By Avery J. Fahrman
Leopold Aschenbrenner, who graduated from Columbia University as valedictorian at age 19, has written a 165-page essay titled “Situational Awareness: The Decade Ahead” in which he posits that the battle for Artificial General Intelligence (AGI) would likely be determined by physical factors (the availability of microchips, the supply of electric power and the data centers where they are housed) rather than by computer algorithms. He founded a hedge fund based upon this premise shortly thereafter. By early March 2026, Situational Awareness LP was being widely followed and investing in power generation firms, manufacturers of chips and companies providing data center services instead of those developing AI models.
For some time the investment strategy seemed valid. After accounting for fees, the fund generated returns of over 1,000% since inception. In the first six months of 2026 alone, the fund generated a return of 439%. By early July 2026, Aschenbrenner’s fund had attracted an estimated $45 billion.
Nearly all of these assets evaporated within a week. The margin calls from Goldman Sachs, J.P. Morgan and Bank of America required Aschenbrenner to sell his publicly traded portfolio to Citadel at a discount on 30 July. Consequently, Aschenbrenner’s remaining assets were reduced to approximately $10 billion and consisted primarily of a private stake in Anthropic. Nothing occurred during July that contradicted Aschenbrenner’s thesis; compute and power may prove to be the scarcest inputs of the decade. However, July demonstrated another principle: that portfolios developed with a ten-year perspective can fail in a single poor month. And they did.
Six Days in July
Aschenbrenner’s fund operated at roughly four times leverage on its publicly traded positions. As such, there are relatively simple mathematical implications. A decline of 30% in Aschenbrenner’s long positions destroys roughly 120% of its capital prior to the application of hedges. Moreover, a 25% decline would completely eliminate the capital.
The decline occurred between 24 and 29 July, when investors suddenly reversed course regarding AI-related infrastructure. The Nasdaq 100 declined by over 10%, with fear and panic throughout Korea’s chipmaking industry. Four of Aschenbrenner’s core holdings, Nebius, Sandisk, Micron and CoreWeave, experienced declines greater than 35% for the month.
Afterward, Aschenbrenner’s thesis became irrelevant. Margin clerks don’t read essays. On the evening of 29 July, Aschenbrenner began seeking bids for his publicly traded positions, and Millennium and Jane Street were reportedly bidding. Before trading opened the next day, Citadel acquired virtually the entirety of Aschenbrenner’s publicly traded positions through a single block trade.
Following Citadel’s acquisition, the stocks that Aschenbrenner sold experienced significant increases in price, ranging from 20% to 27%: Nebius increased by 27%, IREN by 26.5% and Bloom Energy by almost 26%. There were no changes to any of these companies that contributed to their increased prices. The only change that took place was ownership of shares.
Right Thesis, Wrong Vehicle
There is little doubt that Aschenbrenner is intelligent; otherwise the collapse would be less informative. The collapse provides insight into how Aschenbrenner chose to convert his thesis into a portfolio. A ten-year thesis requires capital that can remain invested for ten years, whereas leveraging his capital at a four-to-one level provided him with a time horizon measured in weeks. Several characteristics of his book exacerbated this problem.
One Bet Disguised as a Portfolio
Although Aschenbrenner’s fund appeared diversified on paper across power generation, semiconductors, cloud computing and bitcoin mining, in reality each of his positions was valued relative to how quickly or slowly the market anticipated growth in spending for AI-related technology. His quarterly disclosure statement revealed that 76% of his disclosed long exposure was in five positions: Bloom Energy, Sandisk, CoreWeave, IREN and Core Scientific, with an 8.2% interest in Core Scientific. Therefore, when investor sentiment toward expected AI spending growth shifted negatively, all five positions declined simultaneously.
His hedges offered little comfort either. In addition to purchasing put options protecting approximately $3 billion in a semiconductor exchange-traded fund and $2.6 billion in Nvidia, he also established short positions in legacy software companies (such as Adobe). However, hedging illiquid stocks using the most liquid names available in the market left a gap in coverage that ultimately proved costly when the smaller names declined further than their larger counterparts in the midst of a panicked sell-off.
Sell What You Can, Not What You Should
A third of the portfolio was made up of private investments (most of them Anthropic). Private positions cannot be margin-called, but neither can they be quickly sold. All of the pressure then rested with the public book, the portion of the portfolio that had already fallen.
The Illiquidity That Saved the Firm
Situational Awareness’s ownership in Anthropic preserved approximately $5 billion of equity in a company that had a valuation of $965 billion at the time and is expected to go public. Within days, the fund completed a $400 million investment. It is difficult to miss the irony. In March, Situational Awareness was said to avoid investing in AI model builders. By August, one such model builder was the largest piece of equity in the firm.
The Bull Case
The strongest defense for Situational Awareness is that it was simply early, and that the subsequent rally after Citadel’s purchase vindicated their call. Investors who decided to sit out 1999 were correct about Pets.com and incorrect about Amazon.
This argument is valid; however, it addresses a completely different question. Simply knowing the direction of travel does not tell you anything about how big your position should be. After peaking during the dot-com bubble, Amazon stock declined by more than 90%. The individuals who accumulated wealth from owning shares of Amazon stock held onto those shares throughout that period without using borrowed money. An individual utilizing four times leverage would have been wiped out by the first 25% decline in Amazon stock and would not have seen the subsequent rebound.
Temperance is a fundamental aspect of all successful investments and does not relate to forecasting events. None of these concepts depend upon when we arrive at AGI. A position that cannot withstand a 35% drop in price within a single month will not last long enough for a decade-long thesis to have any chance of unfolding. Leverage should always be based upon the worst conceivable month, not an assumed monthly average. Diversification is not a guarantee against loss. Five publicly traded stocks that increase and decrease in response to the exact same assumptions are essentially one investment.
Aschenbrenner may ultimately prove correct, and Anthropic may eventually turn Situational Awareness into a significant amount of wealth. However, July resolved a far smaller issue: there exists a difference between being correct about AI and generating profits from it.
The views expressed in this article are the author’s own and may not reflect the opinions of The St Andrews Economist.
Image Credit: Unsplash