🎧 Prefer to listen?

The most instructive trade of the AI boom this year wasn’t a buy — it was what a hedge fund refused to sell. Situational Awareness, the fund run by 25-year-old ex-OpenAI researcher Leopold Aschenbrenner, just sold most of its public stock portfolio to Ken Griffin’s Citadel after a brutal month. It kept every share of its Anthropic stake.

If you only read the headline, that looks like a rich kid’s bad month. Look closer and it’s a map of where the smartest money in AI now thinks value actually lives — public AI infrastructure trades versus private AI labs — and the logic behind that split is worth understanding even if you’ll never manage a dollar. It lands in a year when the AI race isn’t about Anthropic vs OpenAI anymore but about entire ecosystems — compute, energy, capital — and this fund just told you which layer it thinks still has upside.

What actually happened

Situational Awareness was the hottest AI fund of the year: 439% returns through June, assets that reportedly swelled to $45 billion at peak. Aschenbrenner — Columbia valedictorian at 19, briefly on OpenAI’s superalignment team before being dismissed — built his thesis on a simple claim: scaling AI would require a massive buildout of semiconductors, compute, memory, and energy. For most of two years, that trade printed money.

Then the AI infrastructure selloff hit. Public investors started asking an uncomfortable question: where’s the revenue? Some of the fund’s biggest positions — memory chipmakers SK Hynix and Sandisk, clean-energy developer Bloom Energy, cloud provider Nebius — all fell more than 30% in a month. The losses were amplified by leverage, the borrowed-money strategy that magnifies both wins and drawdowns. Assets fell from roughly $20 billion to about $10 billion, and Citadel stepped in to buy the bulk of the public portfolio — a classic Griffin move of picking through leveraged unwinds.

Here’s the part worth pausing on. In a July 24 letter to investors, Aschenbrenner called the selloff one of the best buying opportunities since early last year and invited fresh capital from August 1. Bloomberg reported the appeal didn’t bring the commitments he hoped for. His clients declined to double down — and Citadel, notably, did buy, suggesting the sector view isn’t dead so much as the leveraged, public-markets version of it is.

The Anthropic shares he wouldn’t sell

While dumping public stocks, the fund held onto its private positions — most notably a stake in Anthropic now valued around $5 billion. Anthropic was last valued at $965 billion in its May Series H round and is expected to go public as soon as October, potentially at an even higher price. The fund’s other private bets — chip startup MatX and data-center company Fluidstack — sit in the same bucket: private, illiquid, and aligned with the layer of AI that actually builds the models.

Read that split honestly. The public AI infrastructure trade (chips, energy, neoclouds) was a theme anyone could buy through a brokerage app — so everyone did, prices ran ahead of revenue, and the reversal was violent. The private AI lab position is something different: it’s a claim on the companies doing the actual work, locked up until IPO, unavailable to retail money at these prices. Aschenbrenner’s portfolio just admitted which claim he believes in more.

There’s also a less flattering reading worth naming: private marks are appraisals, not market prices. A $5 billion stake is “valued” at the last round’s terms, and it only becomes real money if the October IPO happens at or near that level. Keeping it could be conviction — or the absence of a market to sell into. Both readings coexist, and anyone telling you which one is certain is selling something.

Why builders should care

You’re not trading stocks — but the same divergence is playing out in your world, just smaller:

  • The money says the labs are the asset. Not the tools, not the resellers, not the infrastructure round-trip. If capital concentrates at the model layer — with Anthropic at a $965 billion valuation and an October IPO ahead — expect model companies to keep absorbing more of the value chain, and to keep raising prices for the layers above. We’ve tracked that pressure already in what solo builders are paying for AI coding tools.
  • Themes are not businesses. “AI infrastructure” was a tradeable story until it wasn’t. The same applies to builder trends: “agentic everything” is a theme; a working automation that saves a real business ten hours a week is a business. The fund’s wipeout is what happens when leverage meets a theme.
  • Conviction needs a survival horizon. Aschenbrenner could hold his Anthropic stake through the drawdown because private shares can’t be margin-called the way leveraged public positions can. Your equivalent: don’t build workflows so overextended on one provider’s pricing that a single repricing forces you to unwind everything at once. The memory-chip supply crunch is a reminder that the inputs to your stack are themselves volatile.
  • Discipline beats heroics. The fund’s clients who refused to add capital in August weren’t pessimists — they were applying a risk rule after a 439% run. Solo builders need the same: a rule for when you stop expanding your stack and just run what works.

None of this is investment advice — it’s pattern-reading. The useful question isn’t “should I buy what Citadel bought,” it’s “which parts of my AI dependency are speculative themes, and which are load-bearing?”

The bigger picture

Zoom out and the episode says something about the whole AI economy’s structure. The public markets just repriced AI infrastructure — the picks and shovels — while private capital keeps marking up AI labs. When Anthropic does go public, possibly in October, that tension gets resolved in one direction or another, in public, with a ticker. Until then, the Aschenbrenner trade is the clearest real-money statement of the era’s core bet: that whoever owns the models owns the future, and everything else is cyclical. It also lands amid a regulatory climate where who decides when AI is too dangerous is still an open question — one more variable that could reprice the labs’ valuations overnight.

For a deeper cut on how Anthropic’s positioning got here, our piece on how Anthropic may have talked itself into an AI export ban covers the policy side of the same bet.

The bottom line

A fund built on the AI infrastructure thesis just sold the thesis and kept the lab. That’s not a stock tip — it’s a structural opinion about where durable value sits in the AI economy, and it rhymes with what careful builders already know: bet on the tools you’d still use if the hype cooled, keep your leverage low, and never confuse a theme with a business.

Building your AI stack and want the boring, durable order to do it in? Start at /start-here/ — it routes you to the workflows worth building first.