US power companies face a decision about how much new generation and grid capacity to build for surging datacentre demand, given that some planned AI facilities may never be completed if the AI boom cools. Consulting firm McKinsey & Company argues the bigger danger is under-building rather than over-building, since even if AI compute demand slows, the infrastructure created to serve it would not go to waste and could instead support other demand while improving overall grid reliability.

McKinsey estimates datacentres will drive around 75% of US power demand growth over the next decade, requiring roughly 30 GW of extra capacity annually. The firm projects a nationwide shortfall of 30 to 55 GW by 2030, as ageing coal and gas plants retire faster than new capacity comes online, and notes that 71% of organisations report poor AI implementation outcomes, hinting at bubble-like conditions. To cope with slow grid connections, many datacentre operators are turning to on-site power such as gas turbines, fuel cells and batteries, with McKinsey expecting solar-plus-storage to play a growing role alongside gas beyond 2030.