The AI race stopped being about capability. It is now about who can afford to lose money.

For three years the argument about artificial intelligence was conducted in benchmarks. Whose model reasoned better, coded better, hallucinated less. That framing suited the American laboratories, because on that measure they were winning, and because it kept the conversation away from the balance sheet.

The framing is now breaking down, and three apparently unrelated stories from the past week show how.

The first is Moonshot's Kimi K3 — free, open-weight, and strong enough at coding that the chief executive of the testing firm Arena called it possibly "the single biggest release of the year". It arrived while Google was months late on Gemini 3.5, holding the release to improve coding performance. The significant fact is not that a Chinese model is good. It is that it is free and downloadable, and that Beijing subsidises the computing and power that make giving it away possible. American firms have to charge eventually. Their competitors, for now, do not.

The second is Water UK telling MPs that the government's forecasts for national water supply are "fatally flawed" because they "explicitly exclude" datacentres — and that the AI growth zones policy "makes not a single mention of water". Three-quarters of Britain's datacentres sit in the south and east, where hosepipe bans are already in force. Individual facilities near Slough have asked for up to 3m litres a day, the peak demand of 3,500 homes. Reservoirs take fifteen years to build. Model releases take months.

The third is the quietest. Tech workers, historically the least unionised professionals in the developed world, are organising. Workers at Google DeepMind and Meta in the UK have moved towards recognition; union actions at Alphabet doubled in 2025. "This sense of privilege and immunity for tech workers — that is gone," says Simone Robutti of the Tech Workers Coalition. A survey of IT staff across the University of California found 65% taking on extra work to cover vacant posts that managers expect AI to fill.

What links them is that each represents a cost the industry spent years treating as external. Compute was someone else's subsidy. Water and power were someone else's infrastructure. Labour was a headcount line that AI was supposed to shrink for free. All three are now presenting their bills at once.

That changes the strategic picture more than another benchmark would. If capable models become commodity infrastructure — downloadable, free, running on your own hardware — then the durable advantage is not the model. It is the electricity contract, the water licence, the planning permission and the workforce. Britain has an AI growth strategy that scarcely acknowledges the first three and a labour market discovering the fourth.

Washington's answer so far is to make Chinese models harder to reach. Officials are exploring commerce rules or an executive order; OpenAI's policy head Dean Ball has predicted the administration will "create large amounts of regulatory risk around the use of open-weight Chinese models", warning that Beijing's approach leads to "full AI communism". The Pentagon's own chief technology officer publicly called him "the supreme village idiot" of the sector. A country arguing with itself in this register is not executing a strategy.

Import restrictions might buy time. They do not answer the underlying question, which is what customers are paying a premium for once adequate models are free. Neither does declaring growth zones without asking where the water comes from, nor assuming that AI absorbs the work of posts you have already cut.

The interesting phase of this technology is beginning now — not because the models got better, but because the invoices arrived.