The month the AI trade met a margin call

A star fund unwound, Korea's market broke records falling, and Big Tech kept spending anyway. The gap between those three facts is the story worth watching.

The most instructive AI story this week was not a model release. It was a hedge fund losing most of its money in about three weeks, and what that unwind revealed about how much of the AI boom is running on borrowed conviction and borrowed cash.

Leopold Aschenbrenner built Situational Awareness around a single, coherent idea: that scaling AI would demand a vast build-out of chips, memory, power, and data centers, so you should own the companies supplying all of it. The thesis was not obviously wrong. The problem was the wrapper around it. In a matter of weeks his fund went from managing roughly $45 billion to being forced into a sweeping reduction of its listed-stock positions, as a momentum reversal triggered losses on both sides of a portfolio that owned AI infrastructure names while betting against software firms. Leverage did the rest. Assets dropped to roughly $10 billion after Citadel bought the leveraged positions, among them SK Hynix and CoreWeave, at below-market prices.

What makes this more than a rich-guy cautionary tale is who was standing nearby. The frontier AI world does not just build together, it invests together, often through the same funds and special-purpose vehicles, so a single blow-up sends tremors through a tightly connected community. Worth noting for perspective: even after the wreck, reporting suggests Aschenbrenner remained up meaningfully on the year, and the fund kept its private stakes, chiefly Anthropic. A correct long-term call and a margin call can coexist, and the margin call arrives first.

When two chipmakers are the whole market

The same reversal hit its epicenter in Seoul, and the mechanics rhyme. SK Hynix reported a near sixfold increase in operating profit to a record, roughly $42 billion, but that still fell short of forecasts of around $44 billion. Beating your prior year by 500% and getting sold off is what happens when expectations have run further than even spectacular results. The Kospi slid enough to trigger a circuit breaker for a second straight day and was on course for a record monthly loss of about 35%, after a rally that relied almost entirely on two chipmakers.

The accelerant was familiar. Much of the buying had been done by small investors using borrowed money, and that dynamic, which pushed the rally higher, worsened the selloff as brokers forcibly closed losing positions. South Korea's finance minister even apologized in parliament for approving the leveraged single-stock products that amplified the swings. Concentration plus leverage is a fine engine on the way up and a trapdoor on the way down, whether the vehicle is a $45 billion fund or a retail ETF.

The spenders are not flinching

Here is where the picture gets genuinely interesting, because the companies at the center of the trade are behaving as though nothing happened. Amazon, Microsoft, Alphabet, and Meta all used earnings this week to argue that demand for AI infrastructure justifies extraordinary spending, and none signaled a pullback. Amazon lifted its 2026 capital plan from $200 billion to $220 billion after AWS grew 37% year over year, and Alphabet pushed its own forecast toward roughly $200 billion.

So the market and corporate treasuries are telling opposite stories. Public investors are asking whether the capex converts to revenue soon enough; the hyperscalers are answering by spending more. Both cannot be right for long. If the spenders are correct, this week's rout is a buying opportunity and Citadel just picked up a discounted book. If the skeptics are correct, the capex is early to a demand curve that arrives slower than the depreciation schedule.

There is a quieter data point that cuts toward discipline. A survey of finance leaders found that 79% of enterprises had experienced AI cost overruns in the past year, and stories of firms burning through annual budgets on agentic coding are becoming common. Meanwhile OpenAI cut prices on two GPT-5.6 models, one by 80%, precisely because customers have grown cost-sensitive. Demand is real, but buyers are starting to negotiate, and cheaper open-weight models now match closed ones closely enough on many tasks to give them leverage. That is not the profile of a market with unlimited pricing power.

So what

The useful takeaway is not that the AI boom is fake or that it is invincible. It is that the boom and the trade on the boom are two different things, and this week they came apart. The technology kept advancing; the leveraged, concentrated bets on it did not survive contact with a single disappointing earnings call.

For anyone allocating capital or budget, the discipline is the same one that failed Aschenbrenner and the Korean retail crowd: size positions so a temporary drawdown cannot force you to sell at the bottom, and treat compute spend the way a CFO treats any input cost rather than an article of faith. The firms that endure the next leg will be the ones that can be patient through a drawdown, whether that drawdown shows up in a portfolio or a cloud bill.