Leopold Aschenbrenner built the defining Artificial Intelligence (AI)-infrastructure thesis of the decade, turned it into a hedge fund that returned roughly 439% through June 2026, and scaled it from about $225 million to as much as $45 billion. Six trading days later, Ken Griffin’s Citadel owned the public book. The AI investment theory wasn’t wrong, but the room that ran it had no one who could model the machine on the other side of the table.
A timeless tale of living by the sword, and dying by it - AI Edition.
The decision
Aschenbrenner’s call was pure second-derivative AI. After OpenAI fired him from the Superalignment team in 2024 (he says unjustifiably) after an improper disclosure, he published Situational Awareness: The Decade Ahead, a 165-page argument that Artificial General Intelligence (AGI) is an infrastructure problem: chips, power, memory, data centers. Then he did what theorists and essayists almost never do. He raised a fund on the thesis, backed early by names including Stripe’s Collison brothers and former GitHub Chief Executive Officer (CEO) Nat Friedman, and built the portfolio to match: long the physical layer (SanDisk, Bloom Energy, CoreWeave, IREN, Core Scientific, Applied Digital) hedged with billions in puts on the consensus mega-cap semis and shorts on legacy software.
Through June 30 it was the best trade in the world. Then July hit. Meta’s Meta Compute announcement on July 17 put Big Tech in direct competition with the GPU-cloud names at the heart of the book. Core holdings fell 27–54% while the Nasdaq 100 dropped just over 10%. The hedges failed the wrong way: software shorts moved against the fund as money rotated out of AI hardware, and the semi puts protected a mega-cap crash that never came. At roughly 4x gross leverage, a ~30% hit to the longs is a triple-digit hit to equity. Goldman Sachs, JPMorgan, and Bank of America issued margin calls. On July 24 (already deep in the drawdown) Aschenbrenner wrote investors that the selloff was one of the best buying opportunities in over a year and invited fresh capital from August 1. Five days later the primes forced the hand. Before the July 30 open, [the entire public portfolio, longs and shorts, went to Citadel in a single block. The fund kept its private stakes, including Anthropic. The names he was forced out of bottomed on July 29 and bounced hard as Citadel absorbed the book. IREN (Iris Energy) closed roughly 30% off its liquidation low within a day.
The team’s cognitive composition
Now let’s talk about the room, but stripped to its vectors. Aschenbrenner is the product of a single brilliant track: Columbia valedictorian at 19, OpenAI Superalignment at 21, the most-read AGI infrastructure essayist of 2024, fund founder by 24. The cognition that built Situational Awareness is domain-conviction cognition which equates to 1) map the decade, 2) concentrate the book, and 3) size into the truth. It is the mind that writes the memo everyone else forwards. It is not the mind that sits on a multi-strategy risk desk pulling from decades of real world experience and asks what happens when your famous SEC Form 13F quarterly holdings filing becomes a map the street uses to front-run your forced sales.
There is no public evidence of a counterbalance in that room with equal weight: no Long-Term Capital Management (LTCM) survivor, no prime-brokerage risk operator, no multi-strategy portfolio manager whose job it is to survive being wrong for longer than the essay allows. Conviction was the product. Leverage was the amplifier. The Form 13F made the book legible. When a wounded 4x whale has to sell, every proprietary desk on the street already knows the names.
Now put Ken Griffin across the table. Citadel is not a conspiracy and it is not a morality play. It is the opposite cognitive system: multi-strategy, multi-pod, industrial risk machinery built to shape the weather and warehouse the wreckage. In the days before the AI trade broke, market chatter had Citadel amplifying surprise-rate-hike fears into a hawkish July Federal Reserve (Fed).
Griffin has spent three decades turning narrative, liquidity, and other people’s forced selling into one system. A single-thesis room that cannot model a counterparty who may be both “the weather and the bid” is not competing. It is inventory or the product they buy when you break.
What a Decision Simulation would have surfaced
Load the room as cognitive vectors and you do not get a spread. You get one dominant vector (AGI-infrastructure conviction) at maximum intensity, levered 4x, with a public fingerprint. The stress-test is not “is the AI build-out real?” The stress-test is: “Who in this room is modeling the world where the thesis is early, the theme drops 30%, a hawkish Fed narrative starves the bid, the primes call the loan, and the same multi-strategy firm that helped price the fear is the bid for your book? Who is pricing survival against a counterparty that can move the weather, not just correctness against a spreadsheet?” In Aschenbrenner’s room the answer appears to be no one, because survival under forced liquidation is not how that cognition scores decisions. A simulation flags the gap the day they chose 4x on a crowded, famous book — as a risk conversation. It surfaced instead as a pre-open block to Citadel.
The lesson
Wall Street has a long list of rooms that were right and still died: LTCM, Clarium, now Situational Awareness. The pattern is not stupidity. It is composition. A single unchecked cognition however sharp, however validated by a 439% run only wins the game it was tuned for and loses the one it wasn’t. Aschenbrenner’s game was seeing the AI decade early. Griffin’s game is moving the conditions and owning the other side when the early seer cannot stay solvent. Going up against market movers like Citadel without anyone in the room who thinks like them and who models narrative pressure, prime-broker clocks, and the buyer of last resort as one system, is not courage, it is a predictable output of who sits at the table.
The market does not grade your essay. It grades whether your decision room can survive the man who can help break the book and then buy it.
As executives everywhere clamor to embrace AI, the ones using AI simulations to stress-test who is missing from the room before the margin call may be making the most human choice possible.