Harvard study: 403 US data centers used up to as much electricity as Michigan

In short
  • A Harvard-led study of 403 US hyperscale data centers found they used 68 to 99 terawatt-hours of electricity in a year.
  • The upper figure is close to what states such as Michigan use; about 54% of the power came from fossil fuels.
  • The study cannot split AI from other use, and Harvard's researchers say today's numbers are probably higher.
  • 68-99 TWhPower used by 403 data centersMay 2024 to April 2025
  • 1.8%Share of all US electricityDuring the study period
  • 25.7 MtEstimated CO2 emittedMillion metric tons a year
  • 54%Power from fossil fuelsGas 38%, coal 15%
Poster in green with a science symbol and the words 'Up to as much power as Michigan', on data center electricity use
Photo: TechSpotlight News on Cover card

Every time you stream a film, back up photos or ask a chatbot a question, a building somewhere hums a little louder. Now researchers have put numbers on the biggest of those buildings. A Harvard-led study finds that 403 hyperscale data centers in the US used between 68 and 99 terawatt-hours of electricity in a year.

WIRED reports that no rule makes operators or utilities publish each site’s power use, so data center electricity consumption has been hard to see. The upper figure approaches what states such as Indiana, Michigan and Tennessee use annually, according to Harvard’s T.H. Chan School of Public Health. Harvard’s researchers say the industry is growing fast, and decisions on where to build and how to power sites are being made now. A terawatt-hour is a trillion watt-hours.

Data center electricity consumption, in Harvard’s numbers

The team, led by Gianluca Guidi and Francesca Dominici and published in Nature Sustainability, used satellite images to list 403 hyperscale data centers across the US. These are the largest class of data centers, supporting cloud computing and AI. For each site, the researchers estimated yearly electricity use between May 2024 and April 2025 in several scenarios for how hard the sites were running.

The total came to 68 to 99 TWh, about 1.8% of all US electricity use. They then used Environmental Protection Agency data to link each site to the power plants on its regional grid. The result was about 25.7 million metric tons of CO2, two to three times an earlier estimate for 2018. TIME compares that with the yearly emissions of Denmark.

About 54% of the electricity came from fossil fuels: natural gas 38% and coal 15%. Roughly a quarter was in grid regions dirtier than average.

Where the power goes, and why geography matters

Four states, Virginia, Oregon, Ohio and Iowa, accounted for more than half of the electricity use studied. TIME reports that Virginia has 142 of these facilities using about 21.4 TWh, while Iowa’s 26 use about 7.2 TWh, so Iowa’s average site is substantially larger.

Location also changes the carbon bill. Guidi says the same amount of electricity in Wyoming produces about seven times more CO2 than in Oregon, because of the regional power mix. “Geography matters,” he said, as TIME reports.

What the study cannot tell us yet

The team could not cover every US hyperscale site. They could not separate AI work from other computing. And they did not measure health effects, which they plan to study next.

Dominici, who directs the Harvard Data Science Initiative, notes that fossil-fuel plants also release fine particles and other pollutants linked to heart and lung disease. People living downwind may carry health costs, even if they live far from a data center.

WIRED adds two caveats. The average site in the study is 40 megawatts (a megawatt is a million watts), small next to campuses now planned at hundreds or thousands. And the study leaves out behind-the-meter power plants built on site. Jon Koomey, an energy researcher who was not part of the study, calls that a “wild card”. Harvard’s release quotes Guidi saying the real footprint “could already be substantially larger now in 2026”.

If you pay a power bill or use AI tools

The study does not measure bills, so we will not guess at them. If you live near a proposed data center, you can now ask which grid would power it, and how clean that grid is. The study shows why it matters: the same electricity can carry very different carbon costs, as Wyoming’s grid against Oregon’s shows.

The AI tools we use daily sit on physical buildings and power plants. WIRED reports that the US Energy Information Administration ran pilot surveys in 10 states this year and is designing a possible voluntary survey.

A useful measurement, with a few gaps

The good news is that this is peer-reviewed work that puts careful estimates where there were mostly statements. Dominici told TIME there was very little peer-reviewed data when the work began. The study also shows that the carbon cost is not fixed: cleaner grids, like Oregon’s, make a real difference.

The concern is that the numbers are already a year and a half old in a fast-moving industry. The split between AI and other use remains unknown, and on-site power plants sit outside the count. In our reading, the study is a careful estimate that is probably a floor, not a ceiling, since it leaves out some sites and on-site plants. It is a sound starting point for builders and regulators who want to decide with facts.

The next figure to watch is the health study the team plans, and whether the EIA survey turns pilots into regular public numbers.

Prefer to watch? Follow us on YouTube

The tech stories that matter, explained in plain words. Subscribe so you never miss one.

Subscribe

Quick answers

How much electricity do US data centers use?

A Harvard-led study of 403 hyperscale data centers estimates 68 to 99 terawatt-hours between May 2024 and April 2025. That is about 1.8% of all US electricity use.

Do data centers use as much power as a US state?

Harvard says the upper estimate approaches what states such as Indiana, Michigan and Tennessee use in a year. The lower estimate is close to Wisconsin's roughly 70 TWh, TIME reports.

How much carbon dioxide do data centers produce?

The study links the 403 facilities to about 25.7 million metric tons of CO2 in the year studied. That is two to three times an earlier estimate for 2018.

Is all of this electricity used for AI?

We do not know. The researchers could not separate AI work from general-purpose computing at each site, so the totals cover both.

Words to know

Hyperscale data center
One of the largest data centers, built to run cloud services and to train and run AI.
Terawatt-hour
A unit of energy: one trillion watt-hours, used to count a country's or state's yearly electricity.
Behind-the-meter
Power made on site by a data center's own plant, not bought from the public grid.
Sources (4): every key fact checked against at least two

Related stories