Technology

Study Estimates 403 US Hyperscale Data Centers Used About 1.8% of National Electricity

Martin HollowayPublished 12m ago3 min readBased on 1 source
Reading level
Study Estimates 403 US Hyperscale Data Centers Used About 1.8% of National Electricity
Photo by Brett Sayles on Pexels

A study in Nature Sustainability estimates that 403 US hyperscale data centers likely consumed about 82 terawatt-hours (TWh) of electricity from May 2024 through April 2025. That is roughly 1.8 percent of total US electricity consumption over the same period. The authors also estimate that about 54 percent of that load likely came from gas and coal generation, and that the facilities' demand produced 25.7 million metric tons of CO2.

The lead author is Gianluca Guidi, a postdoctoral fellow in the Department of Biostatistics at the Harvard T.H. Chan School of Public Health. The Verge reported the findings on 16 November 2026.

The researchers used satellite imagery to locate the 403 facilities. They did not work from metered consumption. Instead, they estimated each site's power draw from its electrical capacity, meaning the maximum load its wiring and equipment are rated to carry, often called nameplate capacity. For emissions, they applied EPA data to approximate carbon pollution, assuming each facility is connected to its local power grid. The consumption figure is therefore a modeled estimate, which is why the study's wording relies on "likely."

The 25.7 million metric tons figure is comparable, according to a The Verge report, to Denmark's annual carbon footprint. As a comparison, it is a useful scale marker, though it carries the same modeling assumptions as its inputs.

The emissions estimate depends on the grid assumption. Because each site is attributed to its local grid, the result reflects the regional mix of power plants rather than the electricity any individual operator bought or contracted. Engineers who track carbon intensity per workload will recognize the distinction between location-based and market-based accounting, although the study does not frame its method in those terms.

In my view, the 1.8 percent national share is the number most likely to travel, but it says little about grid stress in the regions where these facilities cluster. A national average can look modest while a single balancing authority, the operator responsible for matching supply and demand in its territory, absorbs a large fraction of new load. The more useful follow-up for utilities and planners is local: how demand from these sites sits against regional capacity and generation mix over time.

The broader context here is that a satellite-based inventory lets analysts count physical facilities without relying on operator disclosures. That kind of baseline makes it possible to track consumption and emissions year over year, and to test whether the gas and coal share shifts as the grid changes. A single year of modeled estimates is a starting point rather than a trend line, but it gives later measurements something to be checked against.

What the study enables is a clearer accounting of where data center load lands on the grid and what it costs in emissions. Better location data, paired with metered figures where operators can provide them, would tighten the estimates considerably. For an industry that builds capacity in megawatts, that visibility is a practical asset.