A $35 Billion Bet That AGI Could Not Wait

ColdFusion reconstructs how Leopold Aschenbrenner built a $45 billion hedge fund on an AGI thesis and lost about $35 billion of it in two weeks, during his own wedding week in Carmel, California.

It is not an AI forecast story. It is a leverage story.

The dispatch opens on the villa and the margin calls. Aschenbrenner, 24, had started Situational Awareness LP less than two years earlier with $225 million, named for his 165-page June 2024 essay that argued AGI would arrive by 2027. That PDF went viral in Silicon Valley after OpenAI fired him in April 2024 from its superalignment team, and podcaster Tim Ferriss branded him the Nostradamus of AI. Money followed the narrative, including Stripe founders Patrick and John Collison, former GitHub CEO Nat Friedman and investor Daniel Gross, even though the firm had eight employees and its founder had no professional trading experience.

His book was simple and concentrated. According to ColdFusion, he bought the physical stack of the AI boom, chip makers, memory and power and cloud capacity, names like SK Hynix, CoreWeave and Nebius and SanDisk, and shorted software incumbents like Adobe and Salesforce that he thought AI would hollow out. Shorting means you profit if the stock falls. ColdFusion frames the portfolio as a hedge on paper that was not a hedge at all, because both sides rode the same conviction that AI would upend everything at once and on his timeline.

Wall Street asked the obvious question Silicon Valley waved past. What if you are early. Blackstone met him and passed. One New York investor asked what happens if adoption slows or AGI takes longer. ColdFusion says Aschenbrenner had no detailed answer beyond belief in the thesis. That belief was then multiplied. The fund ran roughly 400 percent leverage, about four dollars of exposure for every dollar of equity, with options adding more juice. Prime brokers Goldman Sachs, JPMorgan Chase and Bank of America financed it.

For a year the stack worked, until July 2026 did not. ColdFusion points to whispers that capable Chinese open-weight models could do most AI tasks for a fraction of the cost, spooking U.S. markets. The fund's six biggest longs fell 50 to 78 percent from recent peaks by month end. The shorts rose. Both legs lost together. With that much leverage, a single holding falling 57 percent, as SanDisk did in under a week, can erase the equity behind the position entirely.

Lenders knocked. Banks demanded more collateral. When a fund cannot post it, the broker sells. By July 30, about $16 billion of stock was dumped in one overnight auction at more than 10 percent below market, the biggest slice of the $35 billion drawdown in two weeks, bought by Ken Griffin's Citadel. What had been $45 billion was left at about $10 billion. The only reason the fund survived, ColdFusion notes, was an estimated $5 billion illiquid stake in Anthropic, where Aschenbrenner's now wife Avital Balwit is chief of staff to the CEO. Citadel had offered to take the whole public book instead. The sale raised cash, the private shares stayed.

The skeptic's ledger is what ColdFusion leaves unproven. There is no disclosure of the unnamed private company that received $100 million a month earlier and another $400 million days after the Citadel sale, no audited track record behind the claimed more than 1,000 percent return, and no evidence that AGI by 2027 is priced wrong rather than the fund being structurally wrong. The investor letter calling the crash ructions and inviting fresh money from August 1 reads as conviction, but it is also a capital call while collateral was being liquidated. JPMorgan cut the fund off after the losses, a hard constraint ColdFusion flags at the end, and that matters for any rerun.

That rerun is already being attempted. CNBC reported Sept. 11 that Situational Awareness has been buying call options again, in Advanced Micro Devices, Bloom Energy and CoreWeave and SK Hynix and SanDisk, after the forced sale, and has tapped specialist brokerage Clear Street as JPMorgan ended its prime relationship. A separate report said the fund lost about 67 percent in July and remains open, still up roughly 80 percent for the year after earlier gains.

Cliff Asness, the hedge fund manager with three decades of experience, gets the closing lesson in ColdFusion: you never go full Kelly. The Kelly rule tells you how much to bet given an edge, but it is only as good as your assumptions. Leopold's life was hedged as a joke, human capital depreciates or financial capital grows, but the fund was not hedged at all. Silicon Valley may forgive the miss on timing. Leverage does not.