The trade that ended Situational Awareness LP took one afternoon and, according to CNBC’s David Faber, went through one enormous trade. On Wednesday, Leopold Aschenbrenner sold the bulk of his fund’s public equity portfolio to Ken Griffin’s Citadel, six days after telling investors that July’s sell-off had created attractive opportunities and inviting them to add capital on August 1.
That capital never arrived, but the margin calls did. The Situation had changed.
Hold onto one thing before the schadenfreude sets in, because it is the only part of this story that will still matter in a year. Aschenbrenner’s central claim, made in June 2024 when almost nobody was making it, was that the binding constraint on artificial intelligence would be physical rather than algorithmic: chips, memory, data centers, electricity. That claim has been vindicated so comprehensively that it is now the consensus assumption inside every hyperscaler capital budget on earth. Microsoft held its outlook at roughly $190 billion for calendar 2026 on Wednesday night. Alphabet has guided to about $200 billion over twelve months and posted its first negative quarterly free cash flow. The buildout he described is happening at a scale that outran his own forecasts.
Wednesday
The fund was under pressure to either raise fresh capital from investors or offload its entire book, and eventually chose the latter, two sources told Reuters. A number of top Wall Street prime brokers, including Goldman Sachs, JPMorgan Chase, Bank of America and Citigroup, helped facilitate the deal between Citadel and Aschenbrenner’s fund.
The same banks had issued the margin calls that made the deal necessary. The fund ran roughly four times leverage on its public equity positions, which is arithmetic rather than opinion. At 4x, a 25% decline in the underlying consumes the entire equity contribution. The positions in question fell between 35% and 47%.
The largest holdings at the end of the first quarter were Nebius Group, SanDisk, Micron and CoreWeave, all of which are down more than 35% this month. The fund also took heavy losses on infrastructure positions including SK Hynix, while short bets against software companies such as Adobe went the wrong way at the same time. That combination is the worst one available. The longs fell, the shorts rose, and leverage does not care which leg is which.
Millennium and Jane Street, an existing investor in the fund, both bid on the book and were outbid. Roughly two thirds of the fund’s assets under management were public equities, both long and short, according to Faber. Assets peaked near $45 billion. By the time of the sale the figure was just over $20 billion.
The essay that became the market’s script
Aschenbrenner is 25. He was born in Germany, graduated from Columbia as valedictorian at 19, joined OpenAI’s Superalignment team, and was fired in April 2024 over an alleged information security leak, a characterization he has consistently disputed.
Weeks later came Situational Awareness: The Decade Ahead, 165 pages arguing that AGI was plausible by 2027 and that the world was not remotely prepared. He declared that “the AGI race has begun” and forecast trillion-dollar compute clusters and hundreds of millions of GPUs. Of everyone else, he wrote that “few have the faintest glimmer of what is about to hit them”. Michael Dell shared it. So did Ivanka Trump.
Then he traded it. With AI researcher Carl Shulman he raised roughly $225 million in September 2024 from Stripe’s Patrick and John Collison, Nat Friedman and Daniel Gross, on a $25 million minimum and a two-year lock-up. He had no professional trading experience. The most recent filing showed four investment professionals and eight employees in total.
The bet was never on models. It was on what models need, and his formulation was blunt: “It’s not algorithms that are going to be the bottleneck”. It’s electrons. The buildout, he wrote in the essay, would be “a race to mobilize America’s industrial might”.
The record was extraordinary and brief. Up 47% after fees in the first half of 2025. Roughly 200% for the full year. And then, per the July 24 investor letter reviewed by the FT, 439% net through the end of June 2026. Jane Street, which almost never backs outside managers, came in as an investor well after launch. When a stake in T1 Energy surfaced in May the shares rose 23% in a day. A retail platform built a feature letting users mirror his filings.
None of it was visible as leverage. A 13F discloses long US-listed equity and listed options at quarter end. It does not show short positions, swap exposure, financing terms or fund-level performance. The Q1 filing showed $8.46 billion of notional put exposure against chipmakers, roughly 60% of the disclosed book, alongside long positions in Bloom Energy, CoreWeave, Nebius and a set of former miners converting sites to AI hosting. What it never showed was how much of that was financed.
Enter the man who has done this before
Ken Griffin has been buying the wreckage of over-levered funds for twenty years, and he is very good at it.
He started trading convertible bonds from his Harvard dorm room in 1986, having persuaded the university to let him mount a satellite dish on the roof for real-time quotes. He founded Citadel in 1990, a year after graduating, with $4.6 million. Citadel now manages about $71 billion and is the most profitable hedge fund in history by cumulative net gains, with roughly $90 billion returned to investors since inception. Its 2022 gain of $16 billion was the largest single year any hedge fund has produced, surpassing John Paulson’s subprime trade.
The pattern is consistent and it is not sentimental. In 2001, days after Enron’s bankruptcy, Griffin raided its trading floor for talent and built a commodities desk out of it. In 2006, when Amaranth Advisors lost roughly $6 billion and 65% of its assets on natural gas, Citadel and JPMorgan bought the entire energy portfolio at a steep discount and turned a sizable profit on it. In 2007, when Sowood Capital imploded in the early credit dislocation, Citadel assembled a team that worked through the night and put an offer on the table at 3:30 in the morning while competing firms had gone home. It closed. The same year Citadel put $2.5 billion into E*Trade. In 2021 it wrote a $2 billion check into Melvin Capital, the most public casualty of the GameStop squeeze.
Griffin built the firm this way deliberately, having watched Long-Term Capital Management disintegrate in 1998. “We had planned for a repeat of the crash of ’87”, he told Fortune in 2008.
Which brings us to the part of this story that nobody running the Archegos comparison seems to have noticed. In 2008 Citadel itself was leveraged 7:1 into convertible bonds, fell 55%, lost $9 billion, was losing hundreds of millions a week, and gated its investors to stop a forced liquidation. Griffin has called it his low point. The firm survived on prime brokerage terms he had negotiated years earlier precisely because he had studied LTCM, on a $500 million injection from his own partners, and on the fact that he controlled the gate. It rebuilt with materially lower leverage and stringent liquidity requirements, ran 500 stress tests daily, and did not cross its high-water mark again until January 2012.
Griffin knows exactly what happened to Aschenbrenner this week because it very nearly happened to him. The difference is that Griffin had planned for the scenario for a decade before it arrived, and could stop his own investors from leaving. Aschenbrenner had four investment professionals and a prime brokerage agreement.
The Fed, and a matter of timing
Wednesday was already a bad day to be a forced seller. Ahead of the FOMC decision, Citadel Securities had publicly called for a surprise 25 basis point hike against a consensus expecting a hold, arguing that moving immediately would establish new chair Kevin Warsh’s inflation-fighting credibility and end the era of heavy forward guidance. The call landed hard, contributing to record demand for hedges against a surprise increase. The Fed held at 3.50% to 3.75% but paired it with a firm warning on inflation. Thirty-year Treasury yields spiked to their highest since 2007, the Dow fell nearly 800 points at one stage, and equities logged their worst Fed day in more than eighteen months, led lower by exactly the high-growth technology names sitting in Situational Awareness’s book.
Citadel Securities is the market-making arm and a legally separate entity from the Citadel hedge fund, with information barriers between them, and there is no suggestion whatsoever of coordination. But it is a matter of record that on the same day one Griffin firm was publicly stoking rate-hike anxiety into a fragile tech tape, another was buying a distressed AI portfolio from a manager whose collateral was being marked down by that tape. Wall Street will enjoy that symmetry for some time.
What is left
The private book survives, and with it the most interesting position in the affair. Situational Awareness did not sell its private portfolio, which includes a significant stake in Anthropic, expected to go public within months and last valued at $965 billion in the May Series H. The fund participated in that $65 billion round, and Anthropic confidentially submitted a draft registration statement on June 1. The FT puts the stake at around $5 billion, and a spokesman told CNBC that reports the fund was shopping it are “not accurate.”
So Situational Awareness continues, as a private investment vehicle that is substantially a bet on one company going public, with a residual trading operation attached. Aschenbrenner is engaged to Avital Balwit, chief of staff to Anthropic’s chief executive Dario Amodei. A person close to the fund says it kept a small portion of its public book and will keep trading equities.
The comparisons arrived within the hour. ZeroHedge called it “Archegos 2.0”. Semafor framed it as an AI-age sequel to the Bill Hwang collapse. Forbes reached for both Archegos and LTCM before conceding that neither quite fits.
They don’t, for one reason. Hwang was wrong about the companies. Aschenbrenner was right about them, which is why the coda is so strange. The very stocks the fund was forced to dump jumped sharply the next day, with Micron, Nebius, CoreWeave and the broader infrastructure complex up double digits on a reading that the forced-seller overhang had cleared.
His book recovered without him.
What the thesis got right
Strip out the leverage and the essay’s core diagnosis is holding up better than almost any other 2024 forecast about artificial intelligence.
Physical constraints are real and binding. Power, land, memory, cooling and interconnect determine what actually gets built and how fast. The market spent 2024 and most of 2025 pricing AI as a software story. Aschenbrenner priced it as an industrial one, and every hyperscaler capital budget since has agreed with him.
The picks-and-shovels layer was genuinely mispriced. Buying substations, neoclouds and memory rather than the model developers or the consensus chip names was the correct structural call, and it produced 47% in his first half year, roughly 200% in his first full one, and 439% in six months of 2026. That is not luck applied to a wrong idea.
The miner-to-AI pivot was commercially validated, not just narratively. Multi-billion-dollar hosting contracts have landed at converted sites, and the operators he bought early are now infrastructure businesses with real backlogs.
Frontier-lab private exposure was the cleanest expression of the thesis, and it is the only part of the book he still owns. The Anthropic stake was bought at just above a $60 billion valuation in February 2025. It is the position that did not get margin-called, because nobody could mark it against him daily.
What anyone building or investing in AI should take from this
Physical reality bites, and it bites the balance sheet before it bites the technology. Capex can run ahead of proven returns for a long time, but markets eventually demand unit economics, and they demand them on their own schedule rather than yours. Dry powder and liquidity matter as much as conviction, because surviving the cycle is a precondition for harvesting the thesis.
And there is a question this week has raised that nobody yet has the data to answer. If leverage of this order was invisible in the filings of the single most closely watched fund in the AI trade, who else is running it? As BNC argued in April when the first cracks appeared, the revenue underneath this cycle is real in a way the dot-com comparison never was. What is also real is the borrowing stacked on top of it, and that does not appear in anybody’s capex guidance.
Aschenbrenner wrote that few people had the faintest glimmer of what was about to hit them. He was describing an intelligence explosion. What actually hit was a margin clerk. Even so, he may still be right about the decade ahead. Strap in, the situation is still unfolding…
Be the first to comment