A 25-year-old's AI fund went from $45 billion to a Citadel fire sale in July

Leopold Aschenbrenner is 25, a former OpenAI researcher, and in 2024 he published a long essay called Situational Awareness about how fast AI systems would get. He then started a hedge fund with the same name, with early backing from Patrick and John Collison, Nat Friedman, and Daniel Gross. By the start of July the fund was being described as a $45 billion book, after returns that later write-ups put at 439% through June.

The bet itself was straightforward. AI would need chips, memory, and data centers, so the fund bought names such as SK Hynix, CoreWeave, SanDisk, Micron, and Nebius, and it did that with borrowed money. Some positions were levered as high as four times, which means a drop of about 25% can wipe out the equity sitting under the loan. The fund also bet against software companies such as Adobe, and those shorts moved the wrong way in the same month.

In late July the AI-infrastructure stocks fell hard, in some cases 35% to 47%. The prime brokers (Bank of America, Goldman Sachs, and JPMorgan) asked for more cash. That is a margin call: the lender says the collateral is no longer enough, so you send money today or they sell the position. On 30 July, Situational Awareness sold its public stock book in one block to Ken Griffin's Citadel, reportedly at about a 10% discount. After the sale the fund was described as closer to $10 billion. Aschenbrenner wrote to investors that July was down about 67%, and that the year was still up about 80%. He said he would run the public book with cash only, no leverage and no shorts, and he kept the private Anthropic stake.

I keep coming back to the loan, not the essay. Plenty of people still think the AI buildout is real. What broke was buying that view with four times borrowed money, in a concentrated book, in a month when the same names all fell together.

Citadel bought the book and then had a very good July on those same stocks, which is the usual ending when someone is forced to sell. YouTube explainers have already turned it into a fable, and the comment that stuck was a viewer pointing out the fund did not go bankrupt: it was worked out by an acquirer, the way Bear Stearns was in 2008. Maybe that is the cleaner lesson if you are watching AI stocks from India or anywhere else. The thesis can still be right, and you can still get sold out of it.

Aschenbrenner said the fund is not shutting down. Prediction markets still treat a wind-down as unlikely (about 5.5% on Polymarket) and a confirmed Anthropic sale as even less likely, around 3%. One account on X said he later put about $400 million back into the same kind of bet.

The next public tell is a 13-F filing, months from now, and whether he actually stays off leverage. July already taught the expensive version of that lesson.