Skip to content

Research

Powering 403 U.S. data centers produced an estimated 25.7 million metric tons of CO₂

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. A Harvard-led study estimated the electricity used by 403 U.S. hyperscale data centers between May 2024 and April 2025.
  2. The central estimate was about 82 TWh and 25.7 million metric tons of attributable CO₂.
  3. The emissions rate varied by grid. The study covers all power use at the sampled facilities, including computing beyond AI.
PracticalBeginnersDevelopers
Low data center buildings and transmission equipment at dawn
AI-generated conceptual image: Data center buildings and transmission equipment illustrate a campus’s connection to the grid. · AI Ping

A Harvard-led study estimates that electricity used by 403 U.S. hyperscale data centers was associated with 25.7 million metric tons of carbon dioxide over one year. The paper, published November 16 in Nature Sustainability, puts demand at about 82 terawatt-hours between May 2024 and April 2025. That represents roughly 1.8 percent of all U.S. electricity consumption during the period.

Where the facilities drew their power mattered. The researchers linked estimated demand at each site to generation in its regional grid. Virginia, Oregon, Ohio and Iowa used a combined 48.5 TWh, or about 59 percent of the total. The states ranked differently by emissions than by electricity use.

What the researchers counted

Hyperscale facilities support everything from cloud computing to training large language models and serving requests to AI products. The researchers estimated how much electricity the facilities drew as a whole. Harvard’s public health school said the data could not separate AI workloads from ordinary computing at the sites. The total therefore covers the electricity associated with both kinds of work at the facilities in the study.

The team used satellite imagery and other sources to identify sites operating during the study period. It estimated their electricity use from power capacity and assumptions about how much of that capacity was in use. Since actual electricity use at every facility was unavailable, the authors reported several scenarios with different loads.

The estimates ranged from about 68 to 99 TWh of electricity and 21 million to 31 million metric tons of CO₂. The figures of 82 TWh and 25.7 million metric tons are from the same middle scenario. The authors modeled low, central and high loads at 48, 58 and 70 percent of capacity, respectively.

An aisle lined with server cabinets
AI-generated conceptual image: An aisle of server cabinets illustrates computing equipment that draws power at a large facility. · AI Ping

Oregon used more electricity, with lower emissions than Iowa

The researchers used U.S. Environmental Protection Agency data to link demand to power generation. For each grid area, they divided the CO₂ reported by power plants by all the net electricity generated there, including electricity from nuclear plants and renewables. The calculation uses annual grid averages. It assigns emissions to a site’s estimated electricity demand without tracing deliveries from individual power plants.

The authors’ published state table puts Oregon’s study sites at 11.1 TWh and about 1.33 million metric tons of CO₂. Iowa’s sites were estimated at 7.2 TWh and 3.19 million metric tons. The electricity attributed to the facilities averaged about 119 grams of CO₂ per kilowatt-hour in Oregon and 445 grams in Iowa, reversing the states’ order on emissions.

Across all study sites, the average was about 314 grams per kilowatt-hour, compared with 348 grams for the United States under the same accounting method. Some grid areas were above the national figure. Fossil fuels supplied about 54 percent of the generation attributed to the facilities. The overall rate gives more weight to areas where these data centers used more power.

Wind turbines and a gas plant in a landscape crossed by transmission lines
AI-generated conceptual image: Wind turbines and a gas plant in a shared landscape illustrate different sources of electricity in a regional grid. · AI Ping

The published paper changed the accounting method

A June preprint from the same team reported a range of roughly 37 million to 54 million metric tons. In its revision notes, the team says it switched the denominator for its main emissions estimate from power produced by combustion to all net electricity generation. Including sources such as wind and nuclear power changed the rate applied to demand. The estimates describe the same period, so the revision records a change in accounting.

The authors released their scenario totals and aggregates by state and regional grid. Those materials give readers a way to compare electricity use and attributed emissions when assessing where new data centers might get their power. Individual site names, precise locations and certain commercial details are restricted by a data agreement. The public files provide totals at the state and grid-area level.

Sources