Can nuclear power really power AI data centres?

With micro-reactors that can be deployed within eighteen months and conventional reactors that take more than ten years to build, the timescales for nuclear power are poorly aligned with those for AI.

As AI facilities scale up from 25 MW to complexes in the gigawatt range, the idea of directly linking them to nuclear power is no longer merely a provocative suggestion. One megawatt powers around a thousand households, one gigawatt nearly a million: the demand forecast by major operators is now on a par with that of a city.

The problem lies with the timeframe. Micro-reactors, ranging from 1 to 20 MW, could be deployed within eighteen months. Small modular reactors, ranging from 20 to 300 MW, take five to seven years. A conventional reactor unit of over 300 MW takes more than ten years to build. Yet the growth in demand driven by AI is set to occur precisely over the next two or three years — a window during which only the first category can contribute, and on a scale that bears no relation to the actual needs.

The most high-profile project remains the recommissioning of the Three Mile Island power station by Constellation Energy to supply Microsoft’s data centres, expected in 2027: this involves restarting an existing facility, not building a new one. There are still numerous conditions that must be met for the sector to deliver on its promises: initial deployments delivered within the announced budgets, streamlined authorisation procedures, robust safety protocols, mature supply chains and acceptance by the local communities concerned. Industry practitioners are anticipating a gradual roll-out, starting with micro-reactors of around ten megawatts.