US data centres used 82 TWh in a year, study estimates
A study led by Harvard estimates that 403 hyperscale data centres in the US used about 82 TWh in a year. That is 1.8% of national electricity, and the study estimates 25.7 million tonnes of carbon dioxide.

Key takeaways
- The study estimates 403 hyperscale US data centres used about 82 TWh in a year.
- Fossil fuels supplied 53.9% of the estimated electricity, according to the study.
- Virginia hosts 142 sites, while PJM covers about 40% of the estimated load.
- The figures are modelled estimates, and official surveys are still being developed.
In this article
What the study estimates
The research appeared in Nature Sustainability. It examined hyperscale data centres run by cloud providers, content providers or colocation operators. These sites have capacities in the tens of megawatts. A megawatt is a unit of power.
The study defines hyperscale sites as large facilities operated by cloud or content providers. It also includes colocation operators that serve such providers.
The authors began with 675 candidate sites. The list came from data vendor Baxtel. They confirmed 403 sites. They used satellite imagery and building-footprint maps. The confirmed sites had an average capacity of 39.9 megawatts. The median capacity was 36 megawatts.
How the electricity use was calculated
There is no official meter reading for each site. The study therefore used modelled estimates. It took each site's maximum capacity. It multiplied that by the number of hours in one year. It then applied an estimate of how hard each site runs.
For carbon emissions, each site was placed in one of 23 grid regions. The researchers used a regional average fuel mix. It came from the Environmental Protection Agency. They then multiplied and added the results.
These are estimates, not direct measurements. The study notes that such methods can limit how the figures are used in debates over data centre expansion.
Where the load is concentrated
The power use is not evenly spread. Virginia alone has 142 of the sites. They used 21.4 terawatt-hours, about a quarter of the total.
PJM, the grid operator for 13 states and Washington DC, covers 32.8 terawatt-hours. That is about 40% of the estimated total. It is the largest regional share in the study. PJM is the largest grid in the United States.
Virginia's draw is slightly below a 2018 estimate for all US hyperscale sites. That estimate was 22.85 terawatt-hours. This shows how concentrated the load has become.
Carbon output and fuel mix

The study estimates 25.7 million tonnes of carbon dioxide. It gives a range of 21 million to 31 million tonnes. The 2018 estimate was 10.5 million tonnes.
Fossil fuels supplied 53.9% of the estimated electricity. Gas provided 38.2%, while coal provided 15.2%. Renewables covered 25.3%, and nuclear supplied 20.9%.
Estimated carbon intensity was 314 grams of carbon dioxide per kilowatt-hour. A kilowatt-hour is a unit of electricity. The US grid average was 348 grams. That is about 10% lower.
The lower figure may reflect the location of big loads. PJM territory is nuclear-heavy, while Oregon is hydro-heavy. However, nearly 25% of the load is in areas with higher carbon intensity than the US average.
Caveats and next steps
The study covered a period ending in April 2025. Lead author Gianluca Guidi said the footprint may already be much larger in 2026. The authors point to rapid AI development as one reason the footprint may be larger. The median site capacity is also small compared with many planned campuses. Some new campuses are planned at hundreds or thousands of megawatts.
Power generated on site, such as gas plants, may be missing from grid data. The method uses 2023 plant data. It does not link individual plants to individual sites. It also excludes clean-power contracts and credits.
Official numbers may improve. The Energy Information Administration has completed pilot surveys. They covered data centre energy use. The agency is using them to design a wider survey.
The pilots began in March. They used web surveys and in-person interviews. They surveyed 196 companies about power sources, electricity consumption, servers and cooling systems. They did not cover all states.
Study author Francesca Dominici said the next step is to measure health effects.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



