A 49.5 megawatt three-level data center under construction on October 29, 2026 in Vernon, California. Photo by Mario Tama/Getty ImagesToday, a team led by Harvard researchers published a highly anticipated study in Nature Sustainability. For the first time since generative AI took off, we have a peer-reviewed estimate of how much climate pollution comes from keeping America’s biggest data centers running.Headlines about it are already popping up all over. The study estimates that hyperscale data centers across America are emitting up to 31 million metric tons of carbon dioxide per year, roughly as much CO₂ as Denmark emitted in 2024. It also estimates the facilities are taking up roughly 2 percent of U.S. electricity consumption.These are big, important numbers! And yet, they’re also massive underestimations.HEATED is funded by readers—no advertisers, corporations, or billionaires. For more independent journalism exposing the powerful interests shaping the climate crisis, become a subscriber.There are a few reasons why today’s study gets nowhere near the actual amount of climate pollution coming from these ever-buzzing internet slop factories. One is the way the researchers counted the facilities’ emissions. To do this, they estimated how much electricity each data center used, then looked at how the grid in that part of the country makes its electricity—how much comes from gas, coal, renewables, and so on—to estimate how much CO₂ would be released to generate that much power.But not all data center electricity comes from the grid. Indeed, an ever-increasing number of these massive data centers are also building their own, on-site fossil fuel plants. For example, during the study period (May 2024 through April 2025), Elon Musk’s xAI got some of the electricity for its Colossus data center in Memphis from the grid—but it also got some of it by burning gas in turbines right outside the building. Lead author Gianluca Guidi confirmed that Colossus wasn’t among the 403 facilities studied. For the facilities they did study, the researchers couldn’t tell how much power came from the grid and how much was generated on-site. So they applied the regional grid mix to all estimated electricity use, which can undercount emissions at facilities that burn their own gas.Another reason the climate impact of hyperscale data centers is likely understated is that electricity consumption is not the only way these facilities heat the planet. For example, the equipment that keeps the servers from overheating is leaking heat-trapping refrigerants into the atmosphere. And when a data center runs on gas—as so many of them do—methane can escape from wells and pipelines before the gas even reaches the power plant.In addition, there are potentially hundreds of hyperscale data centers that were left out of today’s study on the climate impacts of these facilities. Yes—hundreds. Subscribe nowThat’s because we don’t actually know how many hyperscale data centers are currently up and running in the United States. (Hyperscale means ‘the largest and most powerful class of data centers,’ according to the study. These are the kinds of facilities that can train and run large language models. See footnote to see how the study defines them).1The researchers only landed on their number—403—after painstakingly going through a list of 675 possible hyperscale data center locations given to them by a company called Baxtel, which tracks data centers. And by that I mean they had to pull up satellite images for every single facility to answer questions like Is this the data center, or just another building on the same campus? Has it even been built yet, or is this satellite photo just old? And even after they found the right building, they still had to figure out whether it was running before they could figure out how much electricity it might be using.All told, the researchers had to leave 272 possible hyperscale data centers out of the study because they couldn’t find out enough about them.Finally, the study period ends in April 2025—meaning the climate impact of data centers has likely grown substantially. Gianluca Guidi, the study’s lead author, acknowledged this in Monday’s release. ‘The data center boom hasn’t slowed down since our study period ended in April 2025,’ he said. ‘New facilities have been built, existing campuses have expanded, and demand for AI computing continues to grow. That means the electricity use and carbon footprint we’re reporting could already be substantially larger now in 2026.’And yet, even with all of these limitations (most acknowledged by the study itself, which is how I know about them), the researchers were able to pull out some important findings. They include:The generative AI boom is mostly powered by fossil fuels. 53.9 percent of the electricity consumption of hyperscale data centers is fossil-fuel generated, with 38.2 percent coming from methane gas and 15.2 percent coming from coal. Hyperscale data centers use as much electricity as some U.S. states. They used 82 terawatt-hours of electricity in a year (68–99 terawatt-hours of electricity across scenarios). That’s as much electricity as some individual U.S. states over the course of a year, with the upper estimate approaching the annual electricity consumption of Indiana, Michigan, and Tennessee. It represents about 1.8% of all U.S. electricity consumption.Hyperscale data centers are consuming more than 3 times as much electricity as they were in 2018. An earlier study done before the generative AI boom put U.S. data centers’ electricity use at about 23 terawatt-hours in 2018. The new study’s central estimate is about 3.6 times as high.Virginia alone accounted for about a quarter of the electricity use. The researchers counted 142 facilities in Virginia, more than twice as many as in any other state. Powering those facilities was associated with about 7 million metric tons of CO₂. Together, Virginia, Oregon, Ohio, and Iowa accounted for more than half the electricity used by the data centers in the study.Now, to be totally clear, I’m not saying any of this to bash the study! Quite the contrary: I think it represents an important, Herculean effort, especially given how little information Congress and state governments require tech companies to give us about the unprecedented expansion of data centers across the country. It’s clear the researchers did everything they could possibly do with the paltry information they had.What I’m saying is: It’s supremely messed up that scientists have to work this hard just to get a partial picture of how much an immensely powerful, fast-growing industryis heating the planet we all have to live on—especially when voters just made it clear they want the industry reined in. The public deserves to know exactly how many of these facilities exist, and how much they are polluting, so they can decide whether it’s worth trading the planet’s health for toothbrushes that tell us if we’re doing it right. Come January, Democrats can show voters they were listening by making tech companies publish the data scientists actually need to figure out how many hyperscale data centers are running and how much they pollute. My bet is that once scientists get the full numbers, we’ll need a bigger country than Denmark for the comparison.Subscribe nowBut what are Democrats actually doing on AI right now?Great question! The Senate this week is expected to push through the Bipartisan American Affordability and Jobs Act, or BAAJA—a massive, extremely controversial permitting reform bill that would weaken some of the nation’s oldest environmental laws with the intention of speeding up all forms of energy development, both renewable and fossil fuel. The AI Now Institute, an independent research institute providing expert analysis on AI in the public interest, is calling BAAJA ‘a backdoor bailout of the AI industry’—essentially because it claims the bill would make Big Tech’s bet on AI cheaper and less risky to pursue.One reason is that BAAJA would speed up the buildout of the gas infrastructure that powers AI. The bill would help those projects move faster by ‘shortening approval timelines, eliminating and weakening environmental review processes,’ the institute wrote. One example it flagged is a proposed gas pipeline that would serve Oracle’s Project Jupiter data center campus in New Mexico.Another reason is that BAAJA would make it harder for communities to push back against data centers and the infrastructure that powers them. The bill would shield the industry from ‘the growing financial risk of public opposition,’ the institute wrote. One way it would do that is by making it harder for states and eligible Tribes to deny or put conditions on water-quality approvals for projects such as gas pipelines.A third reason is that BAAJA would make it easier to build data centers on public land. The bill would make federal approvals for those projects ‘much more difficult to challenge or reconsider,’ the institute wrote. One example it flagged is Nevada’s Townsite Data Center, where an Interior Department appeals board temporarily halted construction after finding that federal officials likely erred by relying on an environmental review done for a solar farm.More on BAAJA later this week.(For a general summary, I recommend Aaron Regunberg’s piece for The New Republic, ‘The Democrats’ First Move After Their Big Win? A Giveaway to Big Oil.’)Subscribe nowCorrection: An earlier version of this article misstated why emissions from xAI’s Colossus gas turbines were missing from the study’s estimates. Colossus was outside the study’s sample, so the researchers did not estimate its electricity use or emissions.Previous HEATED coverage of AI and climate1How does the study define a hyperscale data center? ‘We operationally define a hyperscale data center as a facility that satisfies three criteria: (i) it is operated by a hyperscale cloud/content provider or a colocation operator serving hyperscale tenants at scale; (ii) it has reported or inferred facility electrical capacity in the tens of megawatts; and (iii) its building or campus footprint is consistent with large-scale modular hyperscale architecture as verified through imagery and footprint data. We do not impose a single square-footage threshold because newer AI-oriented facilities can be substantially more power-dense than earlier cloud facilities.’

Read original article