403 U.S. data centers used as much electricity as Michigan, Harvard-led study estimates
Researchers reconstructed facilities’ electricity use from commercial records and satellite imagery. Their estimate excludes unverified sites and facilities powered by on-site generators.
Because operators generally do not disclose facility-level power use, a Harvard-led team combined satellite imagery, modeled demand and regional grid data to estimate the footprint of 403 grid-powered hyperscale data centers from May 2024 to April 2025. The facilities used nearly 2% of U.S. electricity and produced an estimated 25.7 million metric tons of CO₂, about 0.5% of U.S. emissions. This is a verified sample, not an AI-only or industrywide tally; its limits make transparency more consequential as newer, much larger campuses come online.
01
Fossil fuels supplied 54% of the sample’s electricity, but emissions varied by location; Oregon’s lower-carbon supply reduced its emissions per unit of power.
02
Virginia, Oregon, Ohio and Iowa used more than half the sample’s electricity; Virginia, Iowa and Ohio produced about half its estimated carbon pollution.
03
The researchers excluded unverified sites and facilities powered by on-site generators; dataset licensing also kept facility names, addresses and owners from being published.
A Harvard-led study has put a number on a largely hidden part of America’s computing infrastructure: 403 large data centers produced an estimated 25.7 million metric tons of carbon dioxide emissions over one year. Published in Nature Sustainability, the research examined May 2024 through April 2025, when those facilities consumed almost 2% of U.S. electricity.
Their annual electricity consumption matched Michigan’s, according to WIRED. The sample covers hyperscale data centers—the largest types of facilities—but it is not a measurement of AI alone. Although such sites often handle AI training or run models for users, not every facility studied was being used for AI.
From satellite images to an emissions estimate
The team, led by Harvard researcher Gianluca Guidi, could not simply download a public inventory and read each facility’s meter. Operators and utilities do not have to disclose facility-specific electricity consumption publicly, and companies often treat those figures as proprietary business information.
Instead, the researchers reconstructed the infrastructure and its likely power demand. Their method linked three kinds of evidence, moving from the buildings themselves to the electricity systems supplying them:
Locate and verify facilities. The team started with a commercial dataset, then examined building footprints and satellite imagery to confirm data-center locations and sizes. There was no public list available for the researchers to use.
Estimate electricity use. Researchers modeled the facilities’ energy consumption under different scenarios, rather than relying on a complete set of publicly disclosed operating records for the sites.
Connect demand to the power supply. They calculated emissions using the mix of power plants supplying each regional grid. That step tied the estimated electricity consumption to the sources generating it.
Even the starting inventory carried a transparency limit. Licensing restrictions on the commercial dataset prevented the researchers—and The New York Times—from publishing facilities’ addresses, names and owners. The work could estimate the collective footprint without making that underlying list public.
The grid changes the carbon cost
Location made a substantial difference. Virginia, Oregon, Ohio and Iowa together accounted for more than half of the sample’s electricity consumption. But the emissions ranking was different: Virginia, Iowa and Ohio alone accounted for about half of estimated carbon pollution, with Virginia responsible for more than a quarter.
Oregon illustrates why electricity demand and carbon emissions are not interchangeable measures. Its data centers emitted less carbon dioxide per unit of electricity because a larger share of their supply came from sources other than fossil fuels. Across the full sample, natural gas and coal supplied most of the fossil-generated electricity.
The sample’s fossil-fuel footprint
54%Electricity from fossil fuels
Fossil fuels supplied 54% of electricity across the 403 facilities studied.
About 0.5%Share of U.S. CO₂ emissions
The study attributed about 0.5% of U.S. carbon dioxide emissions to the facilities.
A baseline, not a complete inventory
The estimate leaves out hundreds of possible data centers that the researchers could not sufficiently verify. It also excludes facilities drawing electricity from on-site generators rather than the grid. The total therefore describes the verified, grid-powered sample, not the entire U.S. data-center industry.
It also captures a smaller scale of development than some projects now being built. The average facility studied had about 40 megawatts of electrical capacity. Developers are building campuses measured in hundreds or thousands of megawatts, Guidi told WIRED, substantially larger than the typical site in this dataset.
The industry has grown since the measurement period, too. At least 250 more data centers have come online since it ended, according to Cleanview data cited by The New York Times. Those additions sit outside this study’s yearlong snapshot.
Guidi’s proposed remedy is direct disclosure: he told The New York Times that the federal government should require large data-center operators to publish electricity use and emissions. For communities trying to understand a facility’s environmental impact, his argument is that obtaining those basics should not require a Harvard team analyzing satellite images.
Sources
nytimes.comCarbon Emissions of Biggest U.S. Data Centers Already Rival Entire Countries, Study Reveals
wired.comJust 400 US Data Centers Use More Electricity Than Michigan
Reader comments
Newest comments first. Replies stay oldest first.