How a single July took apart the fund of the man who wrote the defining text of the AI era
Essay · July 30, 2026
On Wednesday evening a few people on Wall Street sat down and started calling the market. They had a list in front of them: which securities would be sold before tomorrow’s open, in what size, and by what hour it all had to be closed. The list was not drawn up by the seller. It was drawn up by the seller’s lenders.
On July 30, 2026, the fund Situational Awareness exited its entire public book. Longs and shorts, one block, one buyer. Not a rebalancing, not a decision to take down risk, not “we thought we would sit in cash for a while”. This is the kind of position closing that is executed by someone other than the person who opened it.
The fund was run by a twenty-four-year-old whose essay had become, two years earlier, the intellectual foundation of Wall Street’s entire bet on artificial intelligence. The bet worked: roughly 439 percent net through the end of June and more than twenty billion dollars under management. The thesis was right about the thing that mattered. It helped not at all.
What follows is not a drama about a young genius. It is a protocol. Because the interesting part here is not the mistake, which arguably was never made. The interesting part is the machinery: it ran exactly as designed, and it took a man’s portfolio away from him at the precise moment he was right.
The man who wrote the trade
Leopold Aschenbrenner. Columbia University, valedictorian at nineteen. From there straight into OpenAI, onto the Superalignment team. In 2024 he was fired over what the company called improper disclosure of internal information. He disputes the characterization and says he showed outside researchers a document that was largely non-confidential.
What he did next is not what anyone expects from a fired twenty-two-year-old. He did not sue and he did not walk across the street to another lab. He sat down and wrote one hundred and sixty-five pages called Situational Awareness.
The thesis is almost embarrassingly simple. Models will learn to do the work of AI researchers by roughly 2027, after which the system begins improving itself. Until then the binding constraint is not algorithms and not ideas. It is silicon, memory and electricity. Progress runs into physics: into chips, into gigawatts, into substations, into transformers ordered three years ahead.
Notice what that document actually was. It looked like futurism. Functionally it was a shopping list. If you believe the bottleneck of the era is matter rather than software, it follows immediately what you should own. And what you should not.
And that is exactly how it worked. The text was read not only on Twitter. It was read in investment committees. Its phrasing migrated into fund decks, into research notes, into the arguments managers used to explain to clients why this capex cycle was construction rather than a bubble. Within two years “it comes down to hardware and power” went from a bold hypothesis to a consensus nobody bothered to defend anymore. That is a rare position for a twenty-two-year-old: not merely calling the direction, but handing the market the language it used to explain its own buying to itself.
How a text became a fund
Work like that normally ends in a lecture circuit, a podcast tour and a guest column. Aschenbrenner did something else. He opened a fund on top of his own text. He did not sell the thesis to investors as research, he stood behind it with money.
The seed capital, around 225 million dollars, came from Nat Friedman, Daniel Gross and the Collison brothers. By the end of 2024 the fund’s disclosed US positions were about 255 million across six names. By the end of 2025, 5.5 billion. By mid-2026 it ran between 20 and 24 billion, with assets touching 45 billion at the peak. Jane Street was among the investors. Through the end of June the fund showed roughly 439 percent net.
Now the detail without which none of the rest makes sense. That result was built with essentially no Nvidia, Microsoft, Amazon, Google or Meta in the long book. He did not buy the storefront, he bought what stands behind it: memory, second-tier data centers, fuel cells, power. The logic came straight out of the text. If the constraint is physical, the money accrues to whoever sells the physical thing, not to whoever sells the interface.
Let me say this plainly and once, so I do not have to come back to it. The trade was smart. It was smart in 2024, it paid in 2025 and in the first half of 2026, and it may well pay again. He was not taken apart for being stupid.
What was in the book
The construction was two-sided and, in its way, elegant. On the long side, the infrastructure of the era: Bloom Energy, SanDisk, CoreWeave, IREN, miners retooled into compute, and Nebius, disclosed on a separate filing and by most accounts the single largest position in the fund. On the short side, applied enterprise software including Adobe, on the thesis that AI eats packaged software faster than its vendors can raise prices.
The first-quarter filing, submitted on May 18, showed 13.68 billion dollars across 42 disclosed positions. About 8.46 billion of notional sat in puts against semiconductors and tech, roughly 62 percent of reported value. The largest: two billion against the sector semiconductor ETF and 1.6 billion against Nvidia. Then Broadcom, Oracle, AMD, ASML, Intel, Micron, TSMC. Note that Micron, a memory manufacturer, which is to say a literal beneficiary of his own thesis, sat on the put side rather than in his long book.
An honest caveat is required here, because without it this turns into a comic strip. A 13F cannot distinguish a directional short from a hedge: strikes and expirations are not disclosed, and ordinary equity shorts never appear at all. We are not looking at the trade, we are looking at its shadow on the wall. But the shape of that shadow is unmistakable. A man who had built an entire fortune on long infrastructure was simultaneously paying serious money for insurance against a fall in precisely that infrastructure. He was not naive. He knew what he owned.
And in July both sides of the construction turned against him at once. The largest longs, Nebius, SanDisk and CoreWeave, each lost more than thirty-five percent in the month. The software he was short went up.
A hedge is when one leg holds the other. Here both legs buckled in the same direction. That is the first real lesson of the story: a structure assembled out of a single thesis does not stop being one bet just because you wrote it down twice with opposite signs.
The short leg had its own separate defect. Shorting enterprise software is a bet not on price but on the calendar. AI genuinely does erode the packaged software model, but it erodes it slowly, quarter by quarter, through renewal negotiations. The loss on a short, meanwhile, arrives instantly. You can be correct on substance and still pay every single day while your correctness works its way into someone’s income statement.
Then the mechanics compound it. When a large player is unwound, he does not only sell the longs. He has to close the shorts, and closing a short means buying. Your own forced exit pushes up the price of the very names you were positioned against. A leveraged book gets hit on both sides simultaneously not by coincidence but by construction.
The earnings held. The credit did not.
The important thing in July did not happen in corporate earnings. Earnings were fine. What broke was the credit underneath them.
If you have never watched credit markets, here is why they matter more than quotes, in one sentence. A share price tells you what investors will pay for a company’s future. A credit spread tells you what they charge for the risk that there is no future at all. The first is a mood. The second is survival, and it moves first.
The cost of insuring against default by the companies building AI infrastructure went vertical. CoreWeave spreads broke 855 basis points, which on a standard model implies something like a fifty percent probability of default over five years. In the same reporting period interest expense doubled to 536 million dollars per quarter against free cash flow of minus 4.71 billion. Spreads on Nvidia debt hit their own highs, and Nvidia is a company whose profits are in perfect order.
The market changed its question. It stopped asking how fast artificial intelligence is growing and started asking who exactly is funding the capital expenditure and at what rate. Those are two different questions, and the second one kills multiples faster than any bad quarter. Apollo’s chief economist put it best: at these yields the capex cycle begins to throttle itself. A buildout financed with expensive debt gets more expensive at exactly the moment everyone has come to believe in it.
The scoreboard. The semiconductor index lost 18.9 percent in July, heading for its worst month since 2008, and some chip ETFs went below minus twenty-seven, the worst reading since September 2001. Astera Labs minus 48, Marvell minus 45, KLA minus 43.
And alongside it, the number this whole section exists for. In that same worst month since the financial crisis, the largest chip ETF took in 6.1 billion dollars, its biggest monthly inflow in records going back to 2018. The crowd was not leaving. The crowd was adding. It always adds: a decline looks like a discount right up until the day it turns out to be a collateral revaluation.
One more line for the file. While the public was putting record money into chip funds, CoreWeave’s cofounder sold roughly 734 million dollars of stock during the second quarter and the chief executive roughly 447 million. That is not a crime and it is not always a signal. But it is a fact, and it sits in the record next to the others.
The Korean rehearsal
The most interesting part of this story did not happen in New York. It happened two weeks earlier in Seoul, and it was a perfect rehearsal, just performed on different people.
KOSPI started the year near 4,500. It broke 5,000 by late January, 6,000 by late February, 7,000 in May, and touched nine thousand and change in June. One hundred percent in six months. SK Hynix nearly quadrupled over the year, Samsung tripled. Two names carried more than half the index and delivered around seventy percent of its gain.
The money going in was retail. Not funds, ordinary people, through margin loans and single-stock leveraged ETFs. Margin lending set a record at almost 38.6 trillion won at the end of June, and more than forty percent of that sat on four related names around Samsung and SK Hynix. The entire market had become one position bought on credit.
On July 13 the index fell 8.95 percent in a session. The sidecar tripped and so did the first level of the circuit breaker. SK Hynix lost 15.37 percent on the day, the largest single-day drop in its listed history. Samsung fell 10.7. By Goldman Sachs estimates, as of that day more than 1.2 million leveraged retail accounts had cumulatively received margin calls and between 320 and 360 thousand were force-closed. The share of forced liquidations, which had sat quietly near two percent for half a year, went above ten. Over two and a half months, 2.3 trillion won of positions were liquidated involuntarily. In late July the finance minister apologized to parliament that single-stock leveraged products had been allowed to exist at all. They were banned after the second crash.
Now put the two chapters together. SK Hynix sat in Aschenbrenner’s long book. Korean retail and one of the most successful AI funds on the planet held the same stock, on leverage, at a broker. Both were unwound by the identical mechanism: collateral revaluation, call, procedure, sale. The difference was two weeks in time and three orders of magnitude in size.
The mechanism does not distinguish by size. That, if you like, is its defining property and the actual news.
The mechanics of repossession
The fund ran leverage as high as four to one. People tend to think leverage is about position size. Leverage is really about time. It does not enlarge your bet, it puts a timer on it, and the person winding the clock is not the person who opened the trade.
At four to one the portfolio stops being yours the moment the collateral value stops clearing somebody else’s model. Not when you change your mind, not when the thesis breaks, but when a number in someone else’s spreadsheet crosses someone else’s threshold. After that everything proceeds by procedure, and procedure has no field for “but he is right”.
The margin calls came from Bank of America, Goldman Sachs and JPMorgan Chase. The same firms, according to reports, were working with the fund to meet those calls and reduce risk “in an orderly fashion”. Polite phrasing, simple meaning: we are selling this either way, the only open question is whether you are in the room.
And then comes the part this essay opened with, the part that does not make it into interviews. Before Thursday’s open, the fund’s own prime brokers were walking baskets of its long and short positions around to potential buyers.
Read that again slowly. Your lender is out in the market with a list of what you are obliged to sell today, showing that list to the people who will buy it from you. The whole market knows in advance what is for sale, in what size, and by what hour it has to be done.
That is the worst negotiating position that exists in finance. You are no longer selling an asset, you are buying an exit. And the price of the exit is set by whoever knows your deadline.
Who actually played
The buyer has not been named publicly. What is known is that it was a large investment firm and that the trade went through essentially as one package: the long book and the short book, whole, into one pair of hands.
But the buyer is not the protagonist here. He arrived on the final day and collected a price that other people had produced. An honorable role, and a passive one: at the right hour he had the money and the phone.
One thing about him is worth saying, because it explains the size of the discount. When a book moves in a single package, the buyer knows for certain that he is the only one. Not the best of three, the only one, because the seller’s alternative is not another buyer, it is several days of liquidating into an order book that is already reading his flow. Removing that alternative is paid for with a discount, and that discount is calculated not off the fundamental value of the securities but off the value of somebody else’s clock.
The people who beat Aschenbrenner were the people financing him. Bank of America, Goldman Sachs and JPMorgan were simultaneously his lenders, his appraisers, his executing brokers, and the only participants in the market who saw his book in full, both sides, in real time, every day. They extended the leverage on the way up. They marked the collateral on the way down. They went out and sold his positions. Four roles under one license, and, characteristically, not a single breach of procedure.
Goldman deserves a separate mention, because it appears in this story twice. It was Goldman’s estimates that described the scale of the Korean retail liquidation: how many accounts got calls, how many were force-closed. And it was Goldman that sat among the brokers taking the fund out of its positions. One firm at the microscope and at the switch.
There is a profession nobody makes films about: the appraiser at a pawnshop. The man behind the glass takes your object, weighs it, names a number. He does not care that you are right about its true value or that at auction in six months it will fetch double, because six months is not his horizon. The difference with Wall Street is exactly one thing, and it does not favor Wall Street. The pawnshop appraiser had not lent you money against your own object, had not marked its value with his own model, and was not out looking for a buyer for it at the very minute he quoted you a price. Here, the same man does all three.
This is where the romantic version of markets ends, the one in which the better thinker wins. The loser is not the person who got the forecast wrong. The loser is the person who borrowed money from a counterparty that understands his risk better than he does.
What survived had no price
The public book was roughly two thirds of the portfolio. The rest is stakes in private companies: Anthropic, Fluidstack, MatX, T1 Energy, Sharon AI. The largest position, Anthropic, is around a fifth of all fund assets, and it was bought early, at a valuation on the order of sixty billion dollars. In May 2026 Anthropic closed a 65 billion dollar round at 965 billion post-money. Which is to say the paper markup on that stake is a number next to which the entire public book looks modest.
That part survived not because it was smarter or luckier. It survived because it cannot be valued quickly, and therefore cannot be seized quickly. A private stake has no quote, no daily collateral remark and no “sell before the open” button.
Hold on to this conclusion, it is counterintuitive and it is worth more than anything else in this text. The liquidity you pay a premium for across your entire investing life works against you on the day of a margin call. What gets taken is not the worst thing you own and not the wrong thing you own. What gets taken is the most liquid thing, the thing with a price right now. Including the positions that were entirely correct.
That armor has a reverse side, and it needs saying, or this turns into a sermon. Having no quote protects you from news as well as from seizure. A private stake is marked to the last round, and the last round closed in May, in a different world, before credit got expensive. If public infrastructure has fallen by a third while a private position still carries the same mark it carried two months ago, that does not mean it held. It means nobody has asked it yet.
The letter of July 24
Six days before the end of the public book, the fund wrote to its investors. The half-year closed at 439 percent net. The selloff is an opportunity. A potential Anthropic IPO will be a catalyst for the whole sector in the second half. And for anyone who had been waiting for a convenient moment to add, the subscription window opens on August 1.
August 1. The day after tomorrow. A window opened one week before the book it was raising against was sold whole, in a single package.
I do not think this is cynicism. I think it is precisely what he believes. A man who at twenty-two wrote one hundred and sixty-five pages arguing that the world was underestimating the scale of what is happening looks at his own positions falling at twenty-four and sees the same thing: the world is underestimating again. That is consistency, not deceit. Consistency simply is not the quality that saves you at four to one.
And he may well turn out to be right. In a year, in three. The thesis about an AI economy that runs into hardware and power has not gone anywhere. If anything the decline confirms it rather than contradicting it: expensive credit is the physical constraint he was writing about. He described the wall, then drove into it at his own speed.
A margin call simply does not read theses. It reads collateral and the clock.
What to take from this
No metaphors from here on, because this is the part you have to live with.
Being right is not collateral. No prime brokerage on the planet accepts a thesis, a model or the author’s reputation as security. They accept the closing print, and they remark it every day.
Leverage is sold to you as speed. What you are actually buying is time: the right to remain in a position exactly as long as the collateral value satisfies the person calculating it, and that person is not you. Any conversation about leverage that does not name the threshold and does not compute the price at which you are carried out is not a conversation about leverage. It is a conversation about hope.
Your lender sees your whole book. You never see his. The asymmetry here is not about intelligence and not about information on the market. It is about information on you.
A hedge built from one thesis is not a hedge. Long infrastructure and short software look like two sides of a structure. In practice they are one bet written down twice. In calm periods the correlations behave. Under stress they go to one, and you discover there was no diversification.
Liquidity is taken first. When you plan an exit, do not compute how much you make if you are right. Compute what gets taken from you first if the market goes against you for two weeks.
And last. The crowd does not exit into a decline, it averages down: six billion dollars of record inflows into chip funds in their worst month since 2008 is not courage, it is the arithmetic of somebody else’s liquidation. Someone had to be standing on the other side.
Intelligence turned out not to be an asset anyone accepts as collateral. That is not a verdict on intelligence. It is a reminder that the market asks two different questions: are you right, and can you sit through it. The first one your own head answers. The second one is answered by somebody else’s accounting department, and it answers first.
Numbers, not vibes
Sale of the entire public book to a single buyer, leverage as high as four to one, margin calls from Bank of America, Goldman Sachs and JPMorgan, prime brokers marketing baskets of positions before the open. Source: CNBC (David Faber), July 30, 2026.
Assets under management, growth from 225 million to more than twenty billion, a peak near 45 billion, roughly 439 percent net, the search for new capital. Sources: Bloomberg, FT, Hedgeweek.
First-quarter 2026 filing: 42 positions worth 13.68 billion, 8.46 billion of notional in puts, two billion against SMH, 1.6 billion against Nvidia; largest longs Bloom Energy, SanDisk, CoreWeave, IREN. Sources: 13F filed May 18, 2026, and the separate 13G on Nebius.
CoreWeave spreads near 855 basis points, interest expense of 536 million per quarter, free cash flow of minus 4.71 billion, insider sales. Sources: Bloomberg, Investing.com.
Semiconductor indices: minus 18.9 percent in July, worst month since 2008; 6.1 billion dollars of inflows into the largest chip ETF, a record in data going back to 2018. Sources: CNBC, Benzinga.
Korea: KOSPI minus 8.95 percent on July 13, sidecar and circuit breaker triggered, SK Hynix minus 15.37 percent in a session, Samsung minus 10.7; Goldman Sachs estimate of 1.2 million accounts receiving margin calls and 320 to 360 thousand forced closures cumulatively as of July 13; record margin lending of 38.6 trillion won; 2.3 trillion won of liquidations. Sources: KRX, Goldman Sachs, BigGo Finance, 36Kr.
Anthropic: stake acquired at a valuation on the order of 60 billion dollars, around a fifth of fund assets; Series H of 65 billion in May 2026 at 965 billion post-money; investor letter of July 24 and the subscription window from August 1. Sources: Anthropic, CryptoBriefing, Disruption Banking, Hedgeweek.
This is an essay, not investment advice. Everything described above happened to people who count better than you or I do. Leverage turns being right into zero faster than being wrong does. Think for yourself.