Almost nobody disputes that AI is transformative. Less than four years after ChatGPT appeared, roughly 30 percent of Americans are using the technology, five times the share who were on the internet or using electricity at comparable stages of their rollout. That very speed is the heart of a provocative argument in Foreign Affairs by the investor Ruchir Sharma: the faster and more real a technology revolution is, the bigger the bubble it tends to inflate. The innovation is genuine. So, he contends, is the mania built on top of it, and manias end the same way they always have.
Sharma's checklist is what he calls the "Four O's": an object of desire becomes overvalued, overowned, overinvested and overleveraged. The evidence is not hard to find. The ten biggest stocks in the S&P 500 are all AI players, together making up nearly 40 percent of the index and about half its returns this year. American households now hold more than 30 percent of their wealth in stocks, a record, much of it funnelled through passive funds straight into those same AI names. Corporate spending on plants and equipment has hit 6.3 percent of GDP, higher even than at the peak of the dot-com boom, and is still climbing.
The bulls answer that this time the giants are genuinely profitable, so it cannot be a bubble. Sharma's reply is that strong earnings have never been proof of anything; the dot-com leaders like Cisco and Oracle posted real profits too. More unsettling, he argues, is how those earnings are being propped up. There is "circular funding," in which a cloud giant invests in a lab such as OpenAI or Anthropic, and the lab spends much of the money buying computing power back from that same giant. Revaluing those stakes and booking the gain as "other income" has accounted for about a quarter of hyperscaler earnings growth this year. Off-the-books leases could add some $3 trillion in borrowing over the next decade. And government deficits, by one measure, have underwritten more than half of US corporate profits.
If the money is the fuel, the question is what cuts it off. Bubbles do not burst because sentiment politely turns, Sharma writes; they burst when liquidity tightens. The players to watch now are not central bankers but the "bond vigilantes," investors who punish free-spending governments by dumping their debt and driving up long-term rates. Tremors have already run through semiconductor shares this summer, and Bank of America has warned of "bubble-like instability" spreading worldwide. Because so many economies are plugged into the same trade, a crash would radiate outward: from the United States to Japan, South Korea and Taiwan, then down the supply chain to the miners digging lithium and copper and the utilities selling power.
The essay's closing move is the interesting one, because it is not simply doom. Sharma calls AI a "hybrid" bubble: a good bubble in a real, productive technology, like the internet, but financed by debt as well as equity, like the housing crash. Good bubbles leave something valuable behind and a leaner set of survivors; as one strategist he quotes puts it, "the coming massacre of AI companies and the coming growth of the AI economy are the same event." The catch is timing and cushioning. Productivity gains from past breakthroughs often arrived decades late, and this time governments already carrying record debt will have little room to break anyone's fall.