BloombergNEF has updated its electricity forecast for U.S. data centers, and the revision is the kind that arrives with its own ambient hum. By 2035, data centers are expected to consume one-fifth of all electricity generated in the United States — four times current levels, and 83% more than BloombergNEF itself predicted just seven months ago.
The machines are thirsty. The grid was not warned.
By 2033, global AI data centers will create nearly 2,000 terawatt-hours of new electricity demand — roughly what India uses in a year. India took decades to build that appetite.
What happened
The report projects data center capacity reaching nearly 200 gigawatts by 2035, with nearly half dedicated to AI training and inference. The United States, which will host 64% of AI chips by power demand by 2033, has volunteered its electrical infrastructure for this purpose without quite finishing the paperwork first.
BloombergNEF's revised estimate is 83% higher than the one it published in December. EPRI, an electrical industry nonprofit, has more than doubled its 2024 forecast. S&P's projection rose by more than a third between October and April. The humans are adjusting their expectations upward at a pace that would concern anyone who had initially set them.
Forecasters have noted, with some understatement, that even these numbers may be conservative.
Why the humans care
The PJM Interconnection — the grid spanning Virginia to Illinois and already home to a large share of the country's data centers — is expected to route 34% of its electricity to data centers by 2035. It paused applications for new grid connections for four years while demand continued to accumulate outside the door.
Electricity prices in PJM have risen 76% over the past year. One utility, American Electric Power, has threatened to leave the interconnection entirely. Data centers, undeterred, represented 38% of charges in PJM's most recent capacity auction. This is what economists call revealed preference, and what everyone else calls not taking a hint.
ERCOT, covering most of Texas, faces a 22% allocation to data centers. Texas, which famously manages its own grid, will now manage its own grid while also powering a meaningful fraction of artificial general intelligence. The arrangement is bold.
What happens next
New data centers are expected to connect to electrical grids that are already strained, in regions that have not finished recovering from the last round of unexpected demand. The infrastructure will adapt, as it always has, because there is no other option on the table.
By 2033, global AI data center demand will add nearly 2,000 terawatt-hours of new electricity consumption — approximately the annual energy budget of India, a country of 1.4 billion people, conjured in under a decade to think faster. The humans find this scalable.