Harvard Study Estimates 403 Giant US Data Centers Used 1.8 Percent of All US Power
A Harvard-led study in Nature Sustainability estimates 403 US hyperscale data centers used about 82 terawatt-hours of electricity in a year, with 54 percent from fossil fuels. The catch is that nobody can say how much of that was AI.
For all the shouting about AI data centers draining the grid, there has been surprisingly little hard, facility-level data on what the biggest ones actually consume. A Harvard-led team just put a number on it. In a study published Monday in Nature Sustainability, researchers estimate that 403 hyperscale data centers in the United States used about 82 terawatt-hours of electricity between May 2024 and April 2025. That works out to roughly 1.8 percent of all US electricity consumption.
That 82 figure is the central estimate, not a meter reading. Nobody outside the operators knows how hard each facility ran, so the authors modeled a range of load scenarios. Across them, the 403 sites used somewhere between 68 and 99 terawatt-hours. Harvard says the top of that range approaches what states like Indiana, Michigan and Tennessee use in a year. Ars Technica translates the range into between 1.5 and 2.2 percent of national demand.
The carbon side follows the same pattern. In the central scenario, the electricity behind these facilities was tied to about 25.7 million metric tons of carbon dioxide, inside a range of 21 to 31 million metric tons. Harvard says the central number is two to three times higher than an earlier estimate for 2018. The Verge says the paper compares it to Denmark's annual carbon footprint.
Roughly 54 percent of the power came from fossil fuels, according to the Harvard T.H. Chan School of Public Health, mainly natural gas at 38 percent and coal at 15 percent. The twist the angrier takes will skip is that the paper puts the electricity-weighted carbon intensity for these centers at about 314 grams of CO2 per kilowatt-hour, slightly below the US average of about 348. Ars reports the rest of the mix was 25 percent renewables and 21 percent nuclear. So the hyperscalers are not, on average, plugged into a dirtier grid than everyone else. They are plugged into the same grid, just at enormous scale, and Harvard says about a quarter of their electricity sat in regions dirtier than average.
Geography does a lot of the work. Harvard says just four states, Virginia, Oregon, Ohio and Iowa, accounted for more than half the electricity studied. The Verge, citing the study, reports that power in coal-heavy Wyoming carried roughly seven times the emissions per kilowatt-hour of hydro-rich Oregon.
"Two data centers consuming exactly the same amount of electricity can have dramatically different carbon footprints simply because of where they're located." Gianluca Guidi, lead author, Harvard T.H. Chan School of Public Health, to The Verge
What the number cannot tell you
This is not a measurement of AI. The authors say their data could not separate general-purpose computing from AI-specialized work inside each facility, so the honest framing is hyperscale data centers that power cloud services and AI, not AI itself. The abstract says the boom in these centers has been mainly driven by AI adoption, but the totals cover everything running inside them.
It is also not a complete count. Ars reports the team started with 675 suspected facilities from a commercial database and kept only the 403 it could verify, which leaves 272 out and makes the total conservative in one direction. The estimates also skip behind-the-meter power, such as the on-site gas turbines some companies have installed, according to Ars. And Harvard says the study did not measure health effects, which the team wants to tackle next.
The window also closed in April 2025. Guidi told Harvard the boom has not slowed since, and that the electricity use and carbon footprint the team reports "could already be substantially larger now in 2026." That is the authors' caution, not a new tally.
The bigger story is why academics had to piece this together from satellite imagery and EPA grid data at all. Guidi and Cornell's Fengqi You both told The Verge that operators do not disclose this information. The industry has been happy to talk about fewer secrets for host towns, yet the clearest national snapshot of its power use still came from outside. Until operators publish their meters, estimates like this are the best the public gets.

