AI Momentum 104.8 +2.6%GitHub Pulse 5171 -5.9%Capital Flow $206.5B +5650.2%Hiring Heatmap 0.0 -100.0%
Advertisement
Infrastructure

Harvard estimates the biggest US data centers use 1.8% of the nation's electricity, more than half from fossil fuels

The companies spending hundreds of billions of dollars on AI data centers face a basic question from regulators, utilities and local officials: how much power do these buildings actually use? No government agency meters them, so nobody has an official answer.

A Harvard-led study published November 16 in the journal Nature Sustainability offers the closest thing yet. Latitude Media called it the first peer-reviewed, facility-by-facility estimate since the AI boom began. It covers 403 of the largest US data centers, known as hyperscale sites, over the year from May 2024 to April 2025.

The researchers estimate those sites used 68 to 99 terawatt-hours of electricity, depending on how hard they were running. A terawatt-hour is a billion kilowatt-hours. The central estimate, about 82 terawatt-hours, equals roughly 1.8% of US electricity use. The high end approaches the yearly use of a state like Michigan or Indiana.

About 54% of that power came from fossil fuels: 38% from natural gas and 15% from coal. Nuclear supplied about 21% and renewables about 25%. The study puts the resulting carbon dioxide at about 25.7 million metric tons, within a range of 21 to 31 million. The authors say that is two to three times an earlier estimate for 2018.

What the estimate does and does not show

Investors should read the figures as an informed estimate, not a measurement. The team found the sites using satellite images, building maps and permit records. It then multiplied each site's electrical capacity by the hours in a year and an assumed usage rate, under four scenarios ranging from light use to heavy AI work.

The study cannot separate power used for AI from power used for ordinary computing. And it assigns each site the average power mix of its regional grid, rather than tracing which plants actually serve it.

That last choice matters for how the numbers compare with corporate climate reports. Tech companies usually credit themselves for contracts to buy clean power. According to Tom's Hardware, the Harvard study gives no such credit. Its figure reflects the grid the buildings draw from, not the power their owners have paid to add elsewhere.

The snapshot is also dated. Lead author Gianluca Guidi said the footprint could already be substantially larger in 2026, since construction has continued. And hyperscale sites are only one slice of the market. ITPro notes the US has about 5,000 data centers in total. A December 2024 Lawrence Berkeley National Laboratory report put all US data centers at about 176 terawatt-hours in 2023, or 4.4% of national use. It projected 6.7% to 12% by 2028.

Location drives the carbon bill

The power use is highly concentrated. Virginia alone hosts 142 of the sites and about a quarter of the estimated demand, roughly 21 terawatt-hours. Oregon has 56 sites and Ohio 38. Together with Iowa, those four states account for more than half the total.

On average, the sites ran on slightly cleaner power than the US grid as a whole. Tom's Hardware's tally of the data puts the sites at 314 grams of carbon dioxide per kilowatt-hour, against a national average of 348, and credits nuclear power in the grid region that includes Virginia and hydropower in Oregon. But about a quarter of the electricity came from grid regions dirtier than average.

The gap between regions is wide. By that same tally, Oregon sites ran at about 119 grams per kilowatt-hour and Wyoming's two sites at about 848, roughly seven times as much. For a company picking its next campus, siting is now also a choice about how large a carbon figure it will have to defend.

The study may also miss a growing source of emissions. Some operators generate their own power with gas on site. Tom's Hardware notes that such on-site power may not show up in the grid data the study relied on.

The bigger risk is disclosure

The authors argue that the lack of public data is itself the problem. Francesca Dominici, the study's senior author and director of the Harvard Data Science Initiative, told Fast Company the gaps point to a lot of secrecy about these facilities. She asked why there cannot simply be a public database showing where the facilities are.

The federal government has started to collect some of this data. Its Energy Information Administration has run pilot surveys asking 196 companies about their data centers' energy sources, power use, servers and cooling. It is using them to design a broader survey of the sector, Tom's Hardware reported.

The researchers have also released a public tool, Never Just a Building, that scores hypothetical 100-megawatt sites from 0, the best, to 100, the worst. It traces which power plants would supply a site and maps where their pollution would drift. A sample site in Loudoun County, Virginia, scores 76. Switching it to round-the-clock clean power brings that to 54, Latitude Media reported. The researchers say the tool takes no side on whether a project should go ahead.

For developers already facing local opposition, that gives residents and officials a ready-made scorecard. No site owner was quoted in the coverage of the study. The team plans follow-up work on health effects, water use and noise, which would add more numbers to the debates over local approvals.