Leopold Aschenbrenner became well known in the AI community after leaving OpenAI and publishing Situational Awareness, a 165-page argument that AI could reach human-level capabilities by 2027. He predicted an extraordinary expansion in data centres, energy consumption and national investment as the US and China competed to control advanced AI.
Aschenbrenner’s original paper predicted that physical compute would remain important while efficiency gains and agent software supplied an increasing share of AI progress. That forecast may prove more accurate than the fund’s concentration on infrastructure and its bets against software.
From Aschenbrenner’s Situational Awareness report.
When we first wrote about Aschenbrenner, he had decided to turn his analysis into an investment fund of the same name. Situational Awareness would place financial bets on the world described in his paper. Two years later, CNBC reported that the fund had returned 439% net in the first half of 2026 and that its assets had peaked at about $45 billion.
It then reported an unaudited 67% net loss for July.
Situational Awareness invested heavily in the physical infrastructure behind AI. Its reported holdings included SK Hynix and Micron, which manufacture memory chips; Sandisk, which provides storage; CoreWeave and Nebius, which operate AI cloud infrastructure; and Bloom Energy, whose fuel cells can supply electricity to data centres.
These companies were spread across South Korea, Europe and the US, but they depended on the same economic forecast. Microsoft, Amazon, Google, Meta and OpenAI would continue spending enormous amounts on AI computing. That would sustain demand for chips, memory, storage, cloud capacity and electricity.
At their July lows, Nebius, CoreWeave, Micron and Sandisk had each fallen more than 35% from late-June levels. SK Hynix lost 14.65% on 28 July alone. Several had fallen between 50% and 78% from recent peaks.
This was partly a normal cyclical correction. AI infrastructure shares had risen faster than their current profits, while memory manufacturers were investing in new capacity that could eventually turn shortages into surpluses. Higher interest rates and geopolitical tensions added pressure.
But the movement was also directly related to AI’s economics. Reports that Nvidia was in talks to guarantee roughly $250 billion of lease and construction financing for an OpenAI data-centre project raised questions about the quality of demand. If suppliers must finance their customers, investors need to ask whether the customers can produce enough revenue to pay for the equipment.
CNBC reported that Situational Awareness had shorted established software companies, including Adobe. A short seller borrows shares, sells them and hopes to buy them back later at a lower price. The trade reflected the broader thesis that AI agents could replace functions sold by established software companies.
Instead, software shares recovered. Adobe gained about 22% in July, while Salesforce rose about 17% and ServiceNow about 12%, helped by evidence that established vendors could sell AI to their existing customers rather than simply be replaced by it.
The infrastructure holdings and software shorts lost money together. This exposed a weakness in the supposed hedge. Both positions depended on the same prediction about the timing of AI adoption. Infrastructure spending had to grow rapidly while software businesses deteriorated. The price action suggested that infrastructure valuations had moved too far while software disruption might take longer than expected.
Leverage turned these losses into a crisis.
Suppose a fund has $10 billion of investor equity supporting $40 billion of positions. Its exposure is four times its equity. If it loses $4 billion, its equity falls to $6 billion. Even if its positions also shrink to $36 billion, its leverage has increased from four times to six times. The fund has not borrowed more. Its leverage has risen because the capital protecting its lenders has contracted.
Prime brokers provide hedge funds with loans, share borrowing and trading services. They continuously compare the fund’s debts with the value of its positions. As losses reduce the equity cushion, they can demand more collateral through a margin call.
If the fund cannot provide the cash, it must sell. Selling falling shares pushes prices lower, while closing shorts requires buying shares that are rising. Further losses produce further margin calls.
Warren Buffett once attributed a line to Charlie Munger: “There are only three ways a smart person can go broke: liquor, ladies and leverage.” Buffett explained that the first two were included for the alliteration. The serious answer was leverage.
Leverage does more than multiply gains and losses. It gives lenders control over time. Aschenbrenner could believe that his investments would recover, but his prime brokers did not have to wait.
According to the FT, Situational Awareness sold most of its roughly $16 billion public portfolio to Citadel at a discount of more than 10%. The discount gave Citadel a cushion against further volatility.
“We came closer to permanent capital impairment than is acceptable to us.”
After the sale, several of the affected shares rebounded sharply. Citadel’s main Wellington fund returned 5.9% in July, its best month since 2022. Their different outcomes reflected how long each could afford to wait, not necessarily different views of AI.
Reuters reported that Situational Awareness retained a roughly $10 billion book comprising public securities and private investments, including Anthropic. Its unaudited estimate still showed an 80% net gain for the year. The fund said it had removed leverage from its public portfolio, then invested another $400 million in a Sequoia-backed private company.
Takeaways: This was not a verdict against AI progress. It showed that AI infrastructure, software adoption and financial markets operate on different schedules. Investors are starting to distinguish real demand from financed demand, technical importance from shareholder returns, and eventual software disruption from current business performance. Situational Awareness tried to finance a multiyear AI forecast with money that could be withdrawn in a day. Citadel bought the assets after that financing failed.