How Leopold Aschenbrenner’s Hedge Fund Lost 67% in One Month

Leopold Aschenbrenner spent 165 pages explaining how the AI decade would unfold, built a hedge fund around that prediction, returned 439% in six months, then lost 67% in July and sold most of his public-stock portfolio to Citadel.

If somebody wrote that as a Wall Street film, you would tell him to calm down with the plot.

The internet has since made the story even better. Citadel Securities publicly predicted a surprise Federal Reserve rate hike. AI stocks kept falling. Aschenbrenner’s leveraged fund came under pressure and sold most of its public book to Ken Griffin’s Citadel. The Fed did not raise rates, and several of the same AI stocks rebounded hard once the forced selling appeared to be finished.

It looks outrageous, and I understand why people are asking questions. But “this looks suspicious” and “this was market manipulation” are not the same statement. The first is fair. There is currently no public evidence proving the second.

The verified story is already one of the wildest hedge-fund collapses in years. It is also more useful than the conspiracy version, because Aschenbrenner did not blow up by being obviously stupid about AI. His bigger mistake was building a position that could not survive a violent month.

Quick Answer

Leopold Aschenbrenner’s hedge fund, Situational Awareness, reported a 439% net return for the first half of 2026 and then lost 67% in July as leveraged AI-related positions moved against it. Its long AI and infrastructure bets fell, its short side also reversed, liquidity weakened and the fund came under pressure to raise capital or reduce its portfolio.

Reuters reported that Situational Awareness sold most of a roughly $16 billion public-equity portfolio to Citadel. Citadel acquired the portion financed with leverage, while Situational Awareness retained a book of roughly $10 billion containing public shares and private investments including Anthropic. The fund did not shut down, and Citadel did not buy the entire company.

There is no disclosed evidence that Citadel manipulated the market to force the sale. Citadel Securities did make a public call for a surprise Fed hike shortly before the deal, and the Fed held rates instead. That sequence deserves scrutiny, but a wrong public forecast is not proof that it was knowingly false, coordinated with Citadel’s hedge fund or used to manufacture the sale.

Who Is Leopold Aschenbrenner?

Aschenbrenner was an OpenAI researcher on the company’s Superalignment team. OpenAI fired him in April 2024 over an alleged information leak. He disputes the company’s version and has said the document was a benign safety-planning paper shared with three outside researchers for feedback. OpenAI told Business Insider that the internal security concerns he had raised did not cause his dismissal. That disagreement has never been cleanly resolved in public.

Two months later, he published Situational Awareness: The Decade Ahead, a 165-page essay arguing that continued increases in computing power could produce AGI around 2027, followed by an even faster intelligence explosion. The dramatic parts attracted most of the attention, but the investable argument underneath was easy to understand: if AI capability keeps improving, the world will need absurd quantities of chips, memory, data centres, electricity and capital.

He then did what people constantly tell forecasters to do and put money behind the prediction. Situational Awareness launched in late 2024 with roughly $225 million, backed by investors including Stripe founders Patrick and John Collison, Nat Friedman, Daniel Gross and Jane Street. Its first public SEC Form 13F showed about $255 million in reportable US securities at the end of 2024. By the first quarter of 2026, the reported 13F value had reached $13.68 billion.

Those figures do not equal the fund’s net assets. A 13F is a delayed snapshot of certain US-listed holdings. It excludes private investments, cash and conventional short positions, while options can make the reported value look very different from the cash committed. Still, the filings show how violently the public book expanded.

How Did Situational Awareness Return 439%?

The fund took a concentrated version of the AI build-out thesis and added borrowed money. Its publicly associated positions included companies such as CoreWeave, Bloom Energy, Sandisk, Nebius, SK Hynix, Broadcom and Intel. These were not random technology stocks. They sat around the physical AI trade: memory, cloud computing, data centres, power and the equipment needed to keep the whole machine running.

That thesis had real substance. AI companies and hyperscalers were spending at a scale the market had never seen, although some of that growth depended on the same suppliers, customers and investors financing one another. We covered that uncomfortable part in our guide to AI circular financing.

Situational Awareness also ran a short book against companies it expected AI to weaken. The simple description was “long AI infrastructure, short software.” When infrastructure shares rose and the supposed AI losers fell, both sides paid.

Then came leverage. CNBC reported that leverage on the public book had reached as much as four times. Reuters independently confirmed that the fund used borrowed money and that Citadel ultimately acquired the leveraged part of the portfolio, although Reuters did not publish a precise leverage multiple.

For context, global hedge funds averaged about 7% in the first half of 2026. Situational Awareness made 439% after fees.

Nobody produces that gap by discovering one slightly undervalued company and waiting patiently. The AI call was excellent, the concentration was enormous and leverage turned every correct move into something ridiculous. The same mechanism that made the returns look impossible then started working backwards.

Why Did the Trade Collapse From Both Sides?

A long-short fund is supposed to have protection. It owns companies expected to rise and shorts companies expected to fall, so one side should soften the damage when the market moves. That protection becomes less useful when both sides depend on one crowded story.

Situational Awareness was long the physical winners of the AI boom and short parts of the software market expected to lose from AI. In July, investors started questioning AI valuations, capital spending and the debt supporting the build-out. Infrastructure shares fell sharply. At the same time, parts of the software trade rose, hurting the shorts.

In his investor letter, reproduced by Business Insider, Aschenbrenner said many AI names had fallen by half or more while the fund’s positive long-short spread “reversed violently.” He also described increasingly adverse trading in stocks publicly associated with Situational Awareness, comparing the process with a bank run.

That comparison makes sense mechanically. Once other traders know a leveraged fund may have to sell, they do not wait politely for it to recover. They sell or short the same holdings, liquidity disappears and lenders demand more collateral. The weaker the fund becomes, the easier it is to trade against.

This can feel like manipulation to the person trapped inside it. Sometimes predatory trading can cross legal lines. But other investors recognising a forced seller and positioning around it is not automatically illegal. A highly visible, leveraged and concentrated book tells the entire market where the weak point is.

Aschenbrenner’s deeper problem was that leverage removed his right to wait. He could still believe the companies would be worth more in 2028. His lenders cared about the collateral in July 2026.

How Can a Fund Lose 67% and Still Be Up 80%?

The numbers sound contradictory until you apply the losses in the correct order. This is a calculated example using the reported returns, not the fund’s real starting capital:

A 67% fall from $539 leaves about $178, which is still roughly 78% above the original $100. Reporting rounded the remaining annual gain to about 80%.

This is why “he lost everything” is wrong. July destroyed most of the year’s extraordinary gain and came dangerously close to permanent damage, but investors who had been in the fund since the start of 2026 were still ahead on paper after the collapse.

The result also shows how ugly percentages become after a drawdown. A 67% loss requires a gain of slightly more than 203% just to recover. Losing two-thirds and then making two-thirds does not take you back to even.

Did Citadel Buy the Fund for Pennies on the Dollar?

No public reporting supports that exact claim. Citadel did not acquire Situational Awareness itself. It bought most of the public-stock portfolio in a large block transaction arranged with prime brokers including Goldman Sachs, JPMorgan, Bank of America and Citigroup. Reuters described the public book as roughly $16 billion and said Citadel took the portion financed through leverage.

Situational Awareness retained a roughly $10 billion book containing fully paid public positions and private investments, including its Anthropic stake. In his letter, Aschenbrenner said all shorts were closed, leverage was removed and the fund would continue as a hybrid public-private investor.

The transaction probably involved a discount because Citadel was taking a large distressed portfolio from a seller with limited options. That is how block sales work. But the terms have not been published, so “pennies on the dollar” is internet decoration presented as a number.

The widely repeated $35 billion loss has a similar problem. Some reporting placed the fund or its portfolio at $45 billion near the start of July, while other reports described more than $20 billion in assets and a $16 billion public book immediately before the sale. These numbers can refer to different dates and different things: net assets, gross exposure, public positions, private assets or leveraged notional value.

Subtracting the later $10 billion book from the highest reported $45 billion figure creates a dramatic $35 billion headline. It does not prove that investors handed over $35 billion in cash and lost it.

Did Citadel Manipulate the Market to Take Leopold’s Portfolio?

There is no public evidence proving that as of 4 August 2026. The suspicion comes from a real and extremely uncomfortable sequence:

One distinction matters here. The rate forecast came from Citadel Securities, the market-making and research business. The portfolio buyer was Citadel, the hedge fund. Ken Griffin founded both and is CEO of Citadel while serving as non-executive chair of Citadel Securities, according to Citadel’s own leadership page. They are different corporate businesses, even though the shared founder and branding make the timing look worse.

That corporate separation does not make every concern disappear. It means a manipulation claim would need evidence showing that the forecast was knowingly false or misleading, that the two businesses coordinated around the distressed sale, that somebody traded using material non-public knowledge of the block, or that orders were used to create an artificial price.

FINRA’s manipulation rules prohibit false market activity and front-running imminent customer block trades. So there is a real legal question that could be investigated if evidence appears. Right now, no regulator has publicly accused Citadel or Citadel Securities of misconduct in this transaction.

Calling it the biggest market manipulation we have seen is especially premature when Archegos exists. Archegos was a $36 billion collapse whose founder, Bill Hwang, was convicted of fraud and market manipulation. In the Situational Awareness case, there is a suspicious-looking timeline and a lot of inference. Those are not equal categories of proof.

Why Did the AI Stocks Jump After the Sale?

The market no longer had to guess how many billions a desperate seller still needed to dump. When a leveraged fund is unwinding, price stops being only a verdict on the underlying companies. The fund sells because it needs cash, other traders sell because they expect more forced selling, and buyers step back because waiting may produce a lower price tomorrow. Once Citadel absorbed the block, that pressure disappeared in one transaction.

Some shares rebounded violently. Sandisk rose 23% on 30 July, its best day since January, but Barron’s noted that it was still down 45% for July. That rally also followed strong results and a bullish memory-supply outlook from Samsung, which mattered directly to the semiconductor trade.

The wider market had another enormous catalyst. Microsoft jumped 15.5% after strong cloud results and a forecast for lower capital spending, adding roughly $450 billion in market value. Reuters reported that the Nasdaq rose 2.8% and the S&P 500 gained 1.7% that day.

So yes, Citadel bought distressed assets and watched some of them surge. That is exactly why Citadel wanted the trade. The rebound does not prove Citadel caused the preceding collapse. It shows how much forced positioning had distorted prices, followed by company news good enough to bring buyers back.

The cruel part is that Aschenbrenner may have been forced out close to the point where the pressure ended. Markets do this constantly. They can validate your thesis immediately after your financing structure removes you from it.

Was Leopold Aschenbrenner Wrong About AI?

July did not settle that question. His AI thesis can be broadly correct while his fund construction was reckless. The world may spend trillions on chips, power and data centres. Anthropic may become vastly more valuable. AI infrastructure companies may recover and make Citadel a fortune. None of that would retroactively make four-times leverage sensible.

An investor needs two things to be right: the asset and the path. If the path includes a 50% fall before the value appears, a cash investor can wait. A leveraged investor may receive a margin call and become somebody else’s entry price.

Situational Awareness has not vanished. The fund retained private investments, removed all leverage and said it would continue trading public equities with fully paid positions. Its 80% gain for 2026 would be exceptional for any normal fund, although nobody investing at the June peak will feel comforted by that statistic.

Aschenbrenner was not “running Wall Street.” He had one of the most extraordinary short-term runs Wall Street had seen, and then Wall Street reminded him who controls a position bought with borrowed money.

The Real Investing Lesson Is Bigger Than “Take Profits”

Selling part of a winning stock can be sensible when it has grown far beyond the size you planned. It locks in real money, reduces concentration and stops your entire portfolio from becoming one prediction. A screenshot is not a risk-management system.

But “always sell pieces as a stock rises” is too simple. It can also cut a great long-term winner for no reason. Worse, a fund can take profits and then borrow against the remaining position until the risk is right back where it started.

The useful discipline is to decide the risk before the euphoria arrives:

• Set a maximum position size and rebalance when a winner exceeds it.
• Do not use enough leverage for an ordinary violent drawdown to force a sale.
• Assume several positions built around the same thesis can fall together.
• Keep enough cash and liquid assets to survive a bad month without begging investors for emergency capital.
• Judge success by what remains after the drawdown, not by the highest number the account once displayed.

Aschenbrenner’s own investor letter contains the cleanest rule: a fund must be structured so it can take a loss and fight another day.

He may still be right about the AI decade. In July, he built a portfolio that could not wait for it.