Situational Awareness Didn’t Lose to Its Algorithms. It Lost to a Margin Call.

Three corrections to how this story is being told, before anything else. Situational Awareness did not collapse — it nearly did, and it is still operating. The SEC subpoenas did not go to the fund; they went to the banks that lent to it. And the fund has not been accused of any wrongdoing, a point every serious account of the matter makes explicitly.

What actually happened is older and less exotic than an artificial intelligence failure. A concentrated, heavily leveraged book moved against its manager, the lenders asked for their money, and the positions were sold at whatever price was available. The AI part of this story is what Leopold Aschenbrenner owned. It is not why he lost.

The Trade Was Working, Spectacularly

Aschenbrenner, a twentysomething former OpenAI researcher, launched Situational Awareness in 2024 on a straightforward thesis: the constraint on artificial intelligence would be physical — chips, memory, power, data centers — so buy the infrastructure rather than the models.

For eighteen months it worked better than almost anything on Wall Street. The fund returned 439% through June 2026 and peaked at roughly $45 billion in assets. Its public book included SK hynix, Sandisk, Bloom Energy, and Nebius Group — memory, storage, power, and compute, which is to say four expressions of a single idea.

That concentration is the first half of the mechanism. Four positions built on one thesis are not four positions. They are one position held four ways, and they fall together.

The Second Half Was Leverage

Reported leverage on the fund ran as high as 400%. At that ratio the arithmetic stops being about conviction and starts being about survival: a drawdown that a cash fund absorbs and waits out becomes, for a levered one, a liquidity event with a deadline set by somebody else.

In late July the market turned on AI capital expenditure. The names in the fund’s book fell more than 30% in a month. The portfolio dropped 67% in July alone. Margin calls followed, and on July 30 Situational Awareness sold the majority of its public equity portfolio to Ken Griffin’s Citadel — a forced seller meeting the best-capitalized buyer available, which is never a negotiation. Assets went from about $45 billion at the peak, to roughly $20 billion in the months before the sale, to something near $10 billion after it.

Six days before that sale, on July 24, Aschenbrenner wrote to investors describing the downturn as one of the best buying opportunities since early the previous year. He may well be proved right. It did not matter. That is the entire lesson of a margin call: being correct about the destination is worth nothing if you cannot hold the position long enough to arrive.

The Positions Weren’t Even Wrong

It is worth sitting with how defensible the book looks in hindsight. SK hynix is one of three companies now allocating scarce memory output to AI data centers — the shortage currently pushing DRAM contract prices up by triple-digit percentages and forcing consumer device makers to raise prices. The thesis that memory and power would be the bottleneck was not a hallucination. It was, on the evidence of the last month, correct.

No model malfunctioned. No black box produced an inscrutable decision. A manager sized a directional bet at four-to-one against correlated assets and got a one-month drawdown he could not fund. This failure mode predates machine learning by roughly a century.

What the SEC Is Actually Asking

The New York Times first reported the inquiry, and Reuters followed on August 24. The subpoenas went to Goldman Sachs, JPMorgan, Citigroup, and Bank of America — the banks that supervised the trading and channeled the funding.

The regulator wants the timing of the fund’s trades, the details of its leverage, the timing of the margin calls, and the communications between the fund and its lenders about its borrowing. Read that list carefully: none of it concerns the investment thesis, the algorithms, or whether AI can pick stocks. It concerns whether the people financing a $45 billion levered book knew what they were financing, and whether trades were timed in a way that mattered to who absorbed the losses.

Situational Awareness said it would cooperate to the fullest extent with any regulatory request, and characterized the inquiry as routine oversight of a high-profile fund with a significant drawdown. That characterization is plausible. Requests for information do not imply that any firm is a target, and inquiries at this stage frequently end without enforcement.

What Survived, and Why

The remaining roughly $10 billion sits largely in private positions: a stake in Anthropic valued at about $5 billion according to Bloomberg, plus holdings in the chipmaker MatX and the data center firm Fluidstack, valued at $18 billion in April.

Those survived for an unglamorous reason. They are illiquid, so they could not be posted as collateral or sold into a margin call at three days’ notice. The most conservative-looking part of the portfolio came through untouched precisely because nobody could force it out. The liquid, sensible, publicly traded half is the half that got liquidated.

The reckoning here is not for artificial intelligence in finance. It is the same reckoning that follows every levered concentrated fund that meets a fast market, and the only novel detail is that the collateral happened to be AI infrastructure. An AI thesis does not exempt anyone from the mathematics of borrowed money.