On July 24, Leopold Aschenbrenner, the post-2000s AI stock guru, wrote in a letter to investors: "If you've been waiting for the right time to add positions, now might be one." Six days later, he cleared his positions. Forced to.
According to reports, Leopold Aschenbrenner runs a fund called Situational Awareness. He dumped roughly two-thirds of the fund's total U.S. equity holdings in one fell swoop. The fund's total loss is said to be around $20 billion, equivalent to 140 billion yuan. The buyer was another large hedge fund, whose name was not disclosed; market rumors point to Citadel, with Millennium also participating in the bidding.
He took less than two years to go from $200 million to $24 billion. From $24 billion to liquidation took even less time—just six days.
Aschenbrenner is indeed a remarkable figure. Born in 2001, he was a child prodigy: he entered Columbia University at 15 and graduated with top honors at 19. His resume is colorful, and he has been one of Silicon Valley's brightest and most profitable stars in the past two years—yet there is a persistent air of contradiction about him.
For example, he spent nine months at SBF's FTX Future Fund—yes, the now-collapsed cryptocurrency platform. Its affiliated foundation waved the banner of "effective altruism," and founder SBF often said he made money in order to give it away. Later, FTX imploded, having misappropriated billions of customer funds; SBF was sentenced to 25 years, and he himself called that ethical framework "mostly a front."
After leaving FTX, Leopold joined OpenAI's "Superalignment team." The team's job was to ensure AI does not go out of control—to put the brakes on superintelligence. In April 2024, he wrote a safety memo warning that OpenAI's security was inadequate, and the 23-year-old was fired by OpenAI on grounds of leaking information. A month later, the Superalignment team was also disbanded.
Then Leopold turned around and launched one of the most aggressive AI acceleration funds in the market, betting everything on the AI boom. Even the famous hedge fund Jane Street became one of his backers. The "Superalignment team" was supposed to be the brake-pedal crowd; maybe Leopold was aggrieved, or whatever the reason, but he came out and became the accelerator—and the most aggressive one at that, piling on leverage. Fired by OpenAI at 23, just two years later, at 25, Leopold had become the "AI stock god" who turned $200 million into more than $24 billion; at his peak, AUM reportedly exceeded $40 billion.
When he left OpenAI, he wrote a famous essay predicting that AGI would arrive by 2027. As it turned out, AGI has not arrived, but the market has shifted from reward to punishment. And the whole affair is eerily similar to Bill Hwang's case: Hwang, using leverage to bet heavily on Chinese stocks, was hit by a black swan and forced out. Now, the market is worried about the sky-high investment in AI infrastructure and its ROI; Leopold's positions fell sharply from highs, and he also fell to leverage. Chinese stocks have not really recovered since Bill Hwang; as for today's AI, I suspect no one can yet falsify it. But Leopold's forced liquidation is a landmark event; whether it becomes a watershed for this round of AI market remains to be seen.
1. From Brakes to Accelerator
Leopold Aschenbrenner was born in 2001 into a family of doctors in Germany. He entered Columbia University at 15 and graduated at 19 as the top student in his department, earning three degrees in mathematics, statistics, and economics. Such a start would be dazzling in any era.
During his time at Columbia, Leopold co-founded the campus Effective Altruism (EA) chapter, stepping early into this intellectual circle. From 2021 to 2022, he went to Oxford University's Global Priorities Institute as a researcher. This institution is the core think tank of the EA movement, and there he was able to systematically delve into EA theory.
EA originated with Oxford philosophers William MacAskill and Toby Ord, advocating a data-driven, utilitarian approach to "maximizing good" and encouraging "earning to give." FTX's founder was influenced by MacAskill while studying at MIT, and after joining Jane Street he donated about half his salary; FTX's charitable arm also pledged to give over $160 million to hundreds of nonprofits. Leopold later joined the FTX Future Fund and engaged in EA grant-making.
Leopold worked there for nine months, leaving just before the FTX collapse in November 2022. Later, FTX was found to have misappropriated billions in customer funds; SBF was sentenced to 25 years, and the ethical framework was called by himself "mostly a front." Leopold was not a bystander—he stood right inside that door. This experience, being at the center of a financial fraud, has been described by many as giving him an early "taste of the complexity and risk of markets."
In 2023, he moved to OpenAI and joined the Superalignment team. This team was itself very interesting: driven by the huge public reaction to ChatGPT and internal and external pressure, it was led by OpenAI's key figures Ilya Sutskever and Jan Leike in the summer of 2023. Its goal was to tackle an almost unsolvable problem: "When AI's intelligence surpasses humans, how do we ensure it does not go out of control?" OpenAI at the time committed 20% of its computing power and set a four-year deadline to develop technology for aligning superintelligence by 2027. Core challenges included scalable oversight, goal generalization, internal representations, and adversarial robustness.
The problem, however, is that traditional alignment methods—such as RLHF (reinforcement learning from human feedback)—become unreliable once AI's reasoning ability and scale exceed human range, because humans are no longer dependable supervisors.
Then came the famous OpenAI "coup" at the end of 2023. In any case, Altman won, and accelerationism won. In retrospect, the Superalignment plan looks like a joke.
In April 2024, the turn came. Leopold submitted a safety memo to the board warning that the company's security measures were "extremely inadequate," but OpenAI fired him for "leaking internal information." A month later, Jan Leike left; later Ilya left and founded SSI, and the Superalignment team was dissolved. Those who wanted to brake AI were all ushered out of the driver's seat.
After being fired, in June 2024, he published a 165-page essay, Situational Awareness: The Decade Ahead, predicting that AGI would arrive around 2027, and putting forward concepts like "intelligence explosion" and "superintelligence." The essay went viral in Silicon Valley tech circles and paved the way for his later fundraising.
A person who had been in Superalignment and published a safety essay turned around and went to Wall Street to accelerate AI. "If you can't beat them, join them"—it seems the same everywhere; the smartest minds change fastest. Moreover, according to Fortune, Leopold is engaged to Avital Balwit, chief of staff to Anthropic CEO Dario Amodei. Whether this relationship has anything to do with his changing views on AI is unknown. But the result is clear: Anthropic was the largest single position in his fund.
At the end of 2024, Leopold formally established the Situational Awareness fund. Columbia top graduate, EA believer, FTX disciple, Superalignment researcher—and finally an AI accelerator fund manager. Leopold's resume made a full circle around AI. The next step: raise money.
2. How Did $24 Billion Come About?
Situational Awareness is a hedge fund that bet almost 100% on AI themes. The team was minimal: only 4 investment professionals, with 8 employees in total. Its website had just one sentence: "An investment advisor focused on AGI, founded by Leopold Aschenbrenner."
Looking at the fund's early investor list is like reading a Silicon Valley power roster. Stripe co-founders Patrick and John Collison, former GitHub CEO Nat Friedman, tech investor Daniel Gross, plus quantitative giant Jane Street. These are not ordinary LPs; they are core nodes in the EA movement, effective altruist networks, and tech accelerationist circles. In other words, Leopold's fund was backed by this circle from day one.
The Jane Street thread is especially noteworthy: it is SBF's old employer, and it also bought Anthropic shares from the FTX bankruptcy estate—and Anthropic happened to be the fund's largest single holding, accounting for about one-fifth of the fund.
The fund's strategy was straightforward: concentrated bets on AI infrastructure—computing power, electricity, data centers, storage—while using options and shorts for hedging or amplification. Its prime brokers were Bank of America, Goldman Sachs, and JPMorgan. The 13F filing for the first quarter of 2026 showed that the fund's long positions were all AI infrastructure hard assets: Bloom Energy about $879 million, SanDisk about $724 million, CoreWeave about $556 million, Dutch AI infrastructure supplier Nebius with 12.41 million shares worth about $2.6 billion. It also led a $1.6 billion financing round for Sharon AI, taking a 19.9% stake, and became a cornerstone investor in SK Hynix's Nasdaq ADR. The portfolio also included a batch of Bitcoin miners and data center names: Riot Platforms, CleanSpark, Core Scientific, IREN, and Applied Digital.
In addition, its early investment in Anthropic accounted for about one-fifth of fund assets; it entered at a valuation of about $60 billion in February 2025. At the current valuation of around $1 trillion, that represented an estimated floating profit of about 18 times.
As AI soared, the performance of the Situational Awareness fund became legendary. By the end of June 2026, the fund had generated net returns of about 439% year-to-date, with cumulative returns since inception exceeding 1000%. Its size expanded from $200–300 million at launch to roughly $24 billion by mid-2026. A young man in his early twenties, in two years, had turned a small sum into a behemoth that drew Wall Street's attention.
The key was that Leopold's judgment was indeed accurate. Early on, when the market was frenzied over narratives of "power shortages and computing shortages," he said that software can iterate rapidly with AI, but factories, energy, infrastructure—the physical-world things—cannot be built by AI, and that these would be the most valuable assets in the future. This was the basis for his heavy bet on the AI infrastructure chain.
Research firm Citrini once calculated: if you had invested $100 million when the fund was launched, even after losing 90% in July, that investment would still be worth $230 million. Scale itself is part of the narrative.
On July 24, 2026, facing significant market volatility, he wrote in his investor letter that the fund had "not been spared" by the recent market turmoil, but he called the tech sell-off "the most attractive investment opportunity since early 2025." He added at the end: "If you've been waiting for the right time to add positions, now might be one."
But Leopold did not make it to adding positions; instead, six days later, he liquidated.
3. Reflexivity of the Market
George Soros famously proposed the theory of reflexivity. Roughly, it means that share prices are driven not only by fundamentals, but also by participants' expectations and trading behavior, which in turn alter fundamentals and prices, forming a two-way reinforcing loop. Applied to AI, this was a classic upward reflexive cycle. The long-term AGI narrative created expectations of unlimited demand for computing power; massive capital poured into AI hardware, driving stock prices higher; fund net values surged, constantly validating Leopold's thesis, prompting him to increase leverage and concentrate positions.
Leopold himself firmly believed in the narrative that AGI would eventually arrive, and deployed huge sums to trade, using capital to boost the price action corresponding to that narrative. The problem, however, is that the most lethal feature of the reflexive cycle is that as extreme as the upside is, the downside is equally brutal. The long-term industrial vision can be infinitely romantic, but leverage and valuation realization in the capital market do not wait. The higher the stock price, the lower the patience of speculative capital and the lower the tolerance for error of leveraged capital.
Looking at Leopold's 13F holdings, almost every heavy position was bleeding. SanDisk fell about 46% for the month, the biggest loss. Bloom Energy fell about 33%, CoreWeave about 25%, Nebius at one point fell over 46% and retreated about 43% for the month, SK Hynix ADR fell about 35% from its high, Oracle and AMD each fell about 20%. The Nasdaq 100, heavily weighted with tech stocks, fell about 10% cumulatively for the month, while the Korean Composite Index dropped about one-third.
One of the triggers was Meta's plan to lease AI computing power externally, sparking debate over whether computing power is already in excess. Credit markets also tightened: CoreWeave's CDS spread spiked to about 855 basis points, translating into a 50% five-year default probability under market models; Oracle's CDS also rose significantly.
The real culprit was still leverage. Media reports said Leopold used roughly 4 times leverage at the peak—far less than Bill Hwang, of course, but as stock prices fell rapidly, margin calls were triggered; unable to meet them, forced liquidation was inevitable.
Even more ironically, Leopold's so-called hedges. At the end of the first quarter, the fund's top five positions were all put options—on Nvidia, semiconductor ETFs, Oracle, Broadcom, AMD—with notional values worth billions; among them, the VanEck semiconductor ETF puts alone amounted to about $2 billion, and Nvidia puts about $1.6 billion. But the 13F is just a quarter-end snapshot, not showing strike prices or expiration dates, and no one knows how effective those hedges were in July's decline. What is certain is that his hedges did not work, or did not withstand the extreme drawdown and triggered a liquidity crisis; otherwise, Leopold could have adjusted positions to save himself.
Facing margin call pressure, Leopold tried to raise emergency funds from LPs and lenders while negotiating the transfer of private equity assets like Anthropic to raise cash. But ultimately, self-rescue failed. On July 30, Leopold chose to liquidate all U.S. equity holdings, selling the public market portfolio in a package to another large hedge fund. After the public positions were cleared, the fund was left with only unlisted private assets, making it more like a concentrated growth-equity vehicle rather than the all-in AI hedge fund flagship it once was.
Notably, after Leopold's liquidation, the U.S. semiconductor and memory sectors rebounded sharply: SanDisk rose more than 20%, and Micron and SK Hynix both rose more than 15%. Once Leopold's selling pressure was released in one shot, buyers stepped in to pick up the bloodied chips. This also shows that market disagreement and game-playing are far from over; both longs and shorts are placing real-money bets. Underlying the game is a global capital debate over the valuation of the AI narrative.
For the past two years, the narrative was simple: whoever burns the most cash and expands the most aggressively is worth the most. Now the narrative seems to be shifting: whoever can turn computing power into revenue and profit deserves the high valuation. Of course, Jensen Huang would certainly disagree, but the bearish camp is indeed already quite loud.
Goldman Sachs warned in July 2026 that the AI capital expenditure expansion of hyperscale cloud providers is approaching the upper limit of their financing capacity, and the entire tech sector valuation system is about to be restructured, with market sentiment "like an overstretched rubber band." The Bank for International Settlements listed the AI investment boom as one of four pressure points threatening the global economy, noting that the current expansion bears a strong resemblance to the dot-com bubble: real technological breakthroughs have attracted capital far exceeding the capacity of commercial returns to bear. It specifically warned of "circular financing" risks—chipmakers and cloud providers cross-holding shares and tying long-term procurement—and once returns fall short, capital could contract abruptly.
This is also the view of Benchmark veteran partner Bill Gurley. Of course, neutral and optimistic supporters are not without data or narrative support, but they generally recommend names with pricing power rather than broad bullishness on the market, indicating that overall market disagreement on AI is indeed substantial.
At such a time, Leopold was still adding leverage and betting wildly, presumably to earn a few more pennies—oh, of course, not just pennies, but perhaps several, even tens of billions of dollars.
Still, Leopold named his fund Situational Awareness. When he wrote that 165-page essay, he was convinced that he had broken free of mass prejudice and saw clearly the long-term trend driven by AGI. Over the past two years, Leopold was indeed right. But in the end, this fund, named for "perceiving the situation," could not withstand the market turmoil and liquidated. This is precisely the cruelty of Soros's reflexivity: participants are themselves part of the market action; the narrative you believe in and bet on can lift you up, but it can also overturn you in an instant.
This is not an isolated case. In recent years, the world has been immersed in the grand AI narrative, with computing power demand and long-term AGI visions constantly reinforcing each other. When everyone in the market believes the same story, questioning becomes an expensive luxury. I do not know whether the AI narrative ends here, nor do I want it to be so abrupt. But one thing is certain: the rise was never smooth. Leopold is merely the first to fall at a small inflection point in this round of AI's upward cycle—and he will not be the last.
