Leopold Aschenbrenner: The $45B AI Hedge Fund That Fell 67%
8 min readUpdated 22 August 2026
Loading...
8 min readUpdated 22 August 2026
On 24 July 2026, Leopold Aschenbrenner wrote to his investors. The selloff underway, he told them, was creating some of the most attractive opportunities since early 2025.
Six days later he sold his entire public stock portfolio to Citadel.
His next letter read very differently. "We let you down this month."
Leopold Aschenbrenner was born in Germany to a family of medical professionals. He attended the John F. Kennedy School in Berlin before going to Columbia University.
He graduated in 2021 as valedictorian, aged 19, with a degree in mathematics, statistics and economics.
As an undergraduate he co-founded Columbia's Effective Altruism chapter, a movement focused on long-term risk and existential threat. That led to a research role at Oxford's Global Priorities Institute, working on long-run economic growth models.
In early 2022 he joined the team running the FTX Future Fund, a philanthropic vehicle funded by the FTX Foundation. He resigned in November 2022, immediately before FTX collapsed and filed for bankruptcy.
In 2023 OpenAI recruited him for its newly formed Superalignment team, led jointly by Ilya Sutskever and Jan Leike. The team was allocated 20% of the company's total compute, with the job of controlling AI systems far smarter than their human supervisors.
There he co-authored "Weak-to-Strong Generalization", a paper on how weaker AI models can supervise more capable ones. It was presented at ICML 2024.
Then came a breach. An intruder got into OpenAI's internal communication systems and extracted details of model design. OpenAI chose not to publicise it widely.
Aschenbrenner concluded the lab's security was not strong enough to resist foreign intelligence operations, specifically Chinese state-sponsored ones. He wrote a security memo directly to the Board, comparing model weights to nuclear secrets in the Manhattan Project era. He received a formal reprimand from HR.
In April 2024 he was fired. OpenAI cited a leak of non-public information. Aschenbrenner disputed that firmly, saying what he shared was an unclassified brainstorming document sent to three outside academics for feedback, and that his dismissal was connected to the security memo.
Within a month of his departure, Sutskever and Leike both resigned. OpenAI dissolved the Superalignment team.
In June 2024 he published "Situational Awareness: The Decade Ahead", running to 165 pages.
The core argument rests on three trends compounding at once: raw compute, algorithmic efficiency, and what he calls unhobbling, meaning the conversion of static chatbots into autonomous agents that can work over long horizons.
His conclusion: AGI around 2027 is strikingly plausible, followed by an intelligence explosion once AI begins automating AI research itself.
The part that moved capital markets was not the philosophy but the physics. He argued the binding constraint on superintelligence is not algorithmic ideas but very concrete things: semiconductor fabrication, high-bandwidth memory, giant datacenter construction, and above all electrical power.
The essay became required reading in Silicon Valley. Not everyone agreed. Yann LeCun, Meta's Chief AI Scientist, rejected the premise, arguing that scaling language models alone cannot produce general intelligence because token prediction lacks world models and causal reasoning.
In late 2024, aged 22, Aschenbrenner founded Situational Awareness LP. He had no prior track record managing money professionally.
Seed capital of roughly $225 million came from well-known names in the technology ecosystem: Patrick and John Collison of Stripe, Nat Friedman, and Daniel Gross.
The strategy was his essay translated into positions:
Long the physical suppliers of AI infrastructure. Public filings as of 31 March showed holdings including Nebius, Bloom Energy, Sandisk, CoreWeave, SharonAI and IREN.
Short legacy software companies seen as vulnerable to agentic AI.
The results were extraordinary.
| Measure | Figure |
|---|---|
| First half 2026 return | 439% |
| Return since inception through June 2026 | roughly 1,551% |
| Seed capital | $225 million |
| Peak assets, early July 2026 | $45 billion |
From $225 million to $45 billion in under two years.
In mid-July, institutional investors began taking profits across AI infrastructure stocks.
The problem was that this portfolio lost on both sides at once. The infrastructure names fell. According to CNBC, holdings on its filed list had dropped between 50% and 78% from recent peaks. At the same time the software names it was short rallied.
When you hold longs and shorts and both move against you, no part of the portfolio rescues any other part.
In total, the portfolio fell 67% during July alone.
When prices fell, margin calls followed: lenders demanding additional collateral. If it is not posted, the collateral gets sold. Forced selling pushes prices lower, which triggers the next margin call.
Aschenbrenner described this in his own letter as dynamics resembling a bank run, where vulnerability breeds further vulnerability.
On 24 July, Aschenbrenner wrote to investors framing the fall as a buying opportunity, calling it among the most attractive since early 2025, and asked for additional capital.
His investors declined to put money into a leveraged structure under active margin calls.
On 30 July 2026, Ken Griffin's Citadel bought the entire Situational Awareness public equity book in a single block trade. It ranks among the fastest and largest single-fund liquidations in modern financial history.
His next letter, first reported by the Wall Street Journal, struck a very different tone:
"We let you down this month."
"We came closer to permanent capital impairment than is acceptable to us."
This is the part that headlines tend to skip, and it directly answers what happened to the public money.
What was sold was the public equity book. Assets under management fell from around $45 billion to roughly $10 billion.
What was not sold is the private side, principally a stake in Anthropic worth roughly $5 billion. Precisely because that asset is illiquid, no lender could force it out.
Going forward, the fund is reported to be operating as a private investment vehicle.
And here is the strangest part: because the early-year gains were so large, the fund is reported to be still up roughly 80% for 2026 through the end of July, even after a 67% month.
Leverage does not add return, it multiplies outcomes, in both directions. Up 439% in six months and down 67% in one is not two stories. It is one machine running both ways.
The illiquid asset survived because it was illiquid. Public equities are repriced daily, so they can be margin-called and force-sold. The Anthropic stake could not be. The hardest part of the portfolio to sell is the part that saved it.
What kills you is not the loss, it is being forced to sell. An investor with no debt who is down 50% still gets to choose to wait. A leveraged investor does not. The lender decides when they sell, and it is usually at the worst price.
This is exactly why maximum drawdown matters more than peak return, why volatility should be read next to the Sharpe ratio, and why concentration in one theme is riskier than your ticker count suggests.
Did Situational Awareness go bankrupt? No. The fund did not close. Its public equity book was sold and its leverage repaid, but it continues to operate with roughly $10 billion in assets, principally the Anthropic stake.
Is any wrongdoing alleged? No. On the available reporting this was a trading loss and a forced deleveraging, not an allegation of fraud or any criminal conduct.
How old is Leopold Aschenbrenner? He was born around 2001 to 2002, making him roughly 24. He founded the fund at 22.
Did investors lose everything? No. Because the earlier gains were so large, the fund is reported to remain up roughly 80% for 2026 through the end of July, despite falling 67% in a single month.
Open the Portfolio Health section of your dashboard and look at three figures together: volatility, maximum drawdown and Sharpe ratio.
Then run the honest test. Picture your largest position falling 50%. With no borrowing, that hurts but you can wait it out. If you use margin, work out the point at which you would be forced to sell, and recognise that the point is set by your lender, not by you.
Concentration shows up on the Holdings page in the composition bar for each sleeve. If one theme dominates the chart, you own fewer independent bets than your ticker count suggests.