The AI Fund That Fell 67% in One Month — and Was Right All Along
In June his fund was up 439% and managing $45 billion. Six trading days later, three banks forced him to sell everything. The strangest part is that his big idea about AI may still turn out to be correct.
A 25-year-old and six trading days
Leopold Aschenbrenner was born in Germany around 2001. He entered Columbia University at 15 and graduated as top student of his year at 19. At 21 he joined OpenAI, on the team responsible for keeping powerful AI systems under human control.
In April 2024, OpenAI fired him. The company said he leaked internal information. He said he was pushed out after warning about weak security at the lab.
Two months later he published a 165-page essay called Situational Awareness: The Decade Ahead. Its central claim made him famous: the race to build powerful AI is not really a software race. It is a construction race. Chips, memory, electricity, data centres.
Then he did something most essay writers never do. He turned the essay into a fund and bet real money on it.
What is a hedge fund? It is a private investment fund for wealthy individuals and big institutions. Unlike a unit trust, it is allowed to borrow heavily and to make money when share prices fall. That freedom is the reason returns can be spectacular — and the reason they can vanish.
He started with roughly $225 million. His early backers were serious names: Patrick and John Collison, the brothers who founded the payments company Stripe, along with technology investors Nat Friedman and Daniel Gross, and the trading firm Jane Street.
By the end of June 2026, the fund was reported to be up 439%. At the start of July it managed a reported $45 billion.
By 30 July it had sold everything.
What he actually owned
His portfolio had three moving parts. Understanding all three is the whole story.
Part one — he bought the suppliers, not the stars. Instead of buying the famous AI names, he bought the companies one layer beneath them. Memory chip makers SK Hynix, SanDisk and Micron. Companies renting out computing power, called CoreWeave and Nebius. A power generation company, Bloom Energy. And former bitcoin mining companies converting their sites into AI data centres — IREN, Core Scientific and Applied Digital.
Part two — he bet against software companies. He believed AI would eventually destroy the business model of firms like Adobe, so he positioned to profit if their shares fell.
What does "betting against" mean? It is called short selling. You borrow shares you do not own, sell them today, and hope to buy them back cheaper later. If the price falls, you keep the difference. If the price rises, your loss has no natural limit.
Part three — he bought insurance. He held $8.46 billion worth of protection against a fall in chip shares, including about $2 billion covering a semiconductor fund called SMH and $1.6 billion covering Nvidia.
What is that insurance? They are called put options. You pay a small fee for the right to sell a share at a fixed price. If the share collapses, the option pays out. Think of it as fire insurance on a specific building — which matters later, because the fire started somewhere else.
Then he added the ingredient that decided everything.
Four dollars of shares for every one dollar of his own
The fund ran at roughly four times leverage.
What is leverage? It is investing with borrowed money. At four times leverage, for every $1 of investors' money the fund held about $4 of shares. Gains are multiplied by four. So are losses. A 25% fall in the shares wipes out 100% of your own money.
The money was borrowed from three banks: Goldman Sachs, JPMorgan and Bank of America. In the industry these lenders are called prime brokers. They lend the money and hold the shares as security — and they have the right to demand their money back at any moment.
On top of the borrowing, the portfolio was extremely concentrated. His five largest declared holdings were more than 76% of his entire declared buy list.
| Company | What it does | Share of his buy list | How much of the company he owned |
|---|---|---|---|
| Bloom Energy | Power generation | 22.8% | 2.3% |
| SanDisk | Memory chips | 18.8% | 0.8% |
| CoreWeave | Rents out computing power | 14.4% | 1.7% |
| IREN | Ex-bitcoin miner, now data centres | 10.4% | 3.5% |
| Core Scientific | Ex-bitcoin miner, now data centres | 10.1% | 8.2% |
Look hard at that last column. He owned 8.2% of every share Core Scientific has ever issued. A stake that large cannot be sold quietly. The moment you try to leave, you are the reason the price is falling.
The six days that ended it
His whole thesis rested on one belief: the world does not have enough computing power, so anyone who supplies it will make a fortune.
On 1 July, Meta announced it had built too much.
Meta revealed a new business called Meta Compute, selling its spare AI computing capacity to other companies. The message the market heard was simple and brutal: if Meta has extra, the shortage is over.
| Date | What happened | Effect |
|---|---|---|
| 1 July | Meta announces it will sell spare computing power | Micron −10.6%, CoreWeave −14%, Nebius −17%, SK Hynix −9%. About $200bn of value gone in a day |
| 17 July | Reports of a Meta–Anthropic computing deal | The selling deepens across the sector |
| 22–29 July | The whole chip sector re-prices | Philadelphia Semiconductor Index −28.6% from its 22 June peak; Korea's Kospi −33% |
| 24 July | He writes to investors calling it "one of the best buying opportunities in over a year" and invites new money from 1 August | He is buying, not selling |
| 24–29 July | His core holdings collapse | Individual positions fall between 27% and 54% |
| 29 July | All three banks demand their money back | Margin calls |
| 30 July, before the market opened | The entire portfolio — everything he owned and everything he had bet against — is sold to Citadel in a single block | Assets fall from a reported $45bn to about $10bn |
What is a margin call? When the shares you bought with borrowed money fall far enough, the bank demands more cash immediately. If you cannot pay, the bank sells your shares for you. This is the single most important idea in this article: at that moment, you no longer choose when you sell.
Why he lost on both sides at the same time
Here is what should not have been possible. He owned AI hardware and he had bet against software. Those two positions were supposed to protect each other. If AI was winning, hardware rises. If AI was disappointing, software recovers.
In late July, both moved against him. His hardware shares fell, and the software shares he had bet against went up.
His insurance failed too. He had bought protection on Nvidia and on a broad chip fund. But the shares that fell hardest were memory makers, computing-rental firms and power companies. The insurance was real. It was simply written on the wrong building.
This is the trap. A portfolio can look diversified — different companies, different countries, different industries — and still be a single bet. Everything he owned depended on one sentence being true: the world needs more computing power than it has. When that sentence came into doubt, there was nowhere inside the portfolio to hide.
The cruellest detail in the whole story
Citadel bought his positions before the market opened on 30 July, at a discount.
By 3pm that same afternoon, those shares had jumped.
| Share | Move on the afternoon of 30 July |
|---|---|
| Nebius | +27.1% |
| IREN | +26.5% |
| Bloom Energy | +25.6% |
The entire group rose together. Nothing had changed about any of these businesses in six hours. What changed is that the forced seller was gone.
He was proved right about the bounce within a single afternoon. He just no longer owned any of it.
The number that hides the damage
Now for something that sounds impossible. After falling 67% in July, the fund was still up about 80% for 2026.
Both numbers are true. It was up 439% by June, so even a two-thirds fall left it ahead for the year.
But almost nobody actually earned that 80%.
The fund's return and the investor's return are not the same thing. The fund grew from about $225 million to a reported $45 billion. That means the overwhelming majority of the money arrived after the big gains — attracted by them. A small group of early backers made a fortune. The large crowd that arrived in May and June lived through the 67% fall and none of the rise.
This is not unique to hedge funds. It is the most common way ordinary investors lose money in good funds. The advertised performance belongs to whoever was there first. Money almost always arrives after the returns, which is another way of saying it arrives near the top.
What we can actually learn
1. Being right about the story is not the same as making money from it. His argument may still prove correct. AI may well consume enormous amounts of chips, memory and electricity for the next decade. He was destroyed anyway. The idea was never the problem. The construction of the bet was.
2. Borrowed money takes away your right to be patient. If you own shares with your own cash and they fall 30%, it hurts — but you can wait five years if you want to. With four times borrowed money, a 30% fall does not hurt you, it ends you. The bank sells at the worst possible moment, and it is not your decision. Patience is a luxury that leverage takes away.
3. "Hedged" and "safe" are different words. He had insurance worth $8.46 billion and a second position designed to profit when the first one struggled. Neither worked, because both were tied to the same story. Real protection has to be uncorrelated — it has to be something that does not care about your main idea at all.
4. Size your position for the exit, not for your conviction. Owning 8.2% of a company feels like conviction on the way up. On the way down, it means there is no door. Ask how quickly you could sell before you ask how much you believe.
5. Two years of returns describes the weather, not the sailor. +439% in a period when AI infrastructure shares only went up tells you what the market did. It does not yet tell you whether the manager can handle the market doing something else. A track record needs at least one bad regime in it before it means anything.
6. Adding to a falling position works with your own money and kills you with the bank's. On 24 July he called it a great buying opportunity. He was not wrong about the value. He was wrong about whether he would still be solvent long enough to collect it. Six days later he was liquidated.
7. When everyone knows your positions, they become a weapon. Large US funds must publish their holdings every quarter. His were famous and widely copied. Once traders understood he was in trouble and would have to sell, they sold first. In his own letter he described it as a bank run — "vulnerability begetting more vulnerability".
What he got right, and it matters
It would be easy to write this man off. That would be the wrong lesson.
When the end came, he sold the whole portfolio in one single transaction instead of dribbling it into the market over days. Selling piece by piece would have crushed the entire AI supply chain and taken other investors down with him. It was the responsible way to fail.
He removed all borrowed money from what remained. He protected the private holdings, including a stake in the AI company Anthropic reported at around $5 billion, which may list on the stock market as soon as October.
And he wrote it down plainly. His letter to investors opened with "We let you down this month" and included the sentence "I take full responsibility for these events." He admitted the fund had "worked to keep the portfolio within our risk parameters, but gradually this became more difficult."
He ended with: "we took the steps that were necessary to fight another day." He is 25 years old. He is not finished.
And no, it was not the Fed
A Bloomberg headline circulated widely in those days saying Citadel Securities expected the US central bank to surprise everyone with an interest rate increase on 29 July. Many people have since linked the two events.
It did not happen. The Federal Reserve held rates at 3.50%–3.75% on a 9–3 vote, with three officials dissenting because they wanted a rise. Markets had priced roughly a 1-in-3 chance of that increase.
So the uncertainty was real and it added to the nervousness that week. But the rate rise never came. The fund was destroyed by falling AI shares multiplied by borrowed money — not by the central bank. Worth knowing before you repeat the story.
What this means if you are investing from Malaysia
Start with the good news. You almost certainly do not invest with four times borrowed money. That single fact is the biggest protection you have, and it is the reason this exact disaster cannot happen to a normal unit trust investor.
But the other mistakes are available to everyone, in smaller sizes.
Concentration does not require a hedge fund. Putting most of your money into one technology fund is the same error he made, just without the borrowing. If the story behind that fund weakens, everything inside it falls together — because everything inside it was bought for the same reason.
A thematic fund is a story purchase. When you buy a technology or AI fund, you are not buying 60 different companies. You are buying one idea, expressed 60 times. That is fine — as long as you know that is what you own, and you size it accordingly.
The chase is the real danger. Money flooded into his fund because it was up 439%. If you find yourself moving into a fund mainly because of what it did last year, pause. That is the same instinct that put most of his investors on the wrong side of this.
None of this means avoid AI. The buildout may well be one of the defining investment stories of the decade. It means size the position so that a 30% fall is uncomfortable rather than fatal — so that when the bad month arrives, you are the one who gets to wait, and never the one who is forced to sell.
Bottom Line
Leopold Aschenbrenner may still be proved right about AI, and it did not save him. He lost 67% in a month because he used four times borrowed money, put 76% of his portfolio into five shares, and held a hedge tied to the very story he was betting on. The shares he was forced to sell rose more than 25% the same afternoon. Survive first. Being right is only worth something if you are still holding when it happens.
What happened to the Situational Awareness hedge fund?
In July 2026 the AI-focused fund run by Leopold Aschenbrenner fell 67% in a single month. After AI infrastructure shares dropped sharply, its three lenders demanded their money back and the fund sold its entire public share portfolio to Citadel in one block trade on 30 July 2026. Assets fell from a reported $45 billion to around $10 billion.
Was Leopold Aschenbrenner wrong about AI?
Not necessarily. His argument was that AI needs an enormous buildout of chips, memory, power and data centres, and that may still prove correct. He lost money because of how the bet was built — roughly four times borrowed money, five shares making up 76% of the portfolio, and insurance written on the wrong companies — not because the underlying idea was disproved.
How can a fund be down 67% in a month and still be up for the year?
It was up 439% in the first half of 2026, so even after a 67% fall it remained roughly 80% higher for the year. But most of its money arrived near the peak, attracted by those returns. The published fund return and the return actually experienced by the average investor are two very different numbers.
What can Malaysian investors learn from this collapse?
Three things. Borrowed money removes your ability to wait out a fall. A portfolio built on a single story falls all at once, no matter how many companies it holds. And position size should be set so that a 30% drop is survivable — so you are never the one forced to sell at the bottom.
Sources: The Wall Street Journal; Reuters; CNBC; Bloomberg; TechCrunch; the fund's investor letters of 24 and 30 July 2026; US regulatory holdings filings for the first quarter of 2026; Federal Reserve statement of 29 July 2026. Reported asset figures differ between outlets — Reuters and CNBC cite $45 billion at the start of July, while the Wall Street Journal described the fund as managing "well over $20 billion", a gap likely explained by whether borrowed money is counted. Figures are as reported at 3 August 2026.
This article is for educational and informational purposes only and does not constitute financial advice. Past performance is not indicative of future results. All investments carry risk, including the possible loss of principal. Consult a licensed financial adviser before making investment decisions. Adezeno is a licensed unit trust consultant with Eastspring Investments Berhad.