
US hyperscale data center carbon emissions hit 25.7 million tons a year, Harvard study finds
By Yash Malviya·
A Nature Sustainability study of 403 US hyperscale data centers puts their 2024 to 2025 electricity use at 68 to 99 terawatt-hours and their emissions near Denmark's. The authors call it a floor, and it cannot separate AI from other computing.
The data centers that train and serve AI models now have a peer-reviewed carbon bill. A team led by Harvard researchers estimates that 403 large US hyperscale data centers used between 68 and 99 terawatt-hours of electricity from May 2024 to April 2025 and emitted about 25.7 million metric tons of carbon dioxide, a total comparable to Denmark's annual emissions. The study appeared in Nature Sustainability on November 16, 2026. It is the strongest public number yet on the physical cost of the compute race, and it comes with limits worth reading before anyone repeats the headline.
The numbers
According to the Harvard T.H. Chan School of Public Health release, the 403 facilities used about 1.8% of US electricity. The upper estimate of 99 TWh approaches the annual use of Indiana, Michigan or Tennessee. The Next Web reports a central scenario of about 82 TWh and roughly 25.7 million metric tons of CO2, with a scenario range of 21 to 31 million tons. The authors say that is two to three times an earlier 2018 estimate. Semafor, summarizing the same paper, described the electricity share as "almost 2%."
The energy mix explains the emissions. About 54% of the attributed electricity came from fossil fuels: 38% natural gas and 15% coal. About a quarter of the power came from grid regions dirtier than the national average. Virginia, Oregon, Ohio and Iowa together accounted for more than half of the electricity studied, and Semafor says sites in Iowa, Ohio and Virginia account for about half the emissions. Oregon ranked second for power use but sixth for emissions, because of hydropower.
How they counted
No public list of data centers exists, and operators and utilities are not required to disclose power use. So this is an observational, modeled study. The researchers identified the 403 facilities from satellite imagery and industry data, estimated each one's annual electricity use under a range of operating scenarios, and used EPA data on regional power-plant mixes to turn that electricity into emissions.
That method shapes what the result can say. Emissions are estimated, not metered. Clean-energy contracts were not counted, which could lower the figure for sites whose operators buy renewables. On-site, self-generated power was excluded, which could raise it. The authors call the result a floor, and Semafor reports it as "likely an underestimate" for that reason. The release also says there was not enough data to cover all hyperscale data centers.
“Decisions about where data centers go and how they're powered are critical environmental and public health decisions”
The AI question the study cannot answer
Readers will want to know how much of this is AI. The honest answer is that the paper cannot say. The release states that the data could not distinguish power used for general-purpose computing from power used for AI-specialized operations, and The Next Web notes that not every facility in the study runs AI work. Cloud storage, streaming and ordinary web services share these buildings.
That does not make the study irrelevant to the AI race. The sites are growing, and lead author Gianluca Guidi said current figures could be "substantially larger now in 2026." He also said: "What we're seeing is already striking, but it's only a snapshot of a rapidly expanding industry." Capacity announced since the study window closed, including multi-gigawatt AI campuses, will not appear in these numbers. Our data center power coverage from the Energy Information Administration's November outlook shows the national demand side of the same picture.
What the authors and critics say
Senior author Francesca Dominici said the emissions are "the equivalent to the CO2 emission of an entire state like Michigan or a small European country," and: "Decisions about where data centers go and how they're powered are critical environmental and public health decisions." Health effects were not measured in this paper, though the authors plan to study them.
An industry voice, Caleb Max of the National Artificial Intelligence Association, told The Next Web: "I respect the concerns about energy and water consumption from data centers, but the fears may be overblown." Outside experts quoted in the same report called the modeling a reasonable approach given limited company disclosure. That is the real policy finding. The study exists because the data does not.
Where this lands in the policy fight
The Next Web reported that on the same day COP31 negotiators in Antalya launched a voluntary pledge asking data centers to use clean power and disclose energy, water and emissions data. We have not independently confirmed the pledge's text. If it holds, disclosure is the lever that would let the next study replace estimates with measurements.
Reading the numbers against each other
Three points help. The 25.7 million tons is a model output, while the 1.8% electricity share is a ratio the authors can check against national totals, so the second is sturdier. Semafor's "almost 2%" and the release's 1.8% describe the same finding, not two studies. And the comparison with 2018, at two to three times, says more about growth than about level: the sites in this study are a mix of cloud, enterprise and AI facilities, so a rise in total power does not map neatly onto model training.
Our take
The headline number is credible as an order of magnitude and soft at the decimal. Two things should change how readers use it. First, the geography matters more than the total: four states carry over half of the electricity, and the fossil share in those grids, not the existence of data centers, drives the carbon. Second, the missing AI breakdown is a disclosure failure by the operators, not a flaw in the research. We would watch for state utility commission filings and for any lab or cloud provider that publishes metered, site-level figures. Until then, every number in this debate, including this one, is a model.
Frequently asked questions
How much carbon dioxide do US data centers emit?
A Harvard-led study in Nature Sustainability estimates that 403 US hyperscale data centers emitted about 25.7 million metric tons of CO2 from May 2024 to April 2025, with a scenario range of 21 to 31 million. That is roughly Denmark's annual emissions.
How much electricity do US hyperscale data centers use?
The same study estimates 68 to 99 terawatt-hours across the 403 facilities, about 1.8% of US electricity. The Next Web reports a central scenario of about 82 TWh. The upper figure approaches the annual use of Indiana, Michigan or Tennessee.
Is the 25.7 million ton figure for AI only?
No. The authors say they could not separate power used for AI-specialized work from general-purpose computing, and not every facility in the study runs AI. The sites also host cloud storage, streaming and ordinary web services.
Which states have the most data center emissions?
Virginia, Oregon, Ohio and Iowa accounted for more than half of the electricity studied. Semafor reports Iowa, Ohio and Virginia account for about half of the estimated emissions. Oregon ranks second for power use but sixth for emissions because of hydropower.
Is the estimate too high or too low?
The authors call it a floor. It excludes on-site self-generated power and facilities the researchers could not verify. It also does not count clean-energy contracts, which could lower some sites' figures. Lead author Gianluca Guidi said the current numbers could be substantially larger in 2026.
How did the researchers estimate emissions without meter data?
They identified the 403 facilities from satellite imagery and industry data, estimated each site's electricity use under several operating scenarios, and applied EPA data on regional power-plant mixes to attribute emissions. Operators are not required to disclose their power use.
Sources
What each one is, and whose it is.
- 1
Hyperscale data centers' energy consumption, carbon emissions newly estimated by researchers, Harvard T.H. Chan School of Public Health via EurekAlert (November 16, 2026)
PaperIndependent of the vendorPeer reviewed - 2
The largest US data centres emit as much CO2 as Denmark, Harvard finds, The Next Web (November 16, 2026)
Press reportIndependent of the vendor - 3
US data center emissions rival entire countries, Semafor (November 17, 2026)
Press reportIndependent of the vendor