When AI Hype Outran the Money: The Rise and Fall of Situational Awareness

A 24-year-old’s AGI prophecy turned into a $20 billion fund, then a cautionary tale as markets and leverage reversed.

This article originally appeared on Weijin Research on Sina on July 31, 2026. Original Chinese title: 「AI，开始蹦最靓的崽」. It has been translated and adapted for an English-speaking audience.

This is a story about the brightest star on Wall Street in the AI era.

It is about how AI and investment prodigy Leopold Aschenbrenner turned an AGI prophecy into a $20 billion fund, and how that same narrative and leverage then turned against him.

A month ago, the 24-year-old Leopold Aschenbrenner was still hailed as the oracle of AI investing.

He was young, handsome, became valedictorian at Columbia University at 19, and joined OpenAI at 22 to work on superalignment. After leaving, he wrote Situational Awareness, an AI manifesto that sent shockwaves through Silicon Valley and Washington, then turned that report into a hedge fund. By June 2026, the fund, less than two years old, had grown to manage over $20 billion. Its net return since inception exceeded 1,000%, and it had gained roughly 270% in just the first five months of 2026. Every regulatory filing he made was picked apart by followers as if they were studying scripture.

Professional headshot of a man in business attire against a neutral background. Watermark visible in lower right corner.

But by late July, the story suddenly flipped.

Situational Awareness suffered heavy losses in the AI stock sell-off. It began urgently raising capital from existing investors and lenders, and proposed to some investors that they buy assets directly from the fund’s portfolio. The borrowing and leverage that had helped create its myth now amplified the losses. An AI investing legend quickly turned into a cautionary tale of leverage backfiring.

The market correction is not over yet, and whether the fund will ultimately be wound down remains to be seen. Reliable reports so far only confirm that it took a severe hit, is seeking new money, and is considering selling some assets. The fund’s specific losses, the size of its borrowings, and its latest net asset value have not been disclosed. What can be said with certainty is that what collapsed was its aura of invincibility, along with the illusion that long-term technological judgment can be seamlessly converted into short-term investment gains.

From Berlin prodigy to OpenAI

Aschenbrenner was born in Berlin to parents who are both doctors. He entered Columbia University at 15, majoring in economics and mathematics/statistics, and graduated at 19 as valedictorian. Columbia’s assessment of him was unusually strong: not only did he rank first in his class, but his senior thesis also won the economics department’s highest honor. While at university, he co-founded an effective altruism group. His concerns were not ordinary business questions, but long-term economic growth, technological progress, and existential risks to humanity.

After graduation, he studied economic growth and existential risk at Oxford’s Global Priorities Institute, and also briefly worked at FTX’s Future Fund, leaving before FTX collapsed. Aschenbrenner is first and foremost an intellectual who studies technology trends, long-term growth, and extreme risks, not a traditionally trained securities analyst or trader.

In 2023, he joined OpenAI’s Superalignment team, led by co-founder and CTO Ilya Sutskever and VP Jan Leike. The team tried to answer a near-paradoxical question: if future AI is smarter than humans, how can humans supervise and control it?

Aschenbrenner’s representative contribution was the “weak-to-strong generalization” research: using a weaker model to supervise a stronger one, simulating the difficulty humans will face when supervising superintelligence. The study found that strong models can outperform weak supervisors, but ordinary fine-tuning still fails to unlock their full capabilities, suggesting that existing RLHF methods may not naturally scale to superintelligence.

He did not leave OpenAI voluntarily; he was fired. After OpenAI’s internal communications system was hacked in 2023, Aschenbrenner submitted a security memo to the board, arguing that the company’s measures to protect model weights and critical algorithmic secrets were “grossly inadequate,” especially against state-level industrial espionage. Later, he sent an internal discussion document on AGI preparedness, safety, and security measures to three outside researchers for feedback. OpenAI viewed this as leaking confidential information and fired him in April 2024.

The two sides disagree on the real reason. OpenAI said the firing was unrelated to his raising security concerns with the board; Aschenbrenner said the company explicitly told him at the time that sharing the document would normally have only warranted a warning, and that it was the earlier security memo that escalated the punishment to termination.

He is therefore neither a simple “safety whistleblower” nor can he be reduced to an “employee who leaked information.” What is clear is that his views on AI capability growth, lab safety, and US-China competition put him in serious conflict with OpenAI’s management.

A viral report/investment prospectus

In June 2024, Aschenbrenner published a 165-page document titled “Situational Awareness: The Decade Ahead.” It is often called the “AI Situation” report, but its formal title is “Situational Awareness.”

Intelligence Explosion scenario projecting effective compute (normalized to GPT-4) from 2018 to 2030, with annotations for major AI milestones and thresholds for superintelligence. Source: Leopold Aschenbrenner, Situational Awareness.

The report's most important judgment is that achieving AGI by 2027 is "very credible." His extrapolation rests on three variables: training compute grows by about 0.5 orders of magnitude each year; algorithmic efficiency improves by another 0.5 orders of magnitude annually; and "unshackling" technologies — agents, tool use, long-chain reasoning — will turn chatbots into digital workers that can operate independently. Combined, these factors could deliver roughly a 100,000-fold increase in effective compute between 2023 and 2027.

Once AI can automate AI research, hundreds of millions of AI researchers could compress a decade's worth of algorithmic progress into a single year, rapidly propelling AGI into superintelligence. That transition would be paired with a massive expansion of compute clusters, data centers, and power generation on the scale of hundreds of billions to trillions of dollars, along with an eventual U.S. government-led "AGI Manhattan Project."

Line chart comparing annual electricity generation in China and the USA from 1985 to 2030, with projections for total AI demand and largest training cluster beyond 2025. China's generation (red line) shows steep growth after 2010, while US generation (blue line) remains relatively flat. Source: Leopold Aschenbrenner, Situational Awareness.

The report caused a sensation not just because it predicted AGI. It was the first to weave model capability, compute, algorithmic progress, electricity, data centers, national security, and U.S.-China competition into a single, complete industry map. It was at once an AI technical report, a blueprint for industrial mobilization, and later, in effect, a fund marketing document.

Its weaknesses were just as obvious. Continuous trends in compute and benchmarks were simply extrapolated into qualitative leaps in reliability, autonomy, and economic value. It underestimated hallucinations, the long tail of complex tasks, data bottlenecks, and enterprise deployment friction. And it packaged highly uncertain technology forecasts as an urgent geopolitical imperative. Critics noted that the 165-page report mentioned “hallucination” only once, even though reliability is one of the main obstacles to turning a high-scoring model into an autonomous worker.

Silicon Valley’s money bets on Silicon Valley’s prophecy

After publishing the report, Aschenbrenner founded an investment firm of the same name, Situational Awareness. The first anchor investors were not ordinary financial institutions, but Silicon Valley’s core technical elite: Stripe co-founders the Collison brothers (Patrick Collison and John Collison), former GitHub CEO Nat Friedman, and investor Daniel Gross. Later, Jane Street, a top trading firm that rarely invests in outside funds, also became a backer.

The full list of investors is not public. But SEC filings show that as of the end of December 2025, Situational Awareness Partners had sold roughly $1.026 billion in fund interests to 66 investors, with a minimum investment of $5 million. The $1.5 billion or even $20 billion scale cited by the media includes subsequent fundraising, investment gains, and asset appreciation, and should not be taken to mean the firm started with $20 billion in cash.

His strategy was to turn each chapter of Situational Awareness into a trade:

If AGI is near, go long AI infrastructure. If training clusters are heading toward a trillion dollars, find the bottlenecks in power, data centers, GPU cloud, storage, and memory. If traditional industries will be displaced by AI, short the laggards. If model labs ultimately capture the most value, invest directly in private companies like Anthropic.

By 2026, the fund’s holdings spanned CoreWeave, Applied Digital, IREN, Bloom Energy, SanDisk, Oracle, AMD, and a number of companies pivoting from crypto mining to AI data centers. Anthropic at one point accounted for roughly one-fifth of the fund’s assets, making it one of the largest single investments. Anthropic’s chief of staff is his fiancée, Avital Balwit, the only person who reports directly to founder and CEO Dario Amodei.

The Q1 2026 13F filing also disclosed roughly $7.7 billion in notional exposure to semiconductor-related put options, including Nvidia, the VanEck Semiconductor ETF, Oracle, Broadcom, and AMD. At the same time, the fund remained long GPU cloud, power, and data center companies. However, 13F filings disclose only the notional value of options, not the actual premiums paid, nor whether the fund bought or sold the options, so it cannot be simply concluded from this that Aschenbrenner has turned “bearish on AI.”

The real danger lies not in any single stock, but in the portfolio’s overall use of borrowings and derivatives leverage. It appears to invest across chips, storage, power, cloud, and Asian markets, but in reality these assets are all exposed to the same risk factor: whether the market continues to believe that AI capital expenditure will grow indefinitely.

From 1,000% returns to emergency fundraising

At first, the strategy was spectacularly successful.

In the first half of 2025, the fund rose 47% net of fees, compared with a roughly 6% return for the S&P 500 and about 7% for the tech hedge fund index. By the end of May 2026, Situational Awareness was up roughly 270% for the year, with cumulative returns since inception exceeding 1,000%, and assets under management surpassing $20 billion. This reflected appreciation in private holdings like Anthropic, surging power, storage, and AI infrastructure stocks, and a steady inflow of new capital.

By the end of June, the fund’s net return for the year had even reached 439%. But in July, the AI trade abruptly reversed. Oracle and AMD fell roughly 20% in a month, SanDisk dropped 54% in July, and more volatile names like Nebius, Sharon AI, and Bloom Energy fell even harder. South Korea’s Kospi index at one point fell nearly 40% from its high a month earlier, with the highly concentrated AI memory and semiconductor sector at the center of the storm.

Global AI stocks experience significant correction, with major declines and half-year gains. Horizontal bar chart showing recent price drops (green bars, ranging from -64.66% to -18.14%) and six-month gains (red percentages, ranging from 8.03% to 857.84%) for 25 major semiconductor and AI-related companies. Source: Huahuamao iFinD.

Price declines alone may not be enough to bring down a long-term fund. Leverage turns a price problem into a cash problem. Even if the fund still holds substantial net assets, prime brokers can demand additional margin, reduced risk exposure, or loan repayment. Privately held Anthropic shares cannot be sold at will, so the pressure falls on publicly traded stocks and options that can be liquidated. This explains how a fund that was still up 439% at the end of June needed to raise money from investors and lenders by the end of July.

Mathematically, wiping out a 439% year-to-date gain would require the fund to drop roughly another 81.4% from its end-of-June net asset value. So until specific loss figures are disclosed, it cannot be asserted that the fund has swung from profit to loss, much less that it has formally liquidated. What it appears to be facing right now is a severe liquidity and financing crisis, not a portfolio value that has gone to zero.

Aschenbrenner himself has not conceded defeat. In his July 24 investor letter, he acknowledged that the fund failed to sidestep market turbulence, particularly the selloff in Asian markets, but also said this was one of the best opportunities to add positions since early 2025 and opened a new capital window for August 1. He also pointed to Anthropic's potential public listing in the second half of the year as a catalyst for a rebound.

What he really got wrong

Aschenbrenner's biggest mistake may not have been misreading AI. It was conflating four different clocks: the clock of technical capability progress, the clock of enterprise deployment and diffusion, the clock of revenue and cash flow formation, and the clock of capital market pricing and financing.

Getting the technology direction right does not mean a stock is worth buying at any price. Real long-term scarcity does not mean scarce assets cannot price in a decade's worth of returns ahead of time. And eventually reaching AGI does not mean a highly leveraged fund can survive until the day AGI arrives.

The second lesson is that a large number of stock tickers does not equal true diversification. Chips, memory, data centers, power, GPU clouds, and bitcoin mining rigs belong to different industries, but when their valuations all rest on faith in "sustained upward revisions to AI capex," they are the same trade. When markets rise, they generate excess returns together. When risk appetite reverses, they fall together.

The third lesson is that leverage changes the nature of the investment proposition. Without leverage, AI infrastructure can be an industrial judgment with a five-year horizon. Add heavy borrowing, options, and concentrated positions, and it becomes a trade that must maintain margin and financing channels every single day. Whether you are right in the long run becomes secondary. Whether you can survive in the short run becomes the primary question.

The fourth lesson is that intellectual influence itself creates reflexivity. His papers helped him raise capital. Stellar performance turned regulatory filings into market events. Follower money further pushed up his holdings. Investment ability, personal prestige, and asset prices reinforced each other, until the reflexivity began running in the opposite direction on the way down. The so-called "AI prophet" can easily become the center of a crowded trade.

This is not the end of AI. The pricing logic has changed.

The Situational Awareness crisis does not prove that AI technology has stalled. On the contrary, many of its judgments about capex and infrastructure expansion are materializing. Goldman Sachs estimates that major AI suppliers' 2026 capex will be roughly $755 billion, potentially reaching $920 billion in 2027. U.S. tech investment as a share of GDP has already surpassed the peak of the 1990s internet bubble.

The problem is that the market has already priced enormous technological progress into current values. Since ChatGPT's launch, the market capitalization of AI-related companies has increased by roughly $27 trillion. Under Goldman's baseline scenario, the present value of incremental capital returns from AI-driven productivity gains is roughly $9 trillion. The two figures are not perfectly comparable, but the gap shows that further gains require increasingly optimistic assumptions about margins, market share, and value capture.

The AI industry is therefore moving from a first phase of "if you build it, it will rise" into a second phase of "who can turn the buildout into cash flow." The most vulnerable valuations will be GPU cloud and data center companies that rely on debt expansion, lack long-term customer contracts, and suffer from low utilization rates, along with component suppliers priced for permanent shortage. Those with greater resilience will be companies with real demand, long-term contracts, ample cash, and verifiable AI revenue.

Advances in algorithmic efficiency and Chinese open-source models will further redistribute value. Cheaper inference may expand total computing demand through a Jevons paradox, but it may also drive down per-unit compute prices, shifting profits from chips and infrastructure to models, applications, and platforms that own customer relationships. Growth in AI usage is not the same thing as sustained profit growth for every AI infrastructure company.

Aschenbrenner's most famous line is: "You can see the future first in San Francisco." That may still be true. But real situational awareness means seeing not just the future, but also valuation, debt, liquidity, and time.

A person can correctly predict a technological revolution and still lose the trade before the revolution arrives. The brightest star in AI may not be completely out of the game yet, but his high-speed fall is a reminder to everyone: the market does not let you borrow unlimited money to cash in the future early, just because your judgment about that future is right.