A new analysis of 403 hyperscale data centres in the US suggests that together they consume roughly as much electricity each year as some of the country’s major states. The finding is a rare attempt to put a hard number on an industry that mostly keeps its energy use out of public view, and it arrives just as AI is pushing these facilities to grow at speed.
The numbers behind the headline
Francesca Dominici at Harvard University and her colleagues spent two years assembling information on 403 hyperscale data centres to estimate how much power they draw and how much carbon dioxide that generates. Hyperscale facilities are the very large sites that support AI tools, cloud computing and large language models.
Their estimate: about 82 terawatt-hours (TWh) of electricity between May 2024 and April 2025, roughly 1.8 per cent of total US electricity use. That is more than the annual consumption of states such as Missouri, Kentucky and New Jersey. The upper bound of the team’s range, 99 TWh per year, would put these facilities on par with states like Indiana, Michigan and Washington.
The climate side is just as striking. The researchers estimate that roughly 54 per cent of the generation attributed to these sites came from fossil fuels, producing annual carbon dioxide emissions of 25.7 million tonnes. That is comparable to the output of 67 natural gas power plants.
It helps to see why this number was hard to get at all. Many operators do not publicly report their power consumption or emissions, so outsiders have to reconstruct the picture from other data. Dominici’s description of the work, “an enormous amount of work”, reflects how much effort it takes to see something that is, in physical terms, enormous.
How much should we trust the estimate?
This is the part most coverage skips, and it matters. The study’s authors and outside researchers are candid about its limits.
Alex de Vries-Gao, a data scientist at the Free University of Amsterdam, notes that the results line up with international estimates of data centre power use from the International Energy Agency, so they are “not particularly shocking.” In his reading, the work mainly confirms the magnitude of the impacts. But he also cautions that the analysis relies on commercial data that is hard to verify independently, along with assumptions about the power use and operating capacity of each facility. “You can’t really validate the results,” he warns.
Yury Dvorkin at Johns Hopkins University adds a different critique: the analysis is relatively coarse. It does not account for factors such as how network constraints affect the emissions of grid power. “They ignore a lot of nuance,” he says.
Dominici acknowledges that many assumptions are baked in. Her point is less about defending the exact figure and more about the gap it exposes. She wants operators to be more transparent, because, in her words, the public narrative is not being informed by rigorous scientific data. An estimate built from indirect evidence is, in a sense, a symptom of the problem it describes.
So the sensible reading is this: treat 82 TWh as a credible order of magnitude rather than a precise measurement. The range of 82 to 99 TWh is the honest way to hold it.
Why it matters beyond the megawatts
Hyperscale data centres exist for a good reason. By harnessing economies of scale, they lower the cost of computing power for the companies that use it and run servers as efficiently as possible. The trade-off is that the bigger the facility, the heavier its footprint on the surrounding community.
Power is only one part of that footprint. These sites also need large volumes of water to cool their servers and take up a lot of land. The source points to a proposed data centre in Utah that would cover land twice the size of Manhattan and almost double the state’s annual peak electricity demand, a plan that has set off fierce backlash from residents.
De Vries-Gao describes what people on the ground experience: rising land, electricity and water prices, plus possible increases in air pollution and waste. That, he suggests, is why people protest, and why hyperscale sites may be the most painful type of facility simply because of their sheer size.
The politics are shifting accordingly. Public resistance in the US has risen sharply over the past year, and a recent poll suggests 76 per cent of Americans would oppose the construction of a new data centre in their community. Meanwhile, many new facilities are turning to private gas power generation instead of the grid, which increases their climate impact.
This is the larger lesson for anyone thinking about AI. It is easy to treat AI as weightless software, something that lives in “the cloud.” This research is a reminder that every model answer rests on concrete infrastructure: land, water, wires and fuel. Decisions about where that infrastructure goes, who pays for it and how openly it is measured are not technical footnotes. They are the terms on which communities will decide whether to accept the technology next door.
In short
Hyperscale data centres in the US used an estimated 82 TWh of electricity in a single year, about 1.8 per cent of national use and more than states like Missouri, Kentucky and New Jersey. The associated emissions, around 25.7 million tonnes of carbon dioxide, are significant, though the study rests on assumptions that outside experts say cannot be fully verified. What the work shows most clearly is that the industry’s energy footprint is large, poorly disclosed and increasingly contested by the communities that host it. Better transparency from operators is the missing ingredient for a debate that is currently running ahead of the data.
Based on reporting from New Scientist.
Read the original study.