Power, not chips: grid connections are what hold data centers back

Thesis: accelerators are no longer the bottleneck for AI growth. Grid connections and fuel are. That changes the geography of the industry, the rules for where projects can be built, and energy policy at the same time in Europe, China and the United States.

The clearest example comes from the United Kingdom. The AI supercomputer announced in 2025 as the country's largest, in Lawford in Essex, will not start next year as planned. Its launch may slip to the middle of the 2030s. The distribution network operator UK Power Networks told the investor Nscale that it cannot deliver enough power before the start or the middle of the 2030s. The project needs up to 90 MW, roughly the power used by 315,000 households. In June 2026, 315 data centers with a combined demand of 73 GW were waiting for a connection in the UK, against peak national demand of 45 GW. The regulator Ofgem has warned about the growing backlog.

This is not a British peculiarity. Oracle filed a force majeure notice with an infrastructure supplier for its Project Jupiter in New Mexico. In practice, that means accepting the schedule cannot be met. Oracle's capital expenditure in the first quarter was 28.5 billion dollars, against 8.5 billion a year earlier, and the forecast for fiscal 2027 reaches 90 to 95 billion. An RBC analyst called the announcement a signal of growing delay risk, and the share price fell by around 4 percent. Even a company with cash and signed contracts cannot buy power that physically does not exist.

In the United States the regulatory response is already visible. New Jersey fined a data center operator 1.1 million dollars after satellite and drone images revealed 62 gas generators, 45 of them running, with no permits and above the 37 kW emissions threshold, at installation capacities of around 1,982 kW. The fine concerns a breach of the Clean Air Act. It also reveals a wider pattern: when the grid fails, investors reach for their own power sources, including where the law did not foresee it.

China answers the same problem with a different logic, that of central planning. Energy consumption by computing centers reached 170 TWh in 2025, at the world's second-largest scale of computing capacity. Eight computing hubs grew at around 40 percent a year for three years. The phrase "coordination of energy and computing" entered a government work report for the first time, and the China Electricity Council launched a nationwide index of energy consumption by computing centers. The state grid announces 4 trillion yuan of capital expenditure in the five-year plan, of which more than 700 billion in 2026. The electrification rate of final energy consumption is to rise from 29.5 percent in 2025 to 35 percent in 2030.

Putting these three models side by side gives one answer: competition over AI is now competition over megawatts. A country that cannot quickly connect gigawatts and has no plan to generate them will not have models, no matter how many graphics cards it buys. For Europe, the signal is that a debate about technological sovereignty without a debate about transmission grids and power sources is a debate about something that cannot be built.