On July 30, 2026, before the opening bell, a twenty-five-year-old sold the entire public equity book of a $45 billion hedge fund to Ken Griffin in a single block trade negotiated overnight. Long positions and short positions, everything, at a discount, to the one buyer with a balance sheet deep enough to absorb it on a few hours’ notice. By Thursday afternoon the fund’s assets had fallen to roughly $10 billion. The July loss came in at 67 percent. Citadel beat out Millennium and Jane Street, and the deal closed shortly after dawn.
Six days earlier, that same twenty-five-year-old had written to his investors that the selloff hammering his portfolio represented one of the most attractive entry points since early 2025, and invited them to wire additional capital on August 1. Two of them later told the Financial Times the letter had reassured them. By the following Wednesday those investors could not get him on the phone. One of them put it plainly: Leopold just stopped taking calls.
The financial press has covered this as a story about a wunderkind who flew too close to the sun. That framing is comfortable and it is wrong. The interesting question is not how Leopold Aschenbrenner lost the money. Four times leverage on a concentrated book of illiquid small caps loses money in exactly one way, and everyone who has worked a risk desk knew the shape of it in advance. The interesting question is why anyone gave him the money in the first place, when every idea in the portfolio was published quarterly, for free, on a government website.
Everything He Owned Was a Matter of Public Record
Hedge funds above a certain size file a Form 13F with the SEC every quarter. It lists their disclosed long positions in US-listed equities. It is free. It is machine-readable. Half a dozen websites reformat it into charts within hours of the filing hitting EDGAR.
Aschenbrenner’s book as of the first quarter of 2026 was roughly forty percent Nebius Group, twelve percent Bloom Energy, ten percent SanDisk, eight percent CoreWeave, and six percent SharonAI Holdings. The top five names accounted for more than seventy-six percent of the entire portfolio. Below that sat a dense basket of former bitcoin miners — IREN, Core Scientific, Applied Digital, Riot, CleanSpark, Bitdeer — held not because he liked crypto but because those companies owned contracted megawatts, substations, cooling and interconnection agreements that would otherwise take three to five years and billions of dollars to build from scratch. There was no meaningful position in Nvidia, Microsoft, Google, Amazon or Meta. The single largest holding in a fund built on the artificial intelligence trade was a Yandex spinout, and the second largest was a fuel cell company.
That is a genuinely good piece of research. I want to be fair about this, because the reflexive dunk is that he got lucky, and he did not get lucky on stock selection. Identifying that utility interconnection queues in most American markets now run four to seven years, that hyperscalers need gigawatts online inside twenty-four months, and that a dying bitcoin miner is therefore a data center with a bad income statement — that is a real insight, arrived at early, expressed with conviction. Several of those names were up well over a hundred percent on average even after July wiped them out.
But it was published. Every ninety days. To anyone with a browser.
His followers tracked his quarterly filings for stock tips. That was not a side effect of the strategy. By the end, it was the strategy’s largest single risk.
This is where the popular suggestion — just buy his portfolio yourself — stops being clever and starts being the actual mechanism of the collapse. Thousands of people did buy it. Retail traders, copy-trade accounts, momentum funds, and other hedge funds ran the same names, in the same weights, for the same reasons. When you and ten thousand strangers all own the same thinly traded small-cap and the exit gets narrow, you do not have a portfolio. You have a queue.
Then the Financial Times published the contents of his July 24 letter. Once the market knew a levered holder of those specific names was in trouble, the names themselves became a target. Aschenbrenner described what followed to his own investors as essentially similar to a bank run, vulnerability begetting more vulnerability. He was describing the predictable consequence of running a public portfolio in public. His positions fell thirty percent and more in three trading days. The morning after Citadel took the book off his hands, many of the same stocks rallied fifteen to twenty-five percent.
So yes, anyone could have bought his portfolio. The problem was never access. The problem is that by the time you can see a 13F it is forty-five days stale, and the crowd standing in it with you is the risk you did not price.
Four People, Eight Employees, Forty-Five Billion Dollars
Here is the number that should end the debate about what a modern fund is actually for. Regulatory filings showed Situational Awareness ran with four investment professionals and eight employees in total, against assets that peaked at forty-five billion dollars.
I have spent close to thirty years building production software, most of it on the Microsoft stack, and I have spent the last two building agentic systems that run unattended against real money and real customers. So let me state the thing the finance industry would prefer stayed vague. Monitoring a levered book of Korean memory stocks is not hard anymore. It is a weekend project.
What does that system actually consist of? A market data feed. A position table with cost basis, share count and current mark. A margin engine that recomputes maintenance requirements per prime broker agreement on every tick and tells you your distance to a call in dollars, not in vibes. A correlation matrix that flags when your supposedly diversified sleeve of six different companies is in fact one trade wearing six tickers. A drawdown monitor with hard thresholds. A liquidity model that divides your position size by trailing average daily volume and tells you how many days it would take to exit without moving the price. An agent loop that reads filings, earnings transcripts and news the moment they publish, and escalates when something in the thesis breaks. A scheduler. A logger. An alerting path that reaches a human at three in the morning.
Every one of those components is commodity engineering in 2026. The data is a paid API. The margin math is arithmetic that a junior developer can implement correctly in an afternoon. The correlation clustering is a library call. The document ingestion and escalation logic is the exact class of problem that agent frameworks and Model Context Protocol servers were built to solve, and I have shipped that pattern into production for clients who do not have a billion dollars and cannot afford to be wrong. The entire stack is a few thousand lines of code, some cloud spend, and a competent architect. It is not a firm. It is not a two-and-twenty. It is a service you could rent.
If information was not the moat and monitoring was not the moat, then what exactly was the fee for?
Now the part I refuse to pretend otherwise about, because the honest version is far more damning. Better monitoring would not have saved him.
He was not flying blind. The Financial Times reported that at least one of his prime brokers had already placed the fund on an internal watch list because the portfolio was too concentrated and too volatile, and that the same lender had demanded additional margin repeatedly over the preceding year, including during the fund’s final week. He knew. His lenders knew. Someone who was in the room described him as having the illusion of still being in control, and not having control at all.
That is the whole lesson, and it has nothing to do with dashboards. At four times leverage, a thirty percent decline in your longs against a simultaneous rally in your shorts does not produce a warning you can act on. It produces a solvency event. Information stops mattering the moment your options collapse to a single choice, and the choice is which distressed buyer gets your book. Aschenbrenner’s failure was not that he lacked situational awareness. It is that he had it, wrote a famous document about it, named his fund after it, and it did not help him at all.
The Manifesto Was a Fundraising Instrument
In June 2024, two months after being fired from OpenAI’s Superalignment team over an alleged information leak he disputes to this day, Aschenbrenner published a hundred and sixty-five page essay titled Situational Awareness: The Decade Ahead. It became required reading across Silicon Valley. Ivanka Trump praised it publicly. It is the document that turned a twenty-two-year-old with no track record managing money into a person whom tech chief executives, investors and policymakers took private meetings with.
Read it as an engineer rather than as a believer and the machinery becomes visible.
The core method is extrapolation. Draw the trendlines of compute scaling, algorithmic efficiency and measured capability gains, plot them on a logarithmic axis, extend them forward, and observe where they intersect. From that he concluded that systems capable of matching top AI researchers would arrive around 2027, that automated AI research would then trigger an intelligence explosion, that superintelligence follows within a few years of that, and that the resulting national security emergency would culminate in the American government effectively nationalizing the frontier labs before the decade was out.
None of that is a proof. It is a straight line drawn on a chart with the tone of a classified briefing. There are no error bars. There is no serious treatment of what happens if any single trendline bends, and every trendline in the history of technology has bent. The argument is unfalsifiable on the timescale that mattered for raising the money, which is the most commercially useful property a prediction can have.
It is now August 2026. AGI has not arrived. The 2027 clock has roughly seventeen months left on it and nothing currently shipping resembles a system that matches a top researcher across the board. The infrastructure predictions, on the other hand, were directionally excellent — power really is the binding constraint, memory really did go supply-constrained, the data center buildout really is historic. But here is the part nobody wants to say out loud: you did not need artificial general intelligence by 2027 to get to that conclusion. Any data center operator in Northern Virginia could have told you about the interconnection queue in 2023. The prophecy was not load-bearing for the trade. It was load-bearing for the raise.
Then there is the rhetoric, which is the real product. He wrote that before long the world would wake up, and that only a few hundred people, most of them in the labs, had any real sense of what was coming. He wrote that few have the faintest glimmer of what is about to hit them. That is not analysis. That is an in-group being constructed on the page, and the reader is being invited into it. You are reading this, therefore you are one of the few hundred. It is the single most effective sentence structure ever devised for selling a financial product, and it has been used to sell every one of them, from Dutch tulips to mortgage tranches to a crypto exchange run out of a Bahamas penthouse.
Right Credential, Right Room, Right Moment of Confusion
Consider what the package actually looked like in the second half of 2024, and consider who was buying.
Capital allocators understood that artificial intelligence was going to be the defining trade of the decade. They did not understand the technology, they could not evaluate the claims, and they had no way to distinguish a person with genuine inside knowledge from a person performing it. In that specific condition of informed ignorance, credentials do not supplement judgment. They replace it.
Aschenbrenner’s credentials were close to perfect for the moment. Valedictorian of Columbia at nineteen. A degree in economics and mathematical statistics. A seat on OpenAI’s Superalignment team under Ilya Sutskever. A firing that he framed, plausibly and effectively, as retaliation for warning the board that the company’s security could not withstand Chinese espionage. That is not a résumé. That is a legend, and it functions in a room the way a security clearance functions: it converts assertions into intelligence. When he said the labs were going to need a trillion dollars of compute, it did not land as a forecast. It landed as a leak.
What gets skipped in the retelling is that his prior employer was the FTX Future Fund, the philanthropic arm of Sam Bankman-Fried’s exchange, where he worked alongside William MacAskill and where the money turned out to have been taken from customer accounts. There is no evidence anyone on that team knew or did anything illegal, and I am not suggesting otherwise. I am pointing out that the same combination — extraordinary credential, total certainty, a moral framework that made questioning the numbers feel small — had already produced one of the largest frauds in American history, with some of the same people in the room, and the market extended the pattern full credit a second time inside of two years.
What the Money Was Actually For
Which brings us back to the question I started with. If the ideas were public, and the monitoring is commodity engineering, why did four extremely wealthy and extremely capable people hand a twenty-two-year-old two hundred and twenty-five million dollars instead of simply buying the stocks themselves?
Patrick and John Collison run Stripe. Nat Friedman and Daniel Gross had their own operations. None of them was going to sit in front of a screen managing a levered book of Korean memory stocks, and at their net worth, two hundred and twenty-five million was a call option on a thesis they already half believed, written by someone inside their own social graph. The downside was noise and the upside was being early to the defining trade of the decade. They were, incidentally, right. Even after the collapse, day-one capital is up something like fifteen times.
But the real answer is the one nobody puts in the headline. You cannot get four times leverage in a personal brokerage account. Goldman Sachs, JPMorgan Chase and Bank of America will extend that to a fund with an ISDA, a prime brokerage agreement and an institutional relationship. They will not extend it to a billionaire’s retail account, no matter whose name is on it. A meaningful share of the fifteen hundred and fifty-one percent since inception was not stock picking. It was a multiplier that is only available to a vehicle.
The fund was never a way to access his brain. It was a way to access a balance sheet.
And that multiplier is precisely what turned a thirty percent drawdown into a total forced liquidation. The same instrument that produced the returns produced the wipeout. It was one decision, and it was the only part of the entire enterprise that was not freely available to the public.
Which is the part I cannot replicate, and I will say so directly. I have shipped production software for Microsoft, Sony, Rockstar Games, Best Buy and Allegiant Air. I have spent twenty-nine years learning how systems actually fail in production. I do not have Patrick Collison’s phone number, and no amount of shipping will get it for me. The access that converted an essay into two hundred and twenty-five million dollars, and two hundred and twenty-five million dollars into a prime brokerage relationship, is a closed loop that opens for people who already stand inside it. That is not a complaint about meritocracy. It is a description of how capital formation actually works, and pretending otherwise is how founders waste years.
The Scoreboard
Citadel bought a distressed book from a forced seller at a discount, and watched a good portion of it rally fifteen to twenty-five percent the next morning. Aschenbrenner’s earliest investors remain enormously profitable. The fund finished July still up roughly eighty percent for the calendar year. He retains a private stake in Anthropic worth something like five billion dollars, acquired in early 2025 at a sixty billion dollar valuation, which no margin clerk could touch because nobody marks it every second.
The people who actually lost are the ones who arrived last. Assets went from roughly one and a half billion dollars to forty-five billion in about seven months, which means the overwhelming majority of the dollars in that fund walked in at or near the top and then absorbed a sixty-seven percent month. Jane Street is in that group. So is everyone who read the 439 percent letter and decided to size up.
And the thesis? It may well be correct. Unlevered, or even at a sane two times, he would be having a top-decile year and this piece would not exist. That is the actual lesson, and it is not the one about hubris. Being right is not a risk management strategy. A brilliant essay is a fundraising instrument, not a proof. And when someone tells you that only a few hundred people in the world can see what is coming, the correct response is to ask what they are selling — not because they are necessarily wrong, but because that sentence has never once been uttered by someone who wasn’t.
Disclaimer: This article is commentary and analysis. It is not investment advice, financial advice, or a recommendation to buy, sell or hold any security. I am not a licensed financial advisor, broker or investment professional, and nothing here should be treated as a solicitation or as a suggestion to replicate any strategy or portfolio described. All figures are drawn from public reporting by the Financial Times, the Wall Street Journal, Reuters, CNBC and Bloomberg, and from public SEC filings, and may be incomplete, superseded or wrong. Leverage of the kind described in this article can and does result in the total loss of invested capital. Do your own research and consult a qualified professional before making any investment decision.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment. He has spent nearly thirty years shipping production software on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy and Allegiant Air. He was employee number six at Break.com and a finalist for the Los Angeles Business Journal’s CTO of the Year.
His current work centers on production agentic AI — Model Context Protocol servers, the Microsoft Agent Framework, retrieval pipelines, SQL Server 2025 vector search, and the PLogger observability framework — the unglamorous engineering that determines whether an autonomous system can be trusted with anything that matters. He also builds and operates ShopSnap, a zero-transaction-fee e-commerce platform, along with several other production web properties.
He writes about what actually ships, as opposed to what gets announced. He co-hosts Edge Grip Podcast, trains Brazilian jiu-jitsu in Las Vegas, rides motorcycles, and is the author of the novel The Archive of Lost Suns.