The AI trade’s first big blowup: what checks out, and what doesn’t
A $45bn AI hedge fund sold its book to Citadel after a margin call. We checked the claims circulating about it, most hold, one is inflated.
Situational Awareness, the AI-thesis hedge fund founded by former OpenAI researcher Leopold Aschenbrenner, sold its public equity book to Citadel this week after prime brokers called margin. Assets that peaked around $45 billion are now roughly $10 billion, and Bloomberg and others put the July drawdown at 67%. The mechanism was leverage: roughly 4x, paired with a short against traditional software that reversed at the same time the long side fell, squeezed from both ends at once.
We checked the claims circulating about it. Most stand up. One does not, and it is the one people are repeating hardest.
The fact check
Confirmed: the collapse and the scale
Multiple outlets including Bloomberg report the Citadel sale, the fall to about $10 billion, the 67% July decline, and a 439% return before it. The fund reportedly kept a private Anthropic stake.
Overstated: “the Kospi crashed nearly 40%”
South Korea’s index finished July down about 21%, not 40%, and that includes a 17.9% single-day rebound on the final Friday, meaning it was down roughly a third at the worst point. Over twelve months it is still up more than 100%. A brutal month, not a 40% crash.
Confirmed exactly: Meta’s cash flow
Free cash flow fell to $784 million from $8.55 billion a year earlier, a 91% collapse, as claimed, with quarterly capital expenditure of $31.1 billion and 2026 guidance of $130–145 billion. Meta gave no 2027 capex figure.
Confirmed: the Fed held
On 29 July the FOMC kept the target range at 3.5–3.75%. Claims about the 10-year yield, mortgage rates and prediction-market odds we could not verify in time and have left out.
Unverified: everything about the model layer
The specific claims about DeepSeek’s new model beating benchmarks and OpenAI cutting prices by 80% we could not confirm against primary sources. DeepSeek does publish genuinely low API prices, but treat the head-to-head comparisons as unchecked.
The useful part is the mechanism, not the drama. This was not a fund being wrong about AI (the thesis may yet prove right); it was a fund unable to survive being early while carrying 4x leverage. A 25% drawdown at that leverage does not dent your position, it ends it. That is a lesson about financing, not about whether the technology works.
And if you are not trading any of this, the read-through is narrow: capital-market stress changes who owns the infrastructure, not whether the tools keep running. Watch pricing rather than headlines. The scenario that would actually reach you is subsidised inference getting more expensive, not your assistant disappearing.
Questions people ask
- What happened to the Situational Awareness hedge fund?
- After heavy July losses in AI-related stocks, prime brokers issued margin calls and the fund sold its public equity portfolio to Citadel. Reported assets fell from a peak near $45 billion to roughly $10 billion, with a July decline of about 67% following a 439% run.
- Did the Kospi really crash 40%?
- No. South Korea’s index ended July down roughly 21%, after a 17.9% rebound on the month’s final trading day, so the trough was near a third, not 40%. Measured over the past year the index is still up more than 100%.
- Does this mean the AI bubble is popping?
- It shows leveraged positions being unwound, which is not the same thing. Meta’s free cash flow did collapse 91% under capital spending, but usage and cloud revenue keep growing. Stress is being priced in public markets; no major AI infrastructure borrower has actually defaulted.
Verified directly: Meta’s Q2 2026 earnings release (free cash flow, capex, guidance); the FOMC statement of 29 July 2026; Trading Economics data for the Kospi as of 31 July 2026. Fund details are from press reporting including Bloomberg, which we could read only as headlines and summaries rather than full articles. Claims we could not trace to a primary source are labelled unverified above rather than repeated as fact.
Trying to work out which AI tools fit your task? Describe the outcome and get a Blueprint: the tools, the prompt, and the steps, with pricing we verified ourselves.