The One Bet That Could Not Be Margin Called

One block trade, before anyone else could react

Before the market opened on 30 July, Situational Awareness sold its entire public equities book in a single block. Ken Griffin's Citadel was the buyer, having outbid Millennium and Jane Street for the package, which Business Insider reported was heavy in AI hardware names including Nebius and Sandisk. Bloomberg put the firm's assets at roughly 10 billion dollars afterwards, down from more than 20 billion, and CNBC reported the fund had reached as much as 45 billion at its high point.

Leopold Aschenbrenner, who founded the firm after leaving OpenAI and published the manifesto the fund is named after, had been running one of the most concentrated expressions of the AI infrastructure thesis in public markets. The trade was not closed by a change of mind. It was closed by prime brokers.

What actually broke was not the thesis

The positions fell hard. Nebius, where the firm disclosed a multi-billion-dollar stake in May, dropped about 48 percent from its peak. Sandisk lost 56 percent in barely a month. SK Hynix contributed further losses, and a bet against software stocks went the wrong way at the same time. When collateral value falls below the required minimum, prime brokers issue margin calls demanding cash or securities, and the borrower either finds them or sells.

The distinction matters more than the drama. Nothing in the last month proved that AI infrastructure is a bad long-term holding, and nothing about this unwind settles that question either way. What the month proved is that the financing could not survive a drawdown of that depth. Those are two separate propositions, and the second one is the one that closed the fund's public book.

The asset that survived was the one that could not be sold

The firm kept its private portfolio, and the centre of it is a stake in Anthropic reported at around 5 billion dollars. Anthropic was last valued at 965 billion dollars in a Series H round in May. A spokesman rejected reports that the firm was marketing that stake, and the plan described is to carry on as a private investment vehicle.

Read that structure carefully, because it inverts the usual assumption. A private holding is normally treated as the risky part of a portfolio: no daily price, no exit on demand, no way out in a hurry. In this episode those same properties were the reason it survived. Nobody can issue a margin call against an asset that has no daily mark and was never posted as collateral. The liquid positions, the ones considered safe because they could always be sold, were sold, and precisely because they could be.

Leverage sets the clock, not the analysis

Business Insider reported the fund had returned more than 400 percent in the year before July, and reports put the leverage at roughly four times. Both numbers describe the same machine. A four-times position turns a 25 percent decline in the underlying into a wipeout of the equity behind it, which is why a 48 percent fall in one holding and a 56 percent fall in another was sufficient, without any need for the thesis to be wrong.

This is the part that transfers out of finance and into operating decisions. Borrowing does not make a plan more likely to be right; it shortens the amount of time the plan is allowed to take. An owner who funds a three-year technology programme with facilities that can be recalled in ninety days has not made a three-year commitment, whatever the board minutes say. The maturity of the money is the real deadline.

Where this lands for anyone financing AI capacity

European firms are signing GPU leases, datacentre commitments and asset-backed facilities against hardware whose resale value moves fast, and the same structure applies at every size. Three questions do most of the work. How far can the collateral fall before someone else decides your timing. What happens to the covenant if the asset is re-marked rather than sold. And which parts of the plan you would still be able to finish on the day the facility is withdrawn.

The practical discipline is to size borrowing against the drawdown you can survive rather than the return you expect, and to keep the piece of the strategy that matters most funded from money that cannot be recalled. Situational Awareness ended up doing exactly that, but by accident rather than by design: the part of the bet it kept is the part it had no mechanism to lose. Deciding which of your commitments belongs in that category is cheaper to do now than during the call.

Read next: Even a Trillion-Dollar Lab Must Rent Its Compute  |  Microsoft Trained Its Sellers to Call Claude Slower