The stock market spent two years treating artificial intelligence as a one-way bet. This week it flinched. A gauge of United States chip stocks tumbled more than 5 percent, the tech-heavy Nasdaq 100 fell 1.6 percent, and the wobble carried into Asia, where Japan and South Korea retreated and the region's best-performing benchmark of the year, the Kospi, slipped as SK Hynix fell with it. The long-term faith in the technology is intact. What cracked was confidence that today's valuations can keep pace with the spending required to justify them.
Read the analyst notes and a specific fear emerges, and it is not that AI will fail. It is that AI will get cheap. "There are concerns that high memory prices will bring AI solutions that need less memory, and that the data-center build-out may not all get built in the end," warned Louis Navellier of Navellier & Associates. Then the sentence that ties the market to the models: "token pricing of AI software will push users to lower-cost versions, especially Chinese offerings." The same open-weight models winning over developers this month are now showing up in the reasoning of people who price semiconductor stocks. If good enough costs a sixth as much, some of the demand underwriting the boom evaporates.
Underneath the day's numbers is a slower change in what investors are grading. For two years the AI industry compared report cards: bigger context windows, higher benchmark scores, another few billion raised. Boardrooms have started asking a duller question, how much money will this make or save us, and markets are beginning to ask it too. Palantir's Alex Karp has been loudly attacking the token-based pricing of frontier labs, but the real force is arithmetic. A chief financial officer does not see a marvel of engineering; she sees a line item that grows with every use. When the metric shifts from intelligence to economics, economics tends to win.
Consumers are about to feel the same recalibration. Flat-rate AI subscriptions have always been sold at a loss: by one estimate a $200-a-month plan, fully used, delivers roughly $14,000 of raw interface access. Now the subsidy is being withdrawn in public. With its most powerful model, Fable, finally back from export limbo, Anthropic told Claude subscribers they would get one month of access rather than the promised longer trial, then face API prices, and that even during that window they could spend only half their usage limits on it. OpenAI's new GPT-5.6 family, the lightweight Luna, the everyday Terra, and the flagship Sol, remains walled off to trusted partners. The all-you-can-eat era is quietly ending.
None of this means the boom is over. Payrolls are still growing, the frontier labs are still shipping, and the appetite for capable models is real. But the first phase of the AI story was about proving the technology could do extraordinary things, and that argument is won. The second phase is about proving those things are worth what they cost, and that verdict is still open. This week the market started marking the exam, and for the first time in a while it did not hand out an easy A. The question now is whether this is a healthy repricing or the first honest look at a bill that has been growing quietly for two years.