America's data centres are about to become one of its biggest power users. A new forecast from research group BloombergNEF projects they will consume about 20% of the country's electricity by 2035, up from 5.9% today and an estimated 12% in 2030. That is roughly four times their current share, and it is being driven almost entirely by artificial intelligence.
A staggering jump
To meet that demand, data-centre power needs are projected to reach 194 gigawatts by 2035, an 83% jump from the firm's own forecast of 106 gigawatts just last December. For scale, a single gigawatt is about the output of one traditional nuclear reactor. The pipeline behind the surge includes projects like Microsoft's planned 2-gigawatt facility in Pecos, Texas, and Meta's data centre of up to 5 gigawatts in Richland Parish, Louisiana. As one BloombergNEF analyst put it: "Every coal plant, every gas plant, every solar farm in the US, one unit of energy out of five generated by them is going to data centres."
The grid cannot keep up
The trouble is supply. After two decades of flat demand, grids are now straining to keep pace, and BloombergNEF expects most new data centres to connect to grids that are already stretched. It warns of a roughly 19-gigawatt shortfall by 2035, even if the industry sustains its record connection pace of 7.1 gigawatts a year. The pressure is worst where data centres cluster: in the grid stretching from Virginia to Illinois, 34% of electricity is expected to go to data centres, and in Texas, 22%. Developers are turning to on-site gas plants, and running into permitting freezes and local opposition.
The strain is not just American. Globally, if AI adoption keeps climbing, data centres could add 1,935 terawatt-hours of new electricity demand by 2033, almost as much as India uses in a year. Other forecasters, from the industry nonprofit EPRI to Bank of America, have been revising upward too. Bank of America reckons the US needs more than 230 gigawatts of new generating capacity within five years, while regulated utilities are on track to add only about 93 gigawatts, a sign the AI power crunch is only starting to bite.
