KEY POINTS
- Researchers identified 403 US hyperscale data centers by satellite and estimated they used about 82 TWh between May 2024 and April 2025.
- The estimated emissions were 25.7 million metric tons of CO2, with about 54 percent of the power likely coming from gas and coal.
- Local grids matter: Wyoming's electricity produced roughly seven times more carbon per kilowatt-hour than Oregon's.
Hyperscale data centers in the United States consumed an estimated 82 terawatt-hours of electricity in one year and were responsible for about 25.7 million metric tons of carbon dioxide, according to a new paper in Nature Sustainability. The Verge reported on the findings.
The researchers focused on hyperscale facilities, the very large buildings covering millions of square feet that developers are rushing to construct, largely to support generative AI. They used satellite imagery to find 403 such sites and estimated each one's power use from its electrical capacity.
To work out emissions, the team used Environmental Protection Agency data to approximate what each site would have emitted if it drew from its local grid between May 2024 and April 2025. The 82 terawatt-hours they calculated equal about 1.8 percent of all electricity used in the US.
John Sterman, director of the MIT System Dynamics group, said that figure exceeds the yearly consumption of about 148 countries, including Austria, Israel and Venezuela. The emissions are comparable to Denmark's annual footprint, and Sterman said they surpass the totals of about 113 nations. About 54 percent of the electricity likely came from gas and coal. Sterman was not part of the study and said he used Claude to help make the country comparisons.
Location had a large effect on the results. Generating power in coal-heavy Wyoming produced roughly seven times more emissions per kilowatt-hour than in Oregon, which has plenty of hydropower. Lead author Gianluca Guidi of the Harvard T.H. Chan School of Public Health said identical data centers can have very different footprints depending on where they sit, and he described the finding as encouraging because the impacts are not inevitable.
The paper has limits. It does not cover local air pollution from new gas infrastructure built for data center demand, nor gas turbines that some AI companies have installed to generate their own power off the grid. Hyperscale sites also include more than AI workloads, and they are only a portion of the thousands of data centers in the country.
Fengqi You of Cornell University, who was not involved, warned of long-term carbon lock-in, since power plants built for this demand could run for decades. Sterman said policymakers should require operators to bring new clean energy online, adding that emitted CO2 cannot be taken back.
Both Guidi and You criticized the need for such elaborate estimates, since operators do not disclose their energy use. Guidi's team is building a website to track the environmental impact of individual data centers, and a bipartisan bill in Congress could require wider disclosure of their electricity consumption. Former EPA administrator Gina McCarthy called the study an important step toward reliable, independent data.