A study linked electricity used by 403 U.S. hyperscale data centers to 25.7 million metric tons of carbon dioxide from May 2024 through April 2025, highlighting the scale of their demand on U.S. electricity supply.
Published Nov. 16 in Nature Sustainability, the study’s central estimate was 82 terawatt-hours (TWh), about 1.8% of U.S. electricity consumption during the period. Across scenarios for facility loads, estimated use ranged from 68 TWh to 99 TWh, while associated CO2 emissions ranged from 21 million to 31 million metric tons.
For comparison, World Bank data put Denmark’s 2024 CO2 emissions at 25.57 million metric tons. The comparison concerns CO2 alone, not Denmark’s total greenhouse-gas emissions. The study’s estimate is a modeled attribution to electricity generation; researchers did not measure emissions at the data-center sites.
Fossil fuels supplied about 54% of the generation attributed to the facilities, including 38% from natural gas and 15% from coal. Yet their electricity-weighted carbon intensity was 314 grams of CO2 per kilowatt-hour, below the U.S. total-output average of 348 grams. Virginia, Oregon, Ohio and Iowa together accounted for more than half of the electricity use studied.
The researchers identified facilities using satellite imagery and combined Environmental Protection Agency (EPA) power-plant emissions data with regional grid power mixes to attribute emissions to data-center electricity demand. The paper says its central estimate is two to three times an earlier estimate for 2018.
Available records did not let researchers isolate AI workloads from other computing at individual sites, and the analysis did not include every U.S. hyperscale facility. Lead author Gianluca Guidi, a postdoctoral fellow in Harvard’s Department of Biostatistics, called the study “only a snapshot” of a rapidly expanding industry. The Harvard release said construction and AI computing continued after April 2025, which may have raised electricity demand; the study gives no measured 2026 total.
The study also did not measure health effects. Corresponding author Francesca Dominici, Harvard’s Clarence James Gamble Professor of Biostatistics, Population, and Data Science, said decisions about siting and power sources should be “made with good data.” Researchers identified quantifying health impacts, including those linked to power-plant pollutants, as future work.



