In the classic book “Manias, Panics, and Crashes: A History of Financial Crises,” the economic historian Charles Kindleberger divided episodes of speculation into five stages. Citing the work of Hyman Minsky, the post-Keynesian theorist, Kindleberger writes that the process begins with a displacement, such as the emergence of an important new technology, or a major policy shift, that gets investors excited about the future. They start to bid up asset prices, and capital floods into the affected sector. As optimism spreads, an investment boom—the next stage—develops, much of it facilitated by credit expansion, and the repercussions are felt in the broader economy. A period that he calls “euphoria” ensues. Investors who missed out on the initial gains rush to catch up, prices rise exponentially, and skeptics get dismissed as dinosaurs. At some point, though, the divergence between valuations and economic fundamentals becomes increasingly hard to ignore. Some investors get nervous, and prices start to top out. Overextended individuals or institutions get into trouble. Kindleberger referred to this as the stage of distress. Finally, comes the revulsion, or the bust. Panicked investors try to exit the market. Prices collapse.

Judging by recent events, the A.I. bubble may be approaching the point when euphoria gives way to distress. On the one hand, there is still a great deal of optimism and FOMO. Last week, the Dow and the S. & P. 500 both hit new highs. Purchases of call options, which enable investors to bet on the market going even higher, reached record levels. According to a recent survey, about one in five men aged eighteen to twenty-nine are trading stocks on a daily basis. But some cracks have also started to appear, including a dive in semiconductor stocks that led to a crash in the South Korean market. And now we have before us the spectacle of Leopold Aschenbrenner, a twenty-four-year-old hedge-fund manager who, not so long ago, was hailed as the Nostradamus of A.I.

After graduating from Columbia at nineteen as the class valedictorian, Aschenbrenner worked for FTX, the defunct crypto exchange founded by Sam Bankman-Fried, and at OpenAI, which dismissed him for allegedly leaking sensitive information. In June, 2024, he published a long essay titled “Situational Awareness: The Decade Ahead,” which argued that A.I. models would achieve human-level intelligence by 2027 and superintelligence by 2030. Exploiting a positive reception in Silicon Valley, Aschenbrenner launched a hedge fund, also called Situational Awareness, which loaded up on stocks of companies that stood to benefit from A.I., such as chipmakers and cloud-computing outfits. For a time, the fund proved incredibly successful. By this June, according to the Wall Street Journal, its value had risen more than a thousand per cent since its inception.

Last month, Aschenbrenner got caught out. As semiconductor stocks fell, some of his biggest holdings plunged in value, leaving him exposed. Wall Street banks that had lent Situational Awareness gobs of money to magnify its bets issued the fund margin calls—non-negotiable demands to stump up more cash. After trying fruitlessly to raise some money, Situational Awareness was forced to sell the bulk of its stock at a discount to Citadel, a more established hedge fund founded by the billionaire Ken Griffin. To add insult to injury, these events occurred days before Aschenbrenner was set to get married. The event went ahead, but Aschenbrenner and his wife, Avital Balwit, a high-ranking staffer at Anthropic, reportedly cancelled their honeymoon.

Unlike Long-Term Capital Management, Amaranth, and other big funds that collapsed after suffering big losses, Situational Awareness managed to survive that financial crisis, and Aschenbrenner vowed to “fight another day.” Apparently, his fund retains some very valuable investments in private A.I. companies that haven’t yet issued public stock, including Anthropic. But the market tremors that preceded Aschenbrenner’s humbling highlighted some larger questions about the tech-driven stock boom, one of which is the spectre of heightened competition in the A.I. sector.

The A.I.-to-the-sky narrative was built on the supposition that OpenAI, Anthropic, and other U.S. tech companies would dominate the industry, and their customers would happily pay high prices for access to proprietary, state-of-the-art models. Early last year, the Chinese A.I. company DeepSeek challenged this narrative by releasing a cheap open-source model with impressive capabilities. That was only the beginning. In the past few weeks, two more Chinese firms, Alibaba and Moonshot, have put out models that, based on some metrics, can match the latest and most powerful offerings from Anthropic and OpenAI. DeepSeek also launched a new model, V4 Flash, which, according to Bloomberg, can execute certain tasks for a cost of three cents, compared with $3.15 when using Anthropic’s Claude Fable 5 model. “When China asks for cents where American companies charge dollars, the contest between the two nations for A.I. customers . . . starts to look materially different,” Bloomberg reported.

Other potential sources of vulnerability include the vast sums of money that A.I. companies are spending on infrastructure, such as chips and data centers. A new study from Goldman Sachs predicts that global investment in A.I. will total more than a trillion dollars this year, with outlays in the U.S. making up nearly sixty per cent of the total. From one perspective, the financial system’s ability to fund this scale of spending is impressive, but so is the amount of debt and other liabilities that A.I. companies are taking on, both directly and indirectly, to finance their investments.

The dot-com bubble of the late nineteen-nineties and the real-estate bubble of the early two-thousands both saw the proliferation of financial-engineering techniques such as circular financing and offloading debts from corporate balance sheets to special-purpose vehicles. Both have been resurrected in the A.I. boom. The chipmaker Nvidia is famous for investing in companies that buy its chips, and there are many other interconnections. Microsoft owns twenty-seven per cent of OpenAI, which, according to Bloomberg, “accounted for more than half, and likely about 70%, of Microsoft’s actual AI sales during its most recent fiscal year.” Should anything bad happen to OpenAI, Microsoft would feel the pain. As the sums expended on A.I. investments get larger, the financing that underpins them is getting more complicated and opaque. According to a recent report, Alphabet, Amazon, Meta, Microsoft, and Oracle have a combined total of $1.65 trillion in debt and other obligations that don’t appear on their balance sheets.

To justify all these investments and financing schemes, A.I. will have to create a great deal of economic value. A couple of weeks ago, investors were reassured after Alphabet, Amazon, and Microsoft reported that their respective cloud divisions enjoyed strong revenue growth in the latest quarter. But if A.I. is really going to pay off, its financial benefits will have to show up on the books of its end users—not just A.I. labs, hyper-scalers, and equipment suppliers. A year ago, I incurred the wrath of some people in Silicon Valley by discussing a survey from M.I.T. which found that ninety-five per cent of companies using A.I. tools hadn’t seen any increase in their profits. Perhaps management failures and the novelty of the A.I. models were holding some companies back from finding profitable ways to adopt them, I suggested. But a year later the picture doesn’t seem to have changed dramatically. More recent surveys, including one from the research firm Gartner, still show a surprising dearth of profits from A.I. investments.

What does all this add up to for the stock market? For the moment, A.I. optimism, FOMO, and trend-following are still in ascendancy. It’s hard to say what might change this, but one candidate is a tightening of monetary policy. The market crashes of 1987, 2000, and 2007 were all preceded by periods in which the Federal Reserve raised rates. Until Friday’s disappointing jobs report, which showed employers shedding twenty-three thousand jobs in July, many Wall Street analysts were expecting a rate hike next month because inflation is running above the Fed’s target. Now all bets are off.

Stocks rose again after the jobs report came out, which isn’t surprising. The U.S. market appears to have shrugged off the crash in South Korea and the troubles of Aschenbrenner, an attitude that is encouraging investors to double down on momentum trades that can generate further price appreciation. This type of self-reinforcing dynamic is a recurring feature of speculative booms, and it helps explain why they can last for longer than common sense, and the theory of mean reversion, might predict. And yet each leg up in the market heightens the risk of more Situational Awareness-like squeezes, and of individualized distress giving way to a broader bust. Kindleberger, who died in 2003, would have been an interested onlooker. Surveying the long history of boom and bust cycles, he wrote, “Investors seem not to have learned from experience.” ♦