The thesis and the prodigy
In June 2024, shortly after leaving OpenAI's Superalignment team, published a 165-page essay series called Situational Awareness: The Decade Ahead. Its argument was that continued scaling of compute, algorithmic efficiency and what he called "unhobbling" — turning raw models into usable agents — made AGI plausible by 2027 and superintelligence by 2030. It also predicted the physical consequence: trillion-dollar compute clusters, built at speed, and a US-China race serious enough that the state might eventually take it over. He made the same case at length on the Dwarkesh podcast.
Most people who write manifestos about the future do not then take a position on it. Aschenbrenner did. Situational Awareness Partners LP was incorporated in Delaware on 6 May 2024, filed its first Form D in September 2024, and deployed its first capital on 1 November 2024. The seed was roughly $225 million, from Patrick and John Collison, Nat Friedman, Daniel Gross, and Graham Duncan of East Rock Capital. was reported as an investor, which for a first-time manager in his mid-twenties with no Wall Street background is the detail that made the Wall Street Journal write it up. Later coverage summarising WSJ reporting added Neil Mehta of Greenoaks, Dan Sundheim of D1 Capital, the foundation of XN's Gaurav Kapadia, and former Tiger Global public-equities head Feroz Dewan.
The logic was not complicated, and that was rather the point. If you believe the compute buildout is coming, you buy the things it consumes: memory, chips, power, data centres.
The $45 billion hyper-scaling
The trade worked, and then it worked absurdly. The fund returned about 37% in 2024, finishing the year with around $254 million. It gained 47% net in the first half of 2025, against roughly 6% for the S&P 500, and ended 2025 at about $1.5 billion. Then it returned 439% net in the first half of 2026. The Financial Times reported cumulative gains of 1,551% since opening in 2024; CNBC's coverage put it at more than 1,000% since inception. Assets passed $20 billion by May 2026.
Two things produced that. The first was roughly 4x . The second was concentration of a kind that would get a normal manager fired.
The Q2 2026 13F, filed on 14 August 2026, discloses 26 holdings worth $20.24 billion as of 30 June 2026. SanDisk was about $5.674 billion of it and Micron about $5.574 billion, which is over half the disclosed book in two memory names. Both were recent. In the Q1 2026 filing SanDisk was about $724 million and Micron about $58.6 million. In three months, a position went from $59 million to $5.6 billion. Bloom Energy was about $878.7 million, Taiwan Semiconductor about $1.27 billion, Nebius about $1.23 billion. (The portfolio-share percentages circulating from trackers do not reconcile with figures — one puts Bloom at 22.9% of a $20.2 billion book, which it plainly is not — so we are quoting dollars.)
None of this is exotic. There is no derivative anyone needs explaining. It is a young man with a thesis, four turns of borrowed money, and two enormous positions in memory chips.
The July 2026 margin call
The mechanics that ended it are the same ones that ended LTCM in 1998 and Archegos in 2021: concentration means diversification does not protect you against a common shock, leverage means the equity disappears long before the thesis is disproved, and margin calls mean you sell into the market you are already breaking.
The shock came from the AI infrastructure complex itself. had fallen about 48% from its 52-week high of $153.20 by 22 July 2026, hurt first by its 7 May Q1 report and then by Meta's 17 July launch of Meta Compute, which raised the possibility that a major customer becomes a competitor. On 29 July there was a further leg down attributed to forced liquidation and margin calls. That was the fund.
Situational Awareness lost about 67% in July 2026, roughly $30 billion, according to WSJ reporting on how Wall Street worked out it was in trouble. Both the long and short books moved against it. CNBC reported the fund collapsing from $45 billion to around $10 billion and the portfolio being sold to Citadel in a distressed transaction. The WSJ's account is titled "His Wedding Guests Were Arriving—Just as His $45 Billion Fund Was Collapsing", which we mention only because the timing is real and no invention would have improved it.
One genuinely strange number: after losing 67% in a month, the fund was reported as still up around 80% for the year. That is what a 439% first half buys you. It also explains why the story is not a morality tale. The manager and team had already crystallised fees on the way up; one widely circulated analysis derives a figure above $100 million, though no public source gives the actual contractual fee schedule. For a longer treatment of the finance-professor view, Aswath Damodaran has written his own post-mortem.
What is disputed
The peak AUM figure is not settled. WSJ and FT-sourced reporting, based on investor letters, use $45 billion in early July 2026. Other contemporaneous coverage puts it at around $24 billion in late July. Both can be true if the number was already falling, but nobody outside the firm has reconciled them publicly, and "AUM" for a 4x-levered book is a slippery quantity in the first place.
The other claim to treat carefully is the one in headlines about Silicon Valley "lining up to give him more cash" weeks after the loss. There are no on-the-record quotes from the Collisons, Nat Friedman or any other anchor LP in any of the reporting we have seen saying that. It is a characterisation, not a source. Likewise, a summary claiming he had $100 million committed "before he had turned 23" does not fit the documented timeline, in which he is 24 or 25 throughout.
Why it matters to this crash
The blow-up is the hinge in the story between "expensive" and "financed". Until July 2026 you could argue the infrastructure names were richly valued. Afterwards you had to accept that some of the bid underneath them had been borrowed, and that when it was called, the price found out.
What it left behind is a market that sorts the complex by balance sheet. We wrote on 18 August that the AI trade had split in two: over five days CoreWeave was down 15.6%, Nebius 13.6%, Talen 11.7%, Core Scientific 10.2%, Oracle 6.2%, while Microsoft was up 24% over twenty days and Alphabet was flat. The companies being sold borrow at project level; the ones being bought pay cash. The pricing point was the QTS "Project Odyssey" bond, $3.9bn at five years, rated Baa3/BBB−, marketed around 7.63% — a junk-like yield on an investment-grade rating for a Microsoft-linked data centre.
The fund's own fingerprints were still visible in the options market a month later. Implied volatility on the semiconductor ETF SMH collapsed from 65% in July to 40% on 17 August 2026, the lowest since February, as banks unwound hedges they had put on around the blow-up.
What would make this dangerous
So far this is one fund, one month, one buyer of the wreckage. Four things would change that.
A prime broker disclosing a loss. Archegos was dangerous not because Archegos died but because the banks financing it had each seen only their slice of the book. No lender to Situational Awareness has publicly reported a hit, and the identities and exposures of its prime brokers are not in the public record. If one shows up in a quarterly result, the question stops being about a hedge fund.
A second fund with the same book. WSJ reported billions flowing into new AI-focused hedge funds on the back of these returns. If the next 13F season shows two or three managers with half their disclosed equity in the same memory and power names, the on those positions is not an independent price, it is a queue.
The debt window closing on the infrastructure names. A 7.63% coupon on an investment-grade data centre bond is a warning, not a failure. A failed deal is a failure. If a neocloud or a converted miner cannot price project debt at all, the buildout stops for reasons that have nothing to do with demand for compute.
And the overhang itself. Citadel bought a distressed portfolio; what it does with it, and at what pace, is not public. A large forced seller who has already sold is harmless. A large reluctant holder who has not yet is a different animal, and the tape will tell you which one this is before any filing does.