The AI boom is starting to look financially awkward. The price of AI is collapsing, but the cost of building it is not, creating a widening gap between what companies charge and what their infrastructure demands.
Hyperscalers have used an estimated $132bn in debt to fund the datacentre rollout. At the same time, frenzied demand for datacentre components keeps costs elevated, leaving the companies behind the boom with expensive systems and falling prices.
OpenAI has repeatedly cut its fees to hang on to customers. An index tracking what customers pay for a million tokens has more than halved since June, falling to less than $1.
That is a difficult business pattern to admire. Revenue per unit is falling while the machines, chips, facilities, and other components needed to provide the service remain expensive.
The bubble risk is easier to measure than the slowdown
Calls for a slowdown in AI development have grown from several directions, driven by economic, safety, and geopolitical concerns. Yet the word “slowdown” remains so vague that it risks becoming a slogan rather than a plan.
“Nobody has given a substantive explanation of what ‘slowdown’ means,” Ed Zitron said. That is the central problem: calls to pause development can describe economic restraint, safety controls, or a response to geopolitical competition, but those are not the same decision.
President Trump claimed the US was ahead of China in the AI race and said: “Whoever wins AI, wins.” That framing turns AI development into a contest where speed carries political weight, even as the economics of the industry become harder to defend.
The financial stakes have already reached extraordinary levels. The total market value added to the S&P 500 Index since late 2022 is nearly $33 trillion, placing the AI boom inside a much larger market story rather than leaving it as a narrow technology bet.
When an industry attracts that scale of value, a bubble burst becomes more than an embarrassing correction for a few companies. It can expose debt, weaken confidence, and force investors to reconsider whether falling AI prices can support the cost of the systems behind them.
Safety concerns still refuse to disappear
The financial case for caution does not erase the safety case. Dario Amodei, CEO of Anthropic, and Sam Altman, CEO of OpenAI, sit inside an industry facing questions about how fast development should proceed and what risks companies can accept.
An ex-Anthropic employee said: “There is a possibility of human extinction.” Professor dame Wendy Hall, an AI advisor to the UN, put the investment question more bluntly: “Would you invest in a company that says it’s going to bring about human extinction?”
Those warnings carry a different weight from concerns about token prices or datacentre debt, but they now occupy the same debate. Economic pressure asks whether the boom can pay for itself; safety pressure asks whether the race should continue without stronger limits.
Anthropic told investors that its “adjusted operating income” was positive, excluding many costs. That qualification matters because a positive figure can look healthier when costs are left outside the measure—an accounting detail with a long history of making ambitious businesses appear less uncomfortable than they are.
The industry therefore faces two immediate tests. It must explain how falling prices will cover infrastructure whose costs remain elevated, while also answering the people who say the consequences of unchecked development could reach far beyond company balance sheets.
The collapse of the AI bubble may arrive before anyone agrees on what a slowdown means. If it does, the market will make the decision with less patience than the industry has shown in explaining its numbers.
Based on
