
403 U.S. hyperscale data centers used 82 TWh in one year
Harvard-led researchers estimate 403 U.S. hyperscale data centers used 82 TWh from May 2024 to April 2025, with 25.7 million metric tons of CO2 linked to their grid mix.
- By
- Grace Kim
- Filed
- Read
- 3 min read
By Grace Kim
403 U.S. hyperscale data centers consumed an estimated 82 terawatt-hours of electricity from May 2024 through April 2025, according to a Harvard-led peer-reviewed study published Monday in Nature Sustainability. The load approached Michigan's annual electricity consumption and represented about 1.8% of U.S. power generation.
The researchers estimated 25.7 million metric tons of carbon dioxide from the associated electricity mix, roughly matching Denmark's annual emissions. Fossil fuels supplied 54% of the facilities' modeled electricity. Harvard T.H. Chan School of Public Health researchers led the study, with biostatistician Francesca Dominici as senior author.
What did the study measure?
The team described its work as the first peer-reviewed, facility-level estimate of data-center energy use and emissions since the AI boom began. The May-to-April study period captured the early ramp in AI spending. Its estimated emissions were two to three times those in a 2018 peer-reviewed estimate.
The Harvard team compiled facility-level information on all 403 sites. It estimated electricity use at each site, then traced that demand to the plants serving the local grid. Researchers used plant-level records from the Environmental Protection Agency's eGRID database.
A June preprint had estimated 68 TWh to 99 TWh, depending on how heavily the facilities operated, with 82 TWh as the central case. That version put average carbon intensity at about 545 grams of CO2 per kilowatt-hour, roughly 48% above the national grid average.
The published paper put central emissions at 25.7 million tons, below the preprint's central estimate. MLQ.ai could not see the final paper's revised method.
The estimates combine national scale with local carbon burden. That combination puts local generation mix at the center of the study's siting analysis.
Where does electricity load concentrate?
Most of the energy use sits in Iowa, Ohio, Oregon and Virginia. About one-quarter of demand falls in grid regions dirtier than the national average. The four-state concentration makes local generation mix more consequential than the national share alone suggests.
In Ohio, about 80% of electricity comes from fossil fuels, including about 20% from coal, according to Energy Information Administration figures cited by the researchers.
Can cleaner power change the siting result?
Alongside the paper, the research group released a site-selection explorer. It scores candidate sites for a hypothetical 100-megawatt data center from zero, meaning the lowest impact, to 100, meaning the highest.
The tool traces each candidate to its supplying power plants and maps downwind fine-particle pollution by census tract. A site in Loudoun County, Virginia, scored 76.
Its score fell to 54 if the data center ran on round-the-clock clean power. Francesca Dominici, the study's senior author, said: “Decisions about where data centers go and how they're powered are critical environmental and public health decisions.”
The explorer turns the study's grid estimates into a site-level environmental and public-health screen.
What could this mean for AI capacity?
Power for AI campuses has already become a political issue. Meta has said midterm backlash will make new U.S. data centers harder to build. OpenAI has promised to pay its campuses' power costs.
Those positions show how permitting and electricity costs can shape expansion plans. For developers, the siting tool compares candidate campuses on power source and modeled public-health exposure.
Goldman Sachs also raised its 2030 global data-center capacity forecast to 273 GW that month. As the 273-GW forecast approaches, grid mix, round-the-clock clean supply, permitting and electricity costs will shape where operators can build new AI capacity at scale in the United States.
Source: mlq.ai
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,035 articles
Related articles
eia-sees-u-s-power-records-through-as-commercial-use-grows
EIA Sees U.S. Power Records Through 2027 as Commercial Use Grows
datacenters-will-eat-3-3-of-world-power-gartner-says-06f3688c
Datacenters Will Eat 3.3% of World Power by 2030, Gartner Says
weeteq-targets-the-electricity-ai-servers-waste-before-it-is-used-01cce88c
Weeteq Targets the Electricity AI Servers Waste Before It Is Used
2-1-billion-metric-tons-u-s-power-co2-could-match-mexico
2.1 Billion Metric Tons: U.S. Power CO2 Could Match Mexico



