The 403 hyperscale data centers examined in a new study consumed between 68 and 99 TWh of electricity in one year: the upper end is close to the amount used by states such as Michigan, Indiana, or Tennessee. The study, published in Nature Sustainability, also estimates that in the central scenario their demand equaled 1.8% of total U.S. electricity consumption.
The comparison with Michigan refers to the upper end of a range, not to a single measurement. The central scenario puts consumption at 82 TWh, and the study attributes about 25.7 million metric tons of carbon dioxide to that electricity. Harvard T.H. Chan School of Public Health, which publicized the findings, says the emissions are two to three times higher than an earlier estimate for 2018.
An estimate for each data center
The team gathered information on 403 hyperscale facilities that operated between May 2024 and April 2025. These are large data centers that support services such as cloud computing, language model training, and the deployment of artificial intelligence tools. To identify them, the researchers combined satellite imagery with information about the facilities and estimated their consumption under different operating scenarios.
They then matched each data center to its region’s electricity mix to calculate the associated emissions. The study used data on power plant generation and emissions, as well as an attribution method based on regional balancing authorities. It did not directly measure every outlet or distinguish what share of the load was attributable to AI and what share to other services. Nor were all the hyperscale data centers included dedicated to AI, WIRED notes.
This method makes it possible to compare facilities and regions in greater detail than an aggregate figure for the entire sector, but the results depend on the load scenarios and how grid generation is allocated. The article calculates a consumption-weighted carbon intensity of 314 g of CO₂ per kWh, compared with a U.S. average of 348 g. This national difference does not eliminate regional disparities: nearly a quarter of the electricity attributed to the centers came from regions with an intensity above the average.
| Indicator | Estimate |
|---|---|
| Centers included | 403 |
| Operating period studied | May 2024–April 2025 |
| Consumption, range across scenarios | 68–99 TWh |
| Consumption, central scenario | 82 TWh |
| Attributed emissions, central scenario | 25.7 million metric tons of CO₂ |
| Share of electricity attributed to fossil fuels | Approximately 54% |
Location changes the footprint
The generation mix varies widely depending on the region where each facility connects to the grid. Across the group analyzed, 54% came from fossil fuels, mainly natural gas (38%) and coal (15%). Four states—Virginia, Oregon, Ohio, and Iowa—accounted for more than half of the electricity consumed by the data centers studied. The researchers noted that many centers were in areas where electricity is generated largely by coal-fired power plants.
The study shows that associated emissions vary across regions and that nearly a quarter of the electricity attributed to the centers came from areas with a carbon intensity above the average.
The national average can also obscure local decisions. An attributed intensity below the U.S. average coexists with more polluting regions and a majority share of fossil fuels. For developers and operators, reporting only total consumption or a national emissions figure can therefore leave out differences that an analysis at the facility level can reveal.
What the figure does not include
The observation period ended in April 2025, and the Harvard team says that since then, facilities have been built, existing campuses have expanded, and demand for AI computing has continued to grow. Gianluca Guidi, the lead author, described the result as a snapshot: the current footprint could be substantially larger, although the study does not calculate how much it has increased since then.
There are also coverage limitations. The researchers say they could not gather enough data to include every hyperscale data center in the country, and they did not have information to distinguish general workloads from operations specializing in AI. Thus, the 68–99 TWh figure should not be read as a complete census of all U.S. data centers or as an exclusive measurement of AI consumption.
WIRED points to another factor left out: electricity generation installed alongside a data center and connected behind the meter, known as behind-the-meter. These projects can change the relationship between a facility and the regional grid used by the study’s method. The outlet quotes Jon Koomey, an energy researcher not involved in the study, who called this generation a “wild card” in estimating total consumption. An assessment of the impact of new campuses would therefore also need to consider this option, which the analysis does not incorporate.
“The data center boom hasn’t slowed down since our study period ended in April 2025. New facilities have been built, existing campuses have expanded, and demand for AI computing continues to grow. That means the electricity use and carbon footprint we’re reporting could already be substantially larger now in 2026.”
— Gianluca Guidi, lead author of the study and member of Harvard T.H. Chan School of Public Health
An information gap for the sector
The estimate comes amid a situation in which, according to WIRED, federal and state governments do not break out data center consumption from the rest of the electricity sector, and figures for individual facilities are usually treated as confidential commercial information. The study seeks to address part of this gap through modeling and generation data, rather than complete records published by each operator.
The scale of consumption makes the consequences of this lack of transparency more apparent for investment and planning decisions. If companies and communities cannot compare projected loads with comparable data on use and generation, it is harder to understand what demand is being added to an area and what emissions are attributed to it. The research provides regional and state-level figures, while detailed records for each facility remain restricted by data-use agreements.
There is an early official effort underway: the U.S. Energy Information Administration (EIA), which collects energy data, completed pilot surveys on data center consumption in ten states. Its spokesperson told WIRED that the agency is using those results to design a possible voluntary survey for the sector. It is not yet a mandatory reporting system, but it could provide measurements to complement the study’s estimates.
Harvard also cautions that the study did not measure health effects. Its team plans to investigate them in future research, including the possible impacts of pollutants emitted by power plants associated with this demand. For now, the findings quantify electricity and emissions: in the central scenario, 82 TWh and 25.7 million metric tons of CO₂ for 403 centers over the period analyzed.
Sources: Nature, Harvard T.H. Chan School of Public Health, WIRED