well-received AI products can fail to find a path to profitability. Two frameworks are offered for navigating AI: a trading rotational mentality focused on six-month pipeline and pricing power, or a VC mindset making many bets knowing most will fail. Constructing a passive market-cap-weighted basket is warned against when assets are highly valued, citing a 2017 crypto basket that would have allocated roughly 20 percent to IOTA, which went to near zero.

An investment thesis attributed to Chris Hohn at Founders Fund centers on companies with locked-in distribution including government monopolies, regulatory capture, and expensive-to-replicate physical pipelines. The core logic is that AI cannot easily disrupt distribution companies because replicating local distribution involves local regulation and ordinances town by town. Anything touching human beings is described as the bottleneck for AI, since regulatory approval, lobbying, and union negotiations cannot be accelerated by AI speed alone. Visa is cited as an example where even a competitor offering half the fees might fail to displace it due to brand loyalty and universal acceptance, while Visa itself can adopt stablecoins and AI to cut its own costs.

Paul founded BlockTower as a crypto hedge fund roughly from 2017 to 2023. A 30 percent arbitrage between South Korea and the US persisted into late 2017 due to regulation and lack of allocated arbitrage capital. During a Coinbase API outage in 2017, he click-traded Bitcoin five times in roughly one hour, selling near fifteen thousand dollars and buying near eleven thousand dollars repeatedly, exploiting 20 percent swings that persisted because no one could trade electronically. Paul attributes his disillusionment with crypto to misaligned incentives, with projects raising up to a billion dollars but spending only a few percentage points on quality engineering. Ambiguous regulation under prior administrations pushed ethical actors out and allowed the least ethical participants to lead the industry. The secular tailwind thesis for Bitcoin largely ended by late 2021 and certainly by 2023, with most altcoins in the managed portfolio still far below their all-time highs.

A single AI-empowered analyst can now produce the equivalent output of eight analysts, and LLMs can synthesize roughly 200 peer-reviewed medical studies into a meta-analysis in approximately ten minutes versus ten hours previously. The speakers caution that taking 30 supplements daily risks exposure to toxic heavy metals, and injections carry existential infection risk. D-cycloserine, an antibiotic used for tuberculosis, produces approximately ten times normal neuroplasticity in small doses but is usable only roughly once a week before becoming harmful. The speakers predict social cohesion will continue declining for another five to ten years before stabilizing, describe Trump as the most socialist US president at least since FDR given direct government stakes in multiple companies, and caveat all long-term predictions by noting that AI as a singularity means historical analogies will not repeat in expected ways.