A 2026 study in Nature Sustainability estimated that 403 identified U.S. hyperscale data centers used 68–99 terawatt-hours (TWh) of electricity during May 2024–April 2025, equal to 1.5–2.2% of total U.S. electricity use. Ars Technica reported the estimate. Researchers could not separate AI computing from traditional data-center services, and some included facilities may not be AI data centers.
What the electricity estimate covers
The figure is a modeled estimate based on facility capacity and assumed utilization, not a direct meter reading from every site. It covers the 403 facilities retained for the study, rather than all U.S. data centers.
The study began with 675 candidate facilities. Researchers retained 403 after excluding sites without reported electrical capacity or verifiable characteristics such as a suitably large building, cooling equipment and electrical connections.
Those facilities averaged about 40 megawatts (MW) of electrical capacity. MW measures power capacity; TWh measures energy used over time.
Capacity describes what a facility can draw; TWh records energy use over a period.
Estimated emissions and electricity sources
Ars Technica reported an estimated 21–31 million metric tons of carbon dioxide (CO₂) associated with generating the electricity, equivalent to about 0.5% of U.S. fossil-fuel CO₂ emissions. The emissions calculation applied the carbon intensity of the electricity mix in each facility’s grid region to its estimated electricity use.
The reported grid-supply mix was 54% fossil fuels, predominantly natural gas; 25% renewables; and 21% nuclear. The emissions estimate excludes behind-the-meter generation, or electricity generated on-site rather than supplied through the grid.
What readers can access from the study
The Nature Sustainability data-availability statement says facility-level data are restricted under a data-use agreement. It describes public aggregate exports by state and balancing authority, along with the analysis code; facility identifiers, addresses, coordinates and commercially sensitive attributes are not publicly released. (Nature Sustainability study).