Flo Crivello watched his AI bill and decided it was a survival problem. The CEO of the startup Lindy moved all of his company's traffic off Anthropic's Claude models this month and onto DeepSeek, a cheaper open-weight Chinese alternative, and told CNBC the cost curve "crash to the ground." The switch should save his roughly 25-person firm millions over the coming months. Even so, he still expects to spend more on AI than on payroll. That single detail captures where the industry has landed: the bills are so large that cutting them in half still leaves them larger than the staff.
For three years the prevailing instinct ran the other way. In the rush to build with generative models, especially in coding, companies competed to consume as many tokens as possible, the era that came to be called "tokenmaxxing," where developers were rewarded for using more AI rather than getting more out of it. Uber found out where that leads. The company blew through its entire annual AI budget in four months, its CTO told The Information, and has since imposed spending tiers that start at $1,500 per employee per month. The crackdown is now the trend, and OpenAI and Anthropic, the two labs that grew fastest on the spend-at-all-costs mood, are the ones most exposed to its reversal.
The timing is awkward, because both filed confidentially for what could be historic public offerings in early June, with valuations approaching a trillion dollars each. "Current growth rates for Anthropic and OpenAI are the fastest they will ever be, which is mostly a matter of basic math," said D.A. Davidson analyst Gil Luria, who reads the rush to list partly as a hedge against exactly this: that the labs' biggest enterprise customers are starting to ration what they buy.
Wall Street is asking the same question one layer up. The Nasdaq slid nearly 5 percent this week on the worry that the trillions flowing into AI, Goldman Sachs estimates $7.6 trillion in data-center spending through 2031, will not produce matching revenue. "The returns are not coming in, and the claims that are being made, in terms of efficiency or productivity numbers, are not netting out," said Kate Brennan of the research institute AI Now. A May study from Gartner found that companies replacing workers with AI agents often fail to earn back the cost. Americans are using more AI, much of it because they cannot avoid it, but few are paying for it, and Pew finds 40 percent expect it to be a net negative for society against 16 percent who expect the opposite.
The critic Gary Marcus has a name for the resulting mood: the "Generative AI Fizzle." His argument is not that the technology vanishes but that the economics deflate, slowly, as investors tire of the gap between enormous hype and modest profits. He has long predicted that an industry built almost entirely on large language models would have no moat, that the models would become commodities, and that price wars would follow. The arrival of yet another capable open-source Chinese model, the kind of thing that let Crivello halve his bill overnight, is that prediction arriving on schedule. Tulips did not disappear when the bubble burst, Marcus likes to note; only the astronomical prices did. The open question now is whether AI is facing a true reckoning or merely its first honest conversation about what any of it is worth.