Three months before it lost two-thirds of its value, the fund that became the AI trade’s biggest casualty was positioned against the AI trade. Situational Awareness LP’s own filing for the first quarter shows its five largest disclosed positions were all bearish: puts on the VanEck semiconductor ETF, on Nvidia, on Oracle, on Broadcom and on AMD. Put options made up 61.9% of the $13.7 billion book it reported to regulators.
By June 30, that had inverted. The quarterly filing it submitted in August shows a $20.2 billion portfolio that was 99.6% long, spread across 26 lines instead of 42, with two of them accounting for 55.6% of the whole thing. Micron and Taiwan Semiconductor, names it had held puts against in March, were now outright longs.
Three weeks later the fund lost roughly 67% in a single month, and JPMorgan has now ended its lending relationship with it, Reuters reported on Friday, citing the Financial Times. The filings explain what the headlines have not: this was not a bad bet on AI. It was a fast, total reversal of one.
Article Brief
Key Takeaways
5 Points30s Read
- The reversalOn March 31 the fund held puts on the semiconductor ETF, Nvidia, Oracle, Broadcom and AMD, 61.9% of its disclosed book. By June 30 it was 99.6% long.
- The concentrationSanDisk and Micron were 55.6% of a $20.2 billion disclosed portfolio, down from 42 positions to 26.
- The damageApplying its disclosed share counts to July prices, those two names lost $4.24 billion by July 31 and $5.14 billion at their lows.
- The financingPeak assets near $45 billion fell to roughly $10 billion, with leverage reported up to 400%. JPMorgan has now ended its lending to the fund.
- The aftermathCitadel bought about $16 billion of stock at a roughly 10% discount and has unwound over 80%. The SEC subpoenaed four banks over margin lending.
What the fund actually owned
The disclosed June book was a concentrated wager on memory and the power behind it. SanDisk was 28.0% of reported value, at 2,495,344 shares. Micron was 27.5%, at 4,828,786 shares. Bloom Energy came next at 9.4%, then Taiwan Semiconductor at 6.3% and Nebius at 6.1%. Below that sat a long tail of AI-infrastructure names TECHi readers will recognise: CoreWeave, Core Scientific, Applied Digital, IREN, Riot, CleanSpark, Bitdeer and Hive.
A 13F shows only long US equity and listed options, so it is a partial view by construction. It does not show shorts, swaps, foreign lines or leverage. But it is the fund’s own account of where its disclosed money sat 21 days before the drawdown, and the concentration in it is the part that matters. Two semiconductor names carried more than half the book.
What sat underneath them is its own argument. Bloom Energy, a fuel-cell maker selling on-site power to data centers, was the third-largest line at 9.4%. Nebius, the neocloud spun out of Yandex, was fifth at 6.1%. Add CoreWeave, Core Scientific, Applied Digital, IREN, Riot, CleanSpark, Bitdeer, Hive, Solaris Energy Infrastructure and Keel Infrastructure, and roughly a fifth of the book was a single wager that AI compute is constrained by electricity and the buildings that hold it. That is a coherent thesis, and TECHi has made a version of it repeatedly. The problem was never the idea. It was owning the idea four times over with borrowed money.
The exits say as much as the holdings. Nvidia, Oracle, Broadcom, AMD, ASML, Intel and the VanEck semiconductor ETF all disappear between the March and June filings, because the puts held against them were closed. Micron and Taiwan Semiconductor did not merely reappear; they switched sides. A fund that had been paying premium to bet against the chip complex in March was, by midsummer, its most levered owner. One line survives as a curiosity: 10,000 shares of Cerebras Systems, worth $2.21 million, or 0.01% of the book.
Applying the share counts it disclosed to the prices those two stocks actually printed, the arithmetic is brutal. SanDisk fell 46.6% in July and Micron fell 28.7%. On the June 30 positions, that is a $4.24 billion mark-to-market loss across the two names by the end of the month, and $5.14 billion at their July lows. The rest of the book was correlated to the same trade. Neither name had done anything wrong: SanDisk is still up more than 1,700% over twelve months and Micron more than 500%.
The mechanism was leverage, not stock picking
Reporting on the collapse has put the fund’s peak assets near $45 billion, falling to roughly $10 billion, with leverage of up to 400%. At that ratio, a 15% move against a concentrated book is an extinction event, and July delivered far more than 15% in the two names that were the book.
The sequence that followed is the one prime brokerage is designed to produce. Margin calls arrived from multiple lenders. The fund sold roughly $16 billion of listed equities to Citadel, which bought the positions at a discount of about 10% and has since unwound more than 80% of that risk through nearly 100 block trades worth over $4 billion. Ken Griffin told clients the deal “could not have been completed without the extraordinary cooperation of the trading and prime brokerage teams at the banks serving both firms,” according to Disruption Banking.
That cooperation is now the subject of an investigation. The SEC has subpoenaed Goldman Sachs, JPMorgan, Citigroup and Bank of America, examining the timing of the trades that triggered the margin calls and the banks’ communications with the fund about its use of leverage, Reuters reported in August after the New York Times broke the story. The fund said it expects regulators to examine funds that “produce significant returns, or have particularly dramatic drawdowns” and will cooperate. Leopold Aschenbrenner, the former OpenAI researcher who founded it in 2024, wrote to investors that the firm “came closer to permanent capital impairment than is acceptable to us.”
Why this is a memory-stock story
The damage did not stay inside the fund. Forced selling of a $16 billion book, then weeks of block trades unwinding it, lands on the order book of whatever the book contained. TECHi has tracked both ends of that: SanDisk was the S&P 500’s best stock of the first half before this, and Micron passed its $41 billion test while the stock failed anyway.
The recovery since tells you the selling was technical rather than fundamental. From July 31, SanDisk is up 34.5%, Micron 18.5% and CoreWeave 24.0%. None of those moves required new information about memory pricing or AI capex. They required the seller to finish selling. For anyone holding Micron or SanDisk through that window, the lesson is that a single fund’s margin clerk set the price for several weeks.
The bid that is disappearing
JPMorgan’s withdrawal is the part with forward consequences. Prime brokerage is how concentrated funds turn conviction into size, and the biggest bank on Wall Street has now decided this client is not worth the balance sheet. The fund has moved toward smaller brokers, which means smaller lines, tighter terms and less capacity to absorb a drawdown.
Multiply that across the industry and it becomes a slow drag on exactly the stocks TECHi covers. Every prime broker that repriced AI concentration after July is now lending less against memory, neocloud and data-center power names than it did in June. The same repricing is visible elsewhere in the AI-adjacent complex, from the Fed’s effect on Bitcoin’s golden cross to the ETF bid now setting the price of Zcash. That is not a crash; it is a lower ceiling on how much levered money can chase the next leg. The neocloud names and CoreWeave’s debt-funded backlog are the most exposed, because they are the ones whose equity stories depend on continuous access to capital.
What the fund did next
The most revealing disclosure is the most recent. On Aug. 28, Situational Awareness filed a Schedule 13D reporting a 21.1% stake in SharonAI Holdings, a Nasdaq-listed AI cloud company: 5,396,127 Class A shares plus prefunded warrants for 2,674,823 more, held jointly with Carl Shulman and declared as an investment in stock the filers believed was undervalued.
It already owned the position in June, at $456.8 million, and it has not sold. A fund that was margin-called out of its liquid book has kept and disclosed a fifth of a small-cap AI compute company whose stock trades near $59 against an August high of $68.78, on roughly 1.8 million shares a day. At that volume, the disclosed stake is worth several days of total trading, which makes it the opposite of a liquid position.
SharonAI is not a shell. The Australian neocloud listed on Nasdaq in February, raising $125 million, and has since signed a five-year, $1.25 billion contract with ESDS Software Solutions and a $950 million agreement with an unnamed Asia-Pacific technology company, while expanding planned data-center capacity from 70MW to 100MW and announcing a $350 million convertible note. It is a credible business and an illiquid stock at the same time, which is precisely the combination that a levered fund cannot own at size and an unlevered one can.
Read against the collapse, the position says what Aschenbrenner still believes. The thesis was never the problem. The financing was.
The number that decides what happens next
The fund’s disclosed book will be updated in mid-November, for the quarter ending Sept. 30. That filing will show whether the memory concentration survived, whether the AI-infrastructure tail was rebuilt, and whether anything replaced the leverage JPMorgan pulled.
Until then, the honest read is narrow. The AI trade’s most aggressive expression blew up on financing, not on thesis, and the stocks it was forced to sell have recovered most of the damage without it. What has not come back is the leverage. For the memory and neocloud complex, the marginal buyer of the past year is smaller than it was, and the banks that funded him are answering subpoenas about how they did it.