Citadel’s intervention in the wreckage of Situational Awareness was meant to look like a rescue trade. Its reported rapid retreat from much of the book instead makes the episode a warning about leverage, momentum and the fragility of the AI-infrastructure boom.

Leopold Aschenbrenner, the former OpenAI researcher who founded the fund in 2024, built his wager around the vast computing, chip, memory and energy build-out required for advanced AI. The thesis initially delivered spectacularly: Situational Awareness returned 439% through June, according to the Financial Times, while its assets reportedly climbed as high as $45 billion before the reversal.¹

Then the trade soured. Public investors began questioning whether enormous AI capital spending would produce revenue quickly enough, pulling down infrastructure names including SK Hynix, Sandisk, Bloom Energy and Nebius. The fund’s use of borrowed money amplified the losses. Even as markets fell, Aschenbrenner called the sell-off one of the best buying opportunities since early last year and sought new commitments from investors; Bloomberg reported that appeal failed to draw the hoped-for capital.¹

By the following day, the unwind appeared to have accelerated. Axios reported that Citadel had bought all of Situational Awareness’ public-equity portfolio, citing a source, while other reporting described the deal as a sale of the majority of the book.² The difference matters less than the direction of travel: a fund that had recently managed about $20 billion was shedding its listed holdings after a sharp AI-stock rout.

Citadel’s purchase suggested Ken Griffin saw value in assets dumped under pressure; the firm already held some similar AI-infrastructure bets. But the reported next move was far less patient. Citadel offloaded roughly 80% of the portfolio it had acquired, according to a subsequent report.³

Situational Awareness has not, however, surrendered every long-term bet. Its private holdings—including a reported Anthropic stake valued around $5 billion—remain outside the public-market sale.¹ That leaves Aschenbrenner’s central AI conviction intact, but with far less room for a public-market stumble.