Asian markets opened by repricing the AI trade. South Korea's KOSPI fell 3.7%, with the chipmaker SK Hynix down 5.75%; Taiwan Semiconductor Manufacturing Company lost 1.2%; SoftBank, one of the largest investors in the sector, dropped as much as 13% in Tokyo. The cause was not an earnings miss or a supply problem. Anthropic chief executive Dario Amodei called on the industry to "slow down", Sam Altman and Elon Musk backed him shortly afterwards, and investors began marking down the companies building the infrastructure on the assumption that the people who order it may want less of it.
Amodei's argument is that building AI too quickly is reckless, and that a swarm of AI agents could one day do hundreds of billions of dollars of damage if it gained control of the whole internet. Some AI researchers have disputed the claim. The dispute did not matter to the tape: investors are already pricing in a slower rate of development, which would make it harder to finance the pace of data centre construction now underway.
One caveat belongs on the numbers before they get quoted elsewhere. The largest fall of the session has its own explanation: SoftBank, which holds a stake in OpenAI, slid after Altman said the ChatGPT maker would not go public this year. Anyone citing a 13% drop as the market's verdict on AI safety is stacking two separate stories on top of each other. The moves that actually reflect the safety argument are the chip names, and there the 5.75% at SK Hynix against 1.2% at TSMC is the detail I would watch — the selling concentrated where revenue depends most directly on the next build cycle rather than across semiconductors as a whole.
Ipek Ozkardeskaya, senior analyst at Swissquote, said the mood on the markets this morning was poor, and framed the underlying problem as an accounting one. If the AI race slows appreciably, the question becomes who pays for the infrastructure that has already been built. Rent, debt and power commitments do not disappear because expected demand for compute and revenue growth have come down.
That transmission runs straight into credit. Ozkardeskaya pointed to the risk building in highly leveraged data centre operators and in the lenders financing projects underwritten on optimistic assumptions about future AI demand — and it is accumulating just as interest rates, and therefore the cost of borrowing, are expected to rise.
The shift worth understanding is in the nature of the risk, not its size. AI safety has been treated by markets as a subject for conference panels, orthogonal to the financial case. What Amodei's intervention did was convert it into a demand variable. The build-out was financed against a forecast of compute demand; if the labs themselves decide to grow more slowly, that forecast becomes a discretionary choice made by a handful of executives, while the obligations written against it stay fixed. Fixed costs set against a revenue line that someone can elect to reduce is a different security than the one many of these projects were sold as.
What no one specified is the one thing the models actually need: how much slower, and starting when. None of the executives calling for restraint attached a horizon to it, and the facilities now under construction were financed against demand arriving on a particular schedule. A slowdown of six months and a slowdown of three years produce identical headlines and completely different credit outcomes.
Ozkardeskaya's closing point is the one that should unsettle the sector most. The slowdown, if it comes, will not come from big technology companies running out of money, or from investors declining to fund them. It will come from the model developers themselves — an industry financed on the premise that only capital could stop it, discovering that the builders can.