The Aschenbrenner Cascade: What a 67% Hedge Fund Collapse Actually Tells Us About the AI Trade

Leopold Aschenbrenner was right about AI. He was also leveraged four-to-one into the most crowded trade on Wall Street, which means being right didn't save him.

Situational Awareness, the hedge fund Aschenbrenner built around his 165-page AI infrastructure thesis, lost 67% of its value in July. The fund was forced to sell nearly its entire public equity portfolio to Ken Griffin's Citadel and strip all leverage. At its peak earlier this month, the fund managed $45 billion. That number looks like a disaster movie, and the financial press has already framed it as proof the AI buildout is overextended.

It is neither. The thesis didn't break. The financing did. And that distinction is the difference between a stock-picking problem and a sector-rotation problem.

The mechanics of the cascade

Aschenbrenner's 2024 white paper - "Situational Awareness: The Decade Ahead" - predicted trillion-dollar AI infrastructure spending, massive data center buildouts, and growing government involvement in computing power. Those predictions have aged well. Hyperscalers spent $673 billion in combined capex in 2026, up 76% year-over-year. Nvidia signed a $500 billion supply partnership with SK Group. The AI buildout is real, it is accelerating, and Aschenbrenner saw the trajectory early.

The problem is what happened between the thesis and the margin call.

Situational Awareness concentrated on AI infrastructure names - CoreWeave, Nebius, SK Hynix, Sandisk, Micron, IREN - and used roughly four times leverage to amplify returns. Reports indicate the fund also ran short positions against software companies like Adobe. The long side generated cumulative gains exceeding 1,000% since launch, which is how a fund that started with $225 million from Stripe co-founders and other Silicon Valley insiders ballooned to $45 billion.

Then three things happened simultaneously in July:

Memory stocks cracked. SK Hynix reported a record operating profit of 60.5 trillion won ($42 billion) in Q2 - up 557% year-over-year, with a 76% operating margin. The stock fell 10% because revenue and operating profit both missed analyst estimates (consensus was ~$84 trillion in revenue, ~$64 trillion in operating profit). The Korean market, which had become a leveraged trading frenzy around Samsung and SK Hynix, began to unwind. JPMorgan and Citi estimate that 65% to 75% of Korea's leverage positions have now been liquidated.

Adobe squeezed. The fund's short positions in software companies like Adobe moved sharply against it. The selloff in AI infrastructure created a rotation back into beaten-down software names that Aschenbrenner had bet against.

Leverage did its job. Four-to-one leverage means a 25% decline in portfolio value wipes out equity before margin calls even start. When SK Hynix, Micron, Nebius, and CoreWeave all fell 30% to 35% in July, the math was unavoidable. Prime brokers - Bank of America, Goldman Sachs, JPMorgan - had to demand cash or force sales. Aschenbrenner himself compared the dynamic to a bank run: each sign of vulnerability created more vulnerability.

The thesis survived the plumbing. The plumbing killed the thesis in the market.

What this isn't: a verdict on AI spending

The financial press has already conflated Situational Awareness's collapse with a broader AI slowdown. That conflation is lazy and misleading.

UBS estimates hyperscaler capex will grow 76% in 2026 to $673 billion, then slow to 25% growth in 2027 and just 6% in 2028. Bank of America's July survey found 82% of fund managers view semiconductors as the market's most crowded trade, with none reporting short exposure. These numbers don't say AI is dead. They say the acceleration may be slowing, which is a very different thing. Decelerating from 76% growth to 25% growth is still massive spending growth. Chip companies that have priced in perpetual 50%+ revenue growth will have a problem. Chip companies that can deliver strong absolute numbers on a moderating growth curve will not.

SK Hynix is the case study. The company reported revenue up 257% and operating profit up 557%. It has 10 long-term supply deals signed, 88 trillion won ($63 billion) in net cash, and a $31 billion capex plan for 2026. The stock has shed roughly $500 billion in market cap since June - not because the business deteriorated, but because investors started asking whether such high margins are sustainable and whether Chinese competitors in conventional memory will eventually pressure pricing.

SK Hynix ADRs are down 3.5% today, sitting at $143.73, after an intraday amplitude of nearly 13%. Micron, another core Situational Awareness holding, is at $823 after falling 5.9% on a single day's volume of 54.5 million shares. Micron is up 188% year-to-date and 108% over the past 120 days, while SK Hynix ADRs are down 3.5% year-to-date. These are not companies whose business models have broken.

A forced liquidation at the wrong price for one leveraged fund is not a thesis revision for anyone who owns these stocks outright.

What this actually is: the mechanics of a crowded trade unwinding

The real lesson from the Aschenbrenner collapse isn't about AI. It's about what happens when the most crowded trade in the market stops being one-directional.

Bank of America's data - 82% of fund managers long semiconductors, zero shorts - means there is almost no one left to buy when the music stops. The Korean margin system, which had retail investors heavily leveraged on chip stocks, created a domestic selling vortex that had nothing to do with fundamentals and everything to do with margin mechanics. When those Korean positions unwound, they sold the same names that Western hedge funds were concentrated in. That's the definition of a crowded trade unwinding.

The cross-currents are clear:

• AI infrastructure demand remains strong. SK Hynix's CEO said memory shortages could persist beyond 2030. Hyperscalers are still signing multi-year supply deals with embedded deposits. The buildout is not over.
• Growth rates are moderating. UBS sees capex growth falling from 76% to 25% to 6% over the next two years. That trajectory matters because semiconductor valuations are priced on forward revenue growth, not absolute spend levels.
• The crowded trade is structurally fragile. When 82% of managers are long on the same names, any earnings disappointment or margin concern triggers forced selling that amplifies well beyond the original catalyst.
• Rotation risk has materialized. The Adobe short working against Situational Awareness reveals a broader dynamic: as AI infrastructure names get beaten up, capital flows back into the beaten-down AI-adjacent names that were sold six months ago. That rotation hurts concentrated infrastructure longs.

Directionally, the demand story is intact but the valuation risk is real. Chip companies with strong cash positions and visible multi-year contracts - SK Hynix with 88 trillion won ($63 billion) in net cash, Nvidia with its $500 billion SK Group deal - have room to absorb a growth slowdown. Chip-adjacent infrastructure companies with thin balance sheets and high leverage (the Nebius and CoreWeave names in Situational Awareness's portfolio) are far more vulnerable if hyperscaler spending growth decelerates faster than expected.

The AI trade isn't over. The easy part of the AI trade is over.

The investor implication

Aschenbrenner wrote to investors after July's collapse: "We let you down this month." The fund is still up roughly 80% year-to-date. His thesis may still play out over the next decade. But he was caught at the exact intersection where a correct long-term view meets short-term mechanical risk.

For the ordinary investor, the lesson is simpler than the headlines suggest. If you believe AI infrastructure companies will create enormous value over the next decade, the question is not whether to own them. The question is how concentrated you are, whether you're borrowing against those positions, and whether you can survive a 30% correction without being forced to sell.

The companies Aschenbrenner backed - SK Hynix, Micron, CoreWeave, IREN - are still building real products for real customers at record margins. What changed in July was not the demand curve. What changed was the margin system, the Korean retail leverage, and the realization that the most crowded trade on Wall Street had no one left to absorb selling.

Being right about the future doesn't matter if you get margin-called before the future arrives. That's the oldest lesson on Wall Street, and Aschenbrenner learned it the hard way. The rest of the market is still debating whether it needs to.