The Fund Called Situational Awareness Grew 80-Fold in Six Quarters. It Lost Most of It in One.

In June 2024, two months after OpenAI fired him, Leopold Aschenbrenner published a 165-page essay called Situational Awareness: The Decade Ahead. Its argument was that the race to artificial general intelligence would run into a physical wall before it ran into an algorithmic one. Training a frontier model, he wrote, would soon mean building power plants: “By the end of the decade, we are headed to $1T+ individual training clusters, requiring power equivalent to >20% of US electricity production.” The bottleneck was chips, power, buildings and memory, not ideas.

A month later he put money on it. He founded a hedge fund, gave it the essay’s name, and raised it from backers that reportedly included Patrick and John Collison, Daniel Gross and Nat Friedman (Wikipedia; TechCrunch). Situational Awareness the fund spent two years buying the physical bottleneck Situational Awareness the essay described. Then, over about a month this summer, it lost most of what it had, to a bottleneck the essay never mentioned.

the filings

Every quarter, a fund managing more than $100 million in US-listed equities has to tell the SEC what it owns. I read Situational Awareness’s seven filings myself, straight from EDGAR (CIK 0002045724), and summed the value column on each one.

Situational Awareness LP's disclosed US equity value, seven quarterly SEC filings A log-scale line from December 2024 to June 2026. The fund's 13F-reported value rises from 254.8 million dollars at the end of 2024 to 1.01 billion in March 2025, 2.12 billion in June 2025, 4.14 billion in September 2025, 5.52 billion in December 2025, 13.68 billion in March 2026, and 20.24 billion in June 2026, a roughly 80-fold increase in six quarters. A side panel notes what press reporting says happened after the last filing: total fund assets, a broader figure that includes leverage and private holdings, were reported near 45 billion dollars on 1 July 2026, near 10 billion dollars on 30 July after a forced sale, and that the fund was buying options again by 11 September. SITUATIONAL AWARENESS LP · DISCLOSED US EQUITY VALUE · SEC FORM 13F, LOG SCALE $1B $10B Dec '24 Jun '25 Dec '25 Jun '26 31 Dec 2024: $254.8 million 31 Mar 2025: $1.01 billion 30 Jun 2025: $2.12 billion 30 Sep 2025: $4.14 billion 31 Dec 2025: $5.52 billion 31 Mar 2026: $13.68 billion 30 Jun 2026: $20.24 billion, last filing before the sale $0.25B $20.24B Reported after the last filing: 1 Jul 2026: total fund assets near $45B (not a 13F figure) 30 Jul 2026: near $10B after a forced sale to Citadel 11 Sep 2026: buying options again, same four categories 13F covers only US-listed long equity and options. It excludes leverage, short positions, and private stakes such as Anthropic. Quarterly totals are the writer's sums of each filing's value column. Side-panel figures are as reported, cited below.
The line is what the fund told the SEC it owned in US stocks and options, each quarter, on the record. The side panel is what happened three, four and ten weeks after the last of those filings.

The first filing, for the end of 2024, held six positions worth $254,813,765: Constellation Energy, Talen Energy and Vistra, three power generators, plus Vertiv, Marvell and Modine. That is not a stock picker’s list. It is the essay’s power chapter, bought. By June 2026 the fund’s disclosed US equities were worth $20,242,292,228, and the largest names were SanDisk and Micron, memory chipmakers, at $5.67 billion and $5.57 billion apiece, just over 55% of the whole disclosed portfolio between them, by my own sum of the filing’s line items against its reported total (Moomoo Whale Watch put it the same way; Quartz rounds it to “more than 56%”). Between those two filings sat CoreWeave, Applied Digital, Core Scientific and IREN, the data-center and compute-hosting names, and Bloom Energy, back for the power leg. Read across six quarters, the 13F filings are close to a literal holdings-list translation of the essay’s four bottlenecks: chips, power, buildings, memory.

a letter, then a wall

The 13F figures are only the part of the fund the SEC makes public: US-listed long stock and options, no leverage, no shorts, no private stakes. By the reporting that follows, the fund’s total assets, the broader number that does include leverage, peaked near $45 billion at the start of July 2026 (CNBC, via multiple outlets; corroborated by TechCrunch and Quartz). On 24 July, in a letter to investors seen by the Financial Times, TechCrunch reports, Aschenbrenner called the month’s AI stock selloff one of the best buying opportunities since early last year. Six days later, the fund was forced to sell.

Reporting after the fact, none of it disputed, describes a fund carrying leverage of up to 400%, meaning roughly four dollars of position for every dollar of capital (Quartz). Its largest longs, SK Hynix, CoreWeave, Nebius Group, Bloom Energy and Micron, fell between 35% and 47% in July (Augment Market). TechCrunch put a fourth name, SanDisk, in the same bracket, reporting that SK Hynix, SanDisk, Bloom Energy and Nebius Group had “all… plummeted by more than 30% over the past month.” (TechCrunch). Three prime brokers, Bank of America, Goldman Sachs and JPMorgan Chase, issued margin calls, and by 30 July the fund had sold most of its public stock book to Ken Griffin’s Citadel, a reported $16 billion of positions at around a 10% discount to their already-depressed prices (Crypto Briefing). The fund’s assets stood at about $10 billion.

What the fund sold to Citadel and what it kept, at the end of July 2026 Two columns. What was sold, priced every second, marked to market and pledged as collateral: about 16 billion dollars of public stock and options in SK Hynix, CoreWeave, Nebius Group, Bloom Energy and Micron, which fell 35 to 47 percent in July, plus SanDisk, which fell more than 30 percent, sold to Citadel at around a 10 percent discount. What was kept, priced only when a deal closes and carrying no margin requirement: a reported 5 billion dollar stake in Anthropic, a stake in the chip startup MatX, and a stake in the data-center startup Fluidstack, then in talks near an 18 billion dollar valuation. WHAT MOVED AND WHAT DIDN'T · END OF JULY 2026 · PRESS REPORTING Sold to Citadel, ~$16B PUBLIC · LEVERED · MARKED EVERY SECOND Kept PRIVATE · UNLEVERED · MARKED AT LAST ROUND SK Hynix, CoreWeave, Nebius Group, Bloom Energy, Micron: down 35 to 47% in July SanDisk: down more than 30%, also sold Pledged as collateral; a margin call can reach a price that updates now Anthropic stake, reported ~$5B MatX, a chip-design startup Fluidstack, near an $18B round No margin call can reach a price that only updates when a deal closes Figures as reported: Crypto Briefing (3 Aug 2026), TechCrunch and Augment Market (30 to 31 Jul 2026). Not an official balance sheet.
The line the fire sale drew was not chips versus power versus memory. It was priced-every-second versus priced-when-a-round-closes.

Augment Market’s Paul Smalera drew that distinction to explain why the Anthropic stake survived: “A price that updates every second reflects the marginal seller under margin pressure. A price that updates when a round closes reflects the last negotiated transaction.” The public names were pledged as collateral against borrowed money, so a falling market price forced a sale at that price. The private stakes were not pledged the same way, so nothing forced a sale of them, whatever their underlying value.

the thesis wasn’t the problem

On 25 August, the SEC subpoenaed four of the fund’s banks, Goldman Sachs, JPMorgan Chase, Citigroup and Bank of America, for records of trade timing and communications about the fund’s borrowing (CNBC). No wrongdoing has been alleged, and CNBC reported the inquiry was at an early stage. Days before that, on 5 and 6 August, the fund had already put $400 million into a private company, on top of $100 million committed a month earlier, bringing its stake in that one holding to $500 million. The company was not named as of the most recent reporting cited here; Bloomberg reported that a Sequoia Capital partner, Alfred Lin, confirmed Sequoia backed it, telling reporters Aschenbrenner had “wire[d] $400 million to a company that we invested in” without naming it further (Crypto Briefing). Aschenbrenner told investors: “We took the steps that were necessary to fight another day” (Investing.com, via Yahoo Finance).

By 11 September, per CNBC and Quartz, the fund was buying options again, in AMD, Bloom Energy, CoreWeave, SK Hynix, SanDisk and the Roundhill Memory ETF, the same four categories as before: compute, power, infrastructure, memory. Quartz’s Cris Tolomia noted that whether the money behind those trades survived the summer or came from new investors “remains unknown.”

Aswath Damodaran, an NYU finance professor who writes regularly about corporate and fund blowups, put the diagnosis plainly on 11 August: leverage was the cause, not the thesis. He wrote that he has “no issues” with Aschenbrenner’s underlying view of AI, and that a concentrated bet on a real trend, sized to survive its own volatility, “can deliver significant returns.” What the fund added was leverage on top of an already-concentrated, already-volatile bet, which turns an ordinary bad month into a forced one. Tae Kim made the same point the day of the sale: the fund’s individual picks were, on his accounting, up more than 100% on average even after the crash, and, the same day, several of the names Citadel had just bought were reported to be up 15% to 25% on news of the fund’s own liquidation.

I don’t think the essay was wrong about where the money would go. Chips, power, data centers and memory are exactly where two more years of AI capital went, and the fund’s own portfolio traces that path quarter by quarter. In an earlier essay, I wrote about a different kind of accounting fragility built into that same buildout: Microsoft’s 29 July filing that its buildings would now depreciate over 25 years instead of five to 15. That fragility lives in how the spending is booked. This one lived in how the bet on the spending was financed. The essay named the bottleneck as physical: chips, power, buildings, wire. It said nothing about margin, because margin is not a property of the AI buildout. It is a property of betting on it with someone else’s money.

References

  1. Aschenbrenner, L. Situational Awareness: The Decade Ahead. Self-published, June 2024. Trillion-dollar cluster figure from Racing to the Trillion-Dollar Cluster.
  2. Wikipedia. Leopold Aschenbrenner. Accessed 27 September 2026.
  3. Situational Awareness LP. Form 13F-HR filings, SEC EDGAR, CIK 0002045724, for the periods 31 December 2024 through 30 June 2026. Filing index. Quarterly totals are each filing’s own reported table value total, from its primary_doc.xml cover page.
  4. DeMatteo, F. Aschenbrenner makes quick return after near-collapse of Situational Awareness. Investing.com, via Yahoo Finance, 6 August 2026.
  5. Crypto Briefing editorial staff. Aschenbrenner’s $400M Bet Went to a Sequoia-Backed Company. 6 August 2026.
  6. Temkin, M. AI hedge fund Situational Awareness may have sold its public portfolio, but it still has its Anthropic shares. TechCrunch, 30 July 2026.
  7. Tolomia, C. Leopold Aschenbrenner’s AI hedge fund collapses after margin calls. Quartz, 31 July 2026.
  8. Smalera, P. 45B to 10B: The Anthropic Stake Stayed. Augment Market, 31 July 2026.
  9. Crypto Briefing editorial staff. Situational Awareness loses 67% in July, sells $16B public portfolio to Citadel amid margin call chaos. 3 August 2026.
  10. Damodaran, A. The Situational Awareness Fund Blow-up: Collateral Damage from Investment Conviction! 11 August 2026.
  11. Kim, T. The Big Lesson from the Implosion of Leopold Aschenbrenner’s $20 Billion Situational Awareness Hedge Fund. 30 July 2026.
  12. CNBC. SEC reportedly subpoenas Wall Street banks over AI hedge fund Situational Awareness’s near collapse. 25 August 2026.
  13. Tolomia, C. Leopold Aschenbrenner’s Situational Awareness buying options after collapse. Quartz, 11 September 2026. Citing CNBC’s reporting of the trades.
  14. Moomoo Whale Watch. 13F Tracker: Before the Unwind, Situational Awareness’s Q2 Filing Reveals Peak AI Concentration. 15 August 2026.