The $35 Billion AI Trading Loss Exposes The Real Risk Of Leverage

The $35 Billion AI Trading Loss Exposes The Real Risk Of Leverage

The $35 billion AI trading loss attached to Situational Awareness will get attention because of the number. According to the Financial Times, it now sits at the top of its table of major historical trading losses. I think the most useful part of the story is how it happened.

Leopold Aschenbrenner’s fund reportedly gained 439% from the start of 2026 through June. Then the portfolio fell 67% in July. Even after that collapse, it was still up around 80% for the year. This was not a manager who spent six months being wrong about artificial intelligence. It was a hugely successful trade that became very difficult to control once the market turned.

What matters to me is what happened next. I have always thought investors spend too much time asking who wants to buy and not enough time asking who must sell. By late July, Situational Awareness was no longer making decisions from the comfortable position of a long-term investor. Margin pressure was now a key factor in the decision-making process. Citadel was on the other side. The AI thesis may still prove right. That almost becomes secondary once the financing structure dictates when positions must be sold.

The $35 Billion AI Loss Started With Success

In 2024, Aschenbrenner, a former OpenAI researcher whose views on the development of artificial intelligence attracted considerable attention, launched Situational Awareness. Those views translated into extraordinary returns during the first half of this year.

By the end of June, that success had also changed the shape of the portfolio. Sandisk and Micron Technology reportedly represented more than half of the fund’s disclosed U.S. equity holdings. By July 30, Sandisk had fallen 43%, and Micron was down 24%.

I do not have a problem with concentration. Some of the best investors I know have made their best returns by putting meaningful capital behind the few situations where they had genuine conviction. But concentration behaves very differently when substantial leverage sits beside it.

Situational Awareness reportedly used leverage of as much as 400%. Once its AI-related holdings began falling quickly, the question changed. It was no longer simply whether Micron, Sandisk, or the broader AI infrastructure trade would be worth more in two years. The fund needed enough liquidity to get through the next few days. That is how a good long-term idea can still become a bad position.

A winning investment tends to get larger without anybody doing very much. The price rises, confidence rises with it, and a position that began as one idea among many starts to dominate the portfolio. Add borrowed money, and your tolerance for volatility shrinks just as your confidence is likely at its highest. Situational Awareness may eventually be proved right about AI. Its problem in July was that the structure gave it less time to wait.

Leverage Took Away The Fund’s Time

Owning a stock outright gives an investor one thing that is often underestimated: time.

The stock can fall 30%, and you can decide whether the thesis has changed. You can hold it, reduce it, or, if you still believe the market is wrong, add to the position. You may still make a terrible decision, but it is still your decision.

Borrow heavily against the position and the calculation changes. A lender does not care that you think the stock will recover in six months. If collateral falls far enough, more capital has to appear now.

Reuters reported that Situational Awareness faced margin pressure during the July collapse and subsequently removed all leverage from the portfolio. Aschenbrenner told investors that the fund had come closer to permanent capital impairment than it considered acceptable.

AI did not suddenly stop existing in July. The long-term spending plans around computing, data centers, and infrastructure did not disappear because some related stocks fell sharply. What changed was the fund’s ability to sit through the volatility.

I have seen the same mechanism, on a much smaller scale, throughout my career in special situations. A SpinCo gets distributed to shareholders who never wanted it. Funds sell because the company is too small, falls outside an index or no longer fits a mandate. They are not necessarily telling you anything about what the company is worth. They are telling you something about what they are allowed, or required, to own. Leverage can create an even more brutal version of that problem. The seller loses the luxury of waiting for value to be recognized.

Citadel Bought What Someone Else Had To Sell

Situational Awareness reportedly sold the bulk of roughly $16 billion of listed equity positions to Citadel as it dealt with margin calls. Reuters later reported that some of those positions were acquired at discounts of more than 10%.

Citadel did not need to have a perfect view on where every AI stock would trade next year. It knew it was dealing with a seller under pressure. One side needed liquidity. The other had it. At that point the buyer has the advantage.

When somebody must sell, the price can temporarily tell you more about the seller’s circumstances than the underlying asset. This is one of the reasons I have spent so much time looking at forced selling in spinoffs and other special situations. Markets normally do a reasonable job of finding prices. They can become much less efficient when one participant is operating under a constraint that the other is not.

The consequences spread beyond Situational Awareness. Reuters reported that Jane Street suffered roughly $15 billion in July hit tied partly to its investment in the fund and other technology positions. The Wall Street Journal reported that Jane Street’s original investment of around $2.5 billion had grown to nearly $10 billion before the selloff.

Jane Street’s exposure had become substantial precisely because the investment had gone so well.

We tend to think that risk lives in the positions that are going wrong. Sometimes the position that needs the most scrutiny is the one that has made the most money. It quietly becomes larger, your confidence has been repeatedly rewarded, and the assumptions behind it start to feel less like assumptions.

What The $35 Billion AI Trading Loss Really Says

I do not see the loss as proof that the AI boom is finished. AI remains a real technological and capital spending cycle. Situational Awareness could ultimately be right about much of what it believes. Some securities Citadel bought during the July liquidation may have been sold at attractive prices. The problem was that a long-term AI thesis ended up depending on short-term financing.

Position size matters because even strong conviction can be wrong. Liquidity matters because markets do not always allow you time to change your mind gracefully. Leverage matters because it can take the decision away from you completely.

Markets have a habit of making risk look smaller after a period of strong returns. A position works, gets larger, and appears safer precisely because it has already made so much money. That is usually the point when I want to know what would happen if it fell 30% rather than what happens if it rises another 30%.

Situational Awareness may eventually be proved right on AI. That does not change what happened in July with their AI trading loss. Once the financing dictated the sale, the investment horizon no longer belonged entirely to the investor. Citadel did not need the better five-year AI forecast at that moment. It needed cash and the ability to wait. I would worry less about whether the next great trade is right and spend more time asking what could force me out of it before I ever get the chance to find out.