A study published in Nature Sustainability is the first to count electricity use by US hyperscale data centers one facility at a time. Its finding: from May 2024 to April 2025, the 403 hyperscale data centers in the United States used roughly 68 to 99 terawatt-hours of electricity in a year. That is 1.5% to 2.2% of all US electricity consumption, comparable to the annual use of a single state such as Wisconsin or Indiana.
How the count was done
The team started from a commercial database of 675 facilities, built from news reports, company statements and satellite imagery. Any site with no public power capacity figure, or one that could not be verified as a large building with the matching cooling and power equipment, was dropped. That left 403 facilities with an average capacity of about 40 megawatts.
Capacity is not the same as actual consumption. Drawing on other research, the team assumed a range of utilization rates: about 48% of capacity for conventional data centers and about 70% for AI training workloads.
The numbers can be roughly reproduced. 403 facilities at 40 MW come to about 16 gigawatts of capacity, or around 141 TWh over a year at full load. Applying 48% and 70% utilization lands in the 68 to 99 TWh range. The upper and lower bounds differ by more than 40%, mostly because of that single utilization assumption.
Set against total US data center demand
An earlier Lawrence Berkeley National Laboratory report estimated that, excluding cryptocurrency mining, all US data centers used 76 TWh in 2018 and 192 TWh in 2024, a 2.5-fold increase in six years. By this study's figures, hyperscale facilities account for roughly one-third to one-half of all data center electricity use.
The two reports do not cover exactly the same years, so that share is best read as an order-of-magnitude guide. Many of the 272 excluded facilities are also believed to be large data centers, and press coverage describes the estimate as "very conservative."
Where the power comes from
The study estimated emissions from the grid mix in each data center's region. About 54% of the electricity these facilities use comes from fossil fuels, mainly natural gas, with 25% from renewables and 21% from nuclear, only slightly cleaner than the US average. The corresponding emissions are 21 million to 31 million tonnes of CO2 a year, about 0.5% of US fossil fuel emissions.
That figure does not include power the data centers generate themselves. During the study period, xAI's data center in Tennessee was running on gas turbines. According to the energy data firm Cleanview, 116 on-site power generation projects for data centers are now under construction or planned. Other research has found that such on-site generators emit 60% to 80% more greenhouse gas per kilowatt-hour than utility power plants.
Gaps the study acknowledges
The study cannot separate how much electricity went to AI computing and how much to conventional data services, and some of the 403 facilities may not run AI at all. The authors present the results as a baseline: repeating the count each year with the same method would let the increment from AI expansion be isolated.
This site has previously reported on international agencies' estimates of global AI data center electricity use. Those are global, model-based figures, while this study counts facilities one by one within a single country, so the two sets of numbers cannot be added together or compared directly.
Sources: Nature Sustainability paper, Ars Technica, CocoLoop, Lawrence Berkeley National Laboratory report, Cleanview. Electricity figures are annual estimates for the 403 facilities based on a range of utilization rates; emissions exclude on-site generation.