US data centres could consume one-fifth of electricity by 2035

US data centres could be using about one-fifth of the country’s electricity by 2035, roughly four times today’s consumption, according to a BloombergNEF report cited by TechCrunch. BloombergNEF says the surge in AI computing will push data-centre capacity to nearly 200 gigawatts over the next decade, with nearly half devoted to AI training and inference.

Forecasts are being revised upward in a way that suggests the planning assumptions are moving slower than the construction crews. TechCrunch reports BloombergNEF’s 2035 estimate is 83% higher than its own forecast from December, while other forecasters have also raised their numbers: the Electric Power Research Institute has more than doubled its 2024 demand estimate, and S&P increased its forecast by more than a third between October and April. The revisions track rapid buildouts that are increasingly trying to connect to grids already described as strained.

The pressure is regional before it is national. BloombergNEF expects the PJM Interconnection grid—spanning Virginia to Illinois—to see 34% of its electricity used by data centres, while ERCOT in Texas would have to devote 22% of its generating capacity to data centres. PJM, which already hosts many data centres, has struggled with connection requests from both large generators and large loads; it paused applications for new sources to connect for four years and reopened the queue in April, TechCrunch reports. American Electric Power has threatened to pull out of PJM, and the supply-demand imbalance has been associated with a 76% rise in electricity prices over the past year.

The report also sketches where the chips, and therefore the power demand, are expected to sit. By 2033, BloombergNEF estimates the US will host 64% of AI chips by power demand, even as data-centre growth continues worldwide. Under aggressive AI adoption, global data centres would add 1,935 terawatt-hours of new electricity demand by 2033—nearly as much as India’s annual electricity use, according to the figures cited by TechCrunch.

For utilities and regulators, the near-term question is less about whether AI is valuable and more about who gets priority access to constrained infrastructure, and how the cost of upgrades is allocated when a single new customer can resemble a city.

In PJM’s most recent capacity auction, data centres accounted for 38% of charges, even as the region’s connection queues and power prices tightened.