Leverage does not change what you own. It changes how long you are allowed to be wrong. Situational Awareness, the AI-focused fund founded in 2024 by former OpenAI researcher Leopold Aschenbrenner, found the edge of that window last week.
The fund peaked at roughly $45 billion at the start of July, after a reported 439% return in the first half of 2026. By Thursday its holdings were down to about $10 billion. In between, according to reporting from CNBC and Bloomberg, prime brokers issued margin calls the fund could not meet, and the entire public equity book — valued near $16 billion before the sale — went to Ken Griffin’s Citadel.
The structure that produced the 439% is the same structure that produced the unwind. The book was long AI infrastructure, most visibly SK Hynix, and short software incumbents including Adobe, financed at roughly four times leverage. In July the longs fell hard — core positions dropped by reported margins of 35% to 47% — while the software shorts rallied. At 4x, a 25% decline in the collateral is the entire equity stake. Both legs moving against a levered book at once is not bad luck compounding. It is one macro bet showing up twice on the same balance sheet.
Our take: There is no fraud here and no Archegos-style hole. The counterparties got paid, Citadel got a portfolio at a price it liked, and the fund is still operating. What died is a narrower idea: that conviction about a technology transfers cleanly into leverage on the stocks attached to it. Being directionally right about AI over five years does not survive a 4x margin schedule over five weeks. The fund’s remaining stake in Anthropic, reported at about $5 billion, is the position that worked — and the reason it worked is that nobody can margin-call a private mark.
The part that matters if you never heard of the fund
Forced selling of that size does not stay inside one fund. A concentrated levered holder liquidating into an already-weak tape is a mechanical problem, and it lands in the same names index and retail investors hold. It cuts the other way too: some of the software rally that squeezed the short leg was a large short being covered, not businesses getting better. Anyone reading July’s AI-versus-software rotation as a verdict on fundamentals should discount it by however much of the move was one book being taken apart in public.
What to watch
- Whether other levered AI books face the same collateral math. Situational Awareness was the largest and most visible, which is not the same as the only one.
- The Anthropic stake. A private position that size is now a large share of what is left. The next funding round prices it.
- Prime-broker risk limits on concentrated single-theme books — the tightening usually arrives one blowup late.
- Whether the software rally holds now that the covering pressure is gone. That is the cleanest available test of how much of July was flow rather than fundamentals.
The since-inception number, reported above 1,000% after fees, is still one of the best in the business. It now gets printed next to a $35 billion decline in assets over four weeks. Both of those facts came from the same trade.