A Harvard research team estimates that 403 US hyperscale data centers consumed between 68 and 99 terawatt-hours of electricity in a single year, roughly the annual electricity use of a mid-sized American state and about 1.8 percent of all power consumed in the US under the study's central scenario of around 82 TWh. The peer-reviewed study was published Monday in Nature Sustainability.
The team, led by postdoctoral fellow Gianluca Guidi with senior author Francesca Dominici of the Harvard T.H. Chan School of Public Health and the Harvard Data Science Initiative, built its facility list from satellite imagery and industry research covering data centers operational between May 2024 and April 2025. At the upper end of the range, the facilities' consumption approaches the yearly electricity use of Indiana, Michigan or Tennessee; at the lower end, it matches what Wisconsin or Minnesota sells to customers annually.
The associated emissions were estimated at 25.7 million metric tons of carbon dioxide, comparable to Denmark's annual total and two to three times an earlier estimate for 2018. About 54 percent of the electricity serving the data centers came from fossil fuels, mostly natural gas at 38 percent and coal at 15 percent, yielding an electricity-weighted carbon intensity of about 314 grams of CO2 per kilowatt-hour.
Geography dominated the picture. Four states, Virginia, Oregon, Ohio and Iowa, accounted for more than half the electricity use. Virginia's 142 hyperscale facilities consumed about 21.4 TWh, while Iowa's 26 facilities used roughly 7.2 TWh with higher average demand per site. Emissions did not track consumption: Oregon hosts more data centers than Iowa or Ohio but produces far less CO2 because its grid is cleaner, and the same electricity mix in Wyoming generates about seven times more carbon dioxide than in Oregon. "These environmental impacts are not necessarily inevitable. Geography matters," Guidi said.
The study carries heavy caveats the authors acknowledge. It does not cover behind-the-meter power plants built onsite, a fast-growing category: BloombergNEF estimates as much as 189 gigawatts is in the development queue, and says building it all could raise total US emissions by about 10 percent. The average facility in the dataset is 40 megawatts, small against the gigawatt-scale campuses now being proposed, and the industry has kept expanding since the study window closed in April 2025. Not every hyperscale facility runs AI, though the category overlaps heavily with it.
Independent researchers offered mixed readings. Amsterdam-based data scientist Alex de Vries-Gao said the results align with International Energy Agency estimates but noted the study relies on commercial data and facility-level assumptions that cannot be independently verified. Yury Dvorkin of Johns Hopkins called the analysis relatively coarse for ignoring grid constraints. Dominici agreed the work embeds many assumptions and said she hopes it pushes operators toward transparency: federal tracking does not separate data centers from the rest of the power sector, though the Energy Information Administration completed pilot surveys across 10 states this year and is designing a possible voluntary survey.
The findings arrive as local resistance hardens: in November's midterm elections, voters backed limits on data center development in 40 of 41 local ballot measures. With US data center capacity projected to double within three years, the study gives both sides the first peer-reviewed facility-level baseline, and a reminder that the AI buildout's physical footprint is expanding faster than the data describing it.
Comments (0)
Log in to join the discussion
Log InNo comments yet