by

AI datacentres are eating America’s electricity

Harvard Boffins have emerged from their smoke-filled labs with evidence that America’s hyperscale data centres are consuming enough electricity to power a small country. Unfortunately for the AI industry’s enthusiastic salesmen, the bill is getting bigger and somebody will eventually have to pay it.

According to the study published in Nature Sustainability, 403 large US data centres consumed an estimated 68 to 99 terawatt-hours (TWh) of electricity annually. The central estimate was 82 TWh, almost matching Chile’s annual electricity consumption of 83.2 TWh.

The figures represent approximately 1.8 per cent of total US electricity demand. For a relatively small collection of facilities, that is an impressive achievement in finding new ways to turn electricity into heat. The researchers calculated that the facilities were responsible for approximately 25.7 million tonnes of carbon dioxide emissions under their central scenario. Depending on how heavily the centres were operating, the figure could range from 21 million to 31 million tonnes.

According to the study, around 54 per cent of the electricity these facilities used came from fossil fuels. Nuclear power supplied approximately 21 per cent, while renewable energy contributed the remaining 25 per cent. The industry’s enthusiasm for describing its services as being in the cloud has apparently done little to persuade the coal and gas power stations to disappear.

Harvard’s researchers examined 403 hyperscale facilities operating between May 2024 and April 2025. Since datacentre operators are generally reluctant to reveal how much electricity their machines consume, the team had to reconstruct the picture using facility information, satellite imagery and estimates of installed computing capacity.

Study lead author Gianluca Guidi, a postdoctoral researcher at Harvard’s department of biostatistics, said that public information about individual data centres was extremely limited. America’s Department of Energy publishes national totals, which are of limited use when trying to determine which enormous computer sheds put the greatest strain on local electricity supplies.

The geographical distribution makes the situation particularly interesting. Virginia, Oregon, Ohio and Iowa accounted for more than half the estimated electricity consumption, leaving relatively small parts of the country carrying a substantial share of the industry’s appetite.

Virginia was the clear winner in this competition, with 142 facilities identified by the researchers. Its hyperscale data centres consumed an estimated 21TWh annually, about a quarter of the electricity used by the group.

The researchers warned that their figures could already be out of date. Their observations ended in April 2025, and the subsequent rush to build AI infrastructure means that more facilities have appeared while existing campuses have continued expanding.

Some data centre operators have begun installing their own power-generation equipment because electricity grids cannot supply sufficient capacity. Such arrangements complicate efforts to determine how much energy the industry uses, particularly when operators do not publish detailed information about their generators.

The researchers could not distinguish electricity consumed by AI workloads from that used for conventional cloud computing, and their estimates cover 403 identified facilities without claiming to represent every American hyperscale data centre.

Harvard professor Francesca Dominici said the researchers intend to examine the health consequences of the electricity demand. Power stations supplying data-centre-heavy regions emit pollutants linked to respiratory and cardiovascular disease, meaning communities near the generators may face consequences even when the computer facilities are hundreds of miles away.

 

 

Latest articles

Share

Featured articles

Hot topics

No results found.

Latest reviews