The 403 large hyperscale data centers operating in the United States between May 2024 and April 2025 consumed between 68 and 99 terawatt-hours (TWh) of electricity, according to a study led by researchers at Harvard University. The high end of that range exceeds the annual electricity consumption of an entire country like Chile, around 83.2 TWh, and in the central scenario amounts to roughly 1.8% of total US electricity demand. The authors themselves warn that the current figure is probably higher.
US hyperscale data center power consumption in 30 seconds
- A Harvard-led team analyzed 403 US hyperscale data centers between May 2024 and April 2025.
- It estimates consumption of 68 to 99 TWh, with 82 TWh in the central scenario, around 1.8% of the country’s demand.
- Virginia, Oregon, Ohio, and Iowa account for more than half of the detected consumption.
- The lead author considers the figures conservative, since AI’s expansion hasn’t slowed since April 2025.
The study, whose lead author is Gianluca Guidi, a postdoctoral researcher at the biostatistics department of Harvard’s T. H. Chan School of Public Health, was published in the journal Nature Sustainability after an earlier version was shared in June on the arXiv preprint repository. Researchers from the University of Pisa, the IMT School for Advanced Studies Lucca, mapping company Esri, data center consultancy Baxtel, and the University of California, Los Angeles (UCLA) also contributed.

How to measure what operators don’t publish
The main obstacle to calculating how much energy data centers consume is that operators almost never disclose it facility by facility, and Department of Energy data is only published aggregated at the national level. Guidi sums it up by saying there’s barely any public data. To work around that, the team started from an initial list of 675 data centers drawn from private vendors and open sources, and refined it with manual verification and satellite imagery down to 403 facilities for which floor area or power capacity was known.
To count as hyperscale, each center had to belong to a major cloud or content provider, or to a colocation operator serving those clients; have an electrical capacity of tens of megawatts; and occupy a building or campus of the size typical of this category. According to Wccftech, the facilities in the sample average around 40 MW of installed capacity. Annual consumption was calculated by multiplying that capacity by the year’s 8,760 hours and by a load factor. Since no one knows precisely what percentage each center runs at, the authors offer several scenarios, which is where the 68-to-99 TWh range comes from. In the central scenario, with a factor of 0.58, the result is 82 TWh.
Four states account for more than half the consumption
The distribution is highly uneven. Virginia, Oregon, Ohio, and Iowa together make up more than half of the detected consumption. According to the preliminary version of the study, Virginia — home to the large cluster of data centers in Loudoun County — leads with about 21 TWh, followed by Oregon with 11 TWh, Ohio with 9 TWh, and Iowa with 7 TWh. To give a sense of scale, Wccftech compares the total to the consumption of states such as Indiana, Michigan, or Tennessee.
These figures need to be read with their exact scope in mind. The study doesn’t measure every data center in the country, only hyperscale ones; an earlier study from the same group, covering more than 2,100 facilities of all sizes between September 2023 and August 2024, put the sector as a whole above 4% of US electricity consumption. Nor is this a direct measurement — it’s an estimate based on installed capacity and usage assumptions.
Why the real figure is probably higher
Guidi insists the results are conservative. The observation window ends in April 2025, and AI infrastructure construction hasn’t stopped growing since then, so both the consumption and carbon footprint the study captures could already be understated. Ana Valdivia, a professor of AI, government, and policy at the University of Oxford who wasn’t involved in the research, adds another reason: the analysis focuses on grid electricity and doesn’t capture off-grid, self-generated power, such as what xAI’s Tennessee facility uses.
For cloud operators and for the authorities planning the grid, the study confirms a known difficulty — one that has already led Texas to halt new data center connections over grid capacity — deciding where and how the power grid should grow without reliable data on what each facility consumes. Some operators are trying to ease that strain directly: Google and NVIDIA recently joined an alliance to make AI data centers’ power demand more flexible. The authors’ own method, based on open data and satellite imagery, is one way of filling the data gap while operators don’t publish each facility’s consumption.
Frequently asked questions
How much electricity do US hyperscale data centers consume?
According to the Harvard-led study, between 68 and 99 TWh between May 2024 and April 2025, with 82 TWh in the central scenario.
Why is it compared to Chile?
Because the high end of the estimate, 99 TWh, exceeds Chile’s annual electricity consumption of about 83.2 TWh.
Which states consume the most?
Virginia, Oregon, Ohio, and Iowa, which together account for more than half of the detected consumption.
Does the study cover every data center?
No. It only analyzes 403 hyperscale data centers; it doesn’t include smaller facilities or off-grid self-generated power.
Sources:
- Nature Sustainability: Guidi et al. study
- arXiv: Assessing the Carbon Emissions and Energy Consumption of U.S. Hyperscale Data Centers
- Wccftech
- arXiv: Environmental Burden of United States Data Centers in the AI Era
This article is also available in Spanish: “Los centros de datos hiperescala de EE. UU. consumen casi lo mismo que Chile,” on revistacloud.com.

