Leopold Aschenbrenner's $45 billion AI fund collapsed in days, signaling liquidity mismatch risk for mega-AI-infrastructure deals.

At the end of July 2026, one of the most closely watched funds in global markets lost roughly three-quarters of its assets in a matter of days. Situational Awareness, the artificial intelligence fund founded by former OpenAI researcher Leopold Aschenbrenner, was forced to sell its entire public equity book to Ken Griffin's Citadel at a discount after prime brokers issued margin calls it could not meet.

The story matters to crypto readers for reasons that extend well beyond schadenfreude at a leveraged blow-up in another asset class. The collapse ran on mechanics that anyone who traded through 2022 will recognize instantly, it involves a figure whose career began inside the FTX orbit, and it arrived in the same weeks that Bitcoin quietly broke its correlation with the AI trade. Several crypto-native companies are now carrying AI infrastructure risk directly on their balance sheets.

Aschenbrenner is a German investor and former AI researcher, born in 2001 or 2002 to parents who were both doctors, educated at the John F. Kennedy School in Berlin. He entered Columbia University at 15 and graduated as valedictorian in 2021 at age 19, with degrees in economics and mathematics-statistics.

His early career included a stint at the FTX Future Fund, the philanthropic arm of Sam Bankman-Fried's exchange, where he helped run a charitable operation from the Bahamas. He then joined OpenAI's Superalignment team, tasked with controlling systems more capable than humans. OpenAI dismissed him in April 2024 over an alleged information leak. Aschenbrenner disputes that account, saying he shared a largely non-confidential planning document with outside researchers for feedback and that his dismissal followed tensions over warnings he had raised about the company's security practices. OpenAI has said those concerns were unrelated to his departure.

In June 2024, Aschenbrenner published "Situational Awareness: The Decade Ahead," a 165-page essay arguing that artificial general intelligence was arriving faster than almost anyone understood, and that the resulting demand for compute, energy, and hardware would be historic. The essay became required reading across Silicon Valley. The following month, he turned it into a fund of the same name.

The trade was the essay. If AI capability scaling continued, then semiconductors, memory, data centers, and power infrastructure were the bottleneck, and owning that bottleneck with leverage was the highest-conviction expression of the thesis. Backers included Stripe co-founders Patrick and John Collison, former GitHub CEO Nat Friedman, investor Daniel Gross, and Jane Street. The Wall Street Journal reported gains exceeding 1,000% since inception.

Reported peak assets vary by source. CNBC put the fund's high-water mark at around $45 billion, while other reporting cited roughly $20 billion in assets under management at peak. Either figure represents extraordinary capital for a manager who had never run money before founding the fund at 22.

The unwind ran over roughly two weeks in late July. The fund's concentrated positions in AI infrastructure names—reported to include SK Hynix, CoreWeave, Nebius, Micron, and Bloom Energy—fell between 35% and 47% during the month. The Philadelphia Semiconductor Index dropped 28.6% from its June 22 peak as investors began questioning whether hyperscaler capital expenditure could ever generate adequate returns. A separate short position against software stocks reportedly went against the fund at the same time, compounding damage from both directions.

Then the leverage did what leverage does. Reports put the fund's gearing at as much as 400%. At four times leverage, a 25% decline in underlying positions is mathematically sufficient to erase an investor's entire equity contribution. The positions fell considerably further than 25%.

Prime brokers Goldman Sachs, J.P. Morgan, and Bank of America issued margin calls. The fund attempted several escape routes: a capital raise letter to existing investors, discussions with lenders, and negotiations with Millennium Management and Jane Street Group. According to the Financial Times, all failed. Citadel stepped in and bought the entire public book at a discount.

Assets fell from roughly $45 billion to around $10 billion. Reporting since suggests the fund may still be forced to liquidate further holdings.

Once Citadel had absorbed the position, the Nasdaq gained 3.30% and the semiconductor index rose sharply. Much of the late-July decline in AI infrastructure names had been the market pricing in a large, visible, forced seller. Removing that seller removed the discount.

The timing was unusual in one further respect: Aschenbrenner married Avital Balwit, chief of staff to Anthropic CEO Dario Amodei, in California the same weekend the fund was being unwound.

**Why This Matters**

The first reason is that this is a familiar story with different tickers. A young quantitatively gifted manager builds a totalizing thesis about the future, expresses it through extreme concentration and heavy leverage, produces spectacular returns that attract enormous capital, and then discovers that leverage is symmetrical. Crypto has run this experiment repeatedly. The specific detail that closes the circle is that Aschenbrenner's first significant job was at the FTX Future Fund, and that Jane Street, where Bankman-Fried himself trained, appears in this story both as an investor and as a failed rescue counterparty.

The comparison should not be pushed too far. There is no allegation of fraud, no customer funds, no missing assets. Situational Awareness appears to have been a legitimate fund that took a directional view and lost, which is categorically different from what happened at FTX. But the underlying behavioral pattern—narrative conviction plus leverage minus risk management—is the same one that has cost crypto investors more money than any hack.

The second reason is that the mechanics are identical to a liquidation cascade. Concentrated leveraged longs, a price decline, a margin call, a forced seller who must sell into a falling market, and a well-capitalized buyer waiting to take the other side at a discount. Crypto traders watch this happen on-chain and on exchange liquidation feeds constantly. On July 13, when the Kospi fell 8.95% and SK Hynix dropped 15.37% in its worst session on record, $253 million in leveraged crypto positions were force-liquidated in parallel, with long positions accounting for 76% of the total. Same physics, different venue.

The third, and most consequential, reason is what crypto did not do. For most of 2026, crypto traded as a high-beta expression of the AI trade. It rose when chip stocks rallied and fell when they slipped. That relationship broke in July, and it broke twice in five sessions.

When roughly $797 billion came off the largest U.S. technology stocks in a single Thursday session in late July, Bitcoin barely moved. On July 29, as Asian equities suffered one of their worst two-day stretches of the year and SK Hynix fell nearly a fifth despite growing quarterly profit more than sixfold, Bitcoin rose about 1% to $63,800. Ether added 1% to $1,899, XRP gained 2% to $1.07, and Solana held around $73. When Citadel absorbed the Situational Awareness book and AI infrastructure names rebounded sharply, crypto markets were largely unmoved in the other direction as well.

Across July as a whole, Ether gained 16.29% and Bitcoin 5.61%, while the AI infrastructure complex was being repriced downward.

One reading of this decoupling is that Bitcoin is regaining independence as an asset class, driven now by rate expectations, ETF flows, and its own regulatory calendar rather than by sentiment toward Nvidia's supply chain. Analysts increasingly describe crypto as behaving like a liquidity sponge, expanding and contracting with global money supply and real rates rather than with any individual equity narrative. Research has attributed roughly 45% of weekly Bitcoin price movement in 2026 to ETF flows alone.

A more cautious reading is that two weeks does not constitute a trend, and that decoupling claims have been made and abandoned repeatedly since 2020. The honest position is that the correlation has weakened materially and visibly.