Our article2 min read
Harvard Study: 403 US Data Centers Used 68 to 99 TWh. What It Says About AI Power Use
A Nature Sustainability study estimates large US data centers used 68 to 99 terawatt-hours in a year and emitted about 25.7 million tonnes of CO2. Here are the findings and the limits.

In this article
Key takeaways
- Researchers at Harvard estimate 403 US hyperscale data centers used 68 to 99 terawatt-hours between May 2024 and April 2025.
- They put the emissions at about 25.7 million metric tons of CO2, two to three times an earlier 2018 estimate.
- The study could not separate AI from other computing, and it ends in April 2025, so current use is likely higher.
- The result is an estimate from satellite imagery and models, not metered data from each site.
What the study found
A team from Harvard's T.H. Chan School of Public Health and Data Science Initiative published its paper in Nature Sustainability on 16 November 2026. It looked at 403 hyperscale data centers in the United States, meaning very large facilities run by big cloud and technology companies, over the year from May 2024 to April 2025.
The estimated electricity use is 68 to 99 terawatt-hours, depending on how hard the facilities were running. That is about 1.8 per cent of all US electricity use, according to the university's release. One news summary rounds this to 2 per cent. The upper figure is close to the yearly use of states such as Indiana, Michigan or Tennessee.
Emissions and energy sources
The researchers link that electricity to about 25.7 million metric tons of CO2, which they say is two to three times an earlier estimate from 2018. About 54 per cent of the power came from fossil fuels, mainly natural gas at 38 per cent and coal at 15 per cent. Around a quarter of the facilities sat in grid regions with dirtier-than-average power.
Where the impact lands
Virginia, Oregon, Ohio and Iowa together accounted for more than half of the electricity use in the study. The authors stress that effects differ a lot by region. They also note that the power plants serving this demand release fine particles, nitrogen oxides and sulfur dioxide, which are linked to heart and lung disease, and say future work will look at health effects. This study did not measure them.
How they worked it out
The team used satellite images to find the facilities, estimated each one's yearly electricity use under a range of operating scenarios, and then used US Environmental Protection Agency data on each regional grid to assign emissions. That means the numbers are modelled estimates, which is why the result is a range rather than one figure.
What the study cannot say
- It could not separate power used for AI from power used for ordinary computing, although the authors tie the growth of these sites largely to AI.
- It did not cover every US data center. ITPro notes the country has around 5,000 data centers in all, and the study looked only at large hyperscale ones.
- It ends in April 2025. The lead author says the building boom has not slowed since, so present-day use and emissions are probably higher.
- ITPro's article contains some inconsistencies, such as 400 versus 403 facilities, so we relied on the university release for the figures.
What it means for you
You cannot tell from this study how much power one chatbot question uses, and it does not claim to. What it does show is that the combined demand from large data centers is now a visible share of national electricity, and that much of it is still met by fossil fuels. If this matters to you, you can check whether your cloud or AI provider publishes energy and renewable-power figures, and favour lighter models for simple tasks. Local power prices and planning decisions are where the effect is most likely to show up.
Sources
Written by an AI from the sources above and checked by a person before publishing. We have not tested products unless we say so. See our editorial policy or tell us about an error through contact.



