US hyperscale data centers emit as much CO2 as Denmark, Harvard study estimates

The 403 largest data centers in the United States consumed between 68 and 99 terawatt-hours of electricity per year — with a central estimate of about 82 TWh — and emitted approximately 25.7 million metric tons of CO2, roughly the annual emissions of a country like Denmark. That is the estimate of a study led by the Harvard T.H. Chan School of Public Health and the Harvard Data Science Initiative, published on November 16, 2026, in the journal Nature Sustainability.

The work is one of the first facility-level estimates of energy use and emissions from so-called hyperscale data centers — the largest and most powerful facilities in the sector, which support cloud computing, large language model training and the global deployment of AI tools — since the start of the artificial intelligence boom.

What the study measured

The researchers used satellite imagery to build a list of 403 hyperscale data centers operating in the United States and estimated the annual electricity consumption of each facility between May 2024 and April 2025, under a range of operating scenarios. They then used data from the U.S. Environmental Protection Agency (EPA) to attribute carbon emissions to each facility based on the power mix of its regional grid.

During the study period, electricity demand from these data centers corresponded to about 1.8% of total U.S. electricity consumption. The associated emissions, 25.7 million metric tons of CO2 (within a range of 21 to 31 million metric tons, depending on the scenario), are two to three times higher than the previous estimate, from 2018 — the last peer-reviewed measurement before this one.

Where the power comes from

About 54% of the electricity attributed to the facilities came from fossil fuels — mainly natural gas (38%) and coal (15%). Nuclear plants supplied 20.9%, and renewable sources, including hydro, wind, solar, biomass and geothermal, accounted for 25.3%.

The researchers also found strong regional concentration: just four states — Virginia, Oregon, Ohio and Iowa — accounted for more than half of the electricity use studied. About a quarter of the electricity was in grid regions dirtier than the U.S. average.

The numbers are probably already larger

The study has important limitations, acknowledged by the authors themselves. Not every U.S. hyperscale data center could be studied — 272 facilities believed to be large data centers were left out of the analysis, according to Ars Technica — and the data did not allow the team to separate power used by AI workloads from power used by general-purpose computing. Facilities with on-site generators, such as xAI's gas turbines in Mississippi, were also excluded.

The estimates are not direct measurements of each data center either: they attribute regional power generation to each facility and do not count renewable power purchase agreements or certificates unless they are already reflected in local generation. The figures should therefore be read as estimates, not exact measurements.

Even so, the authors warn that the picture is already out of date: the data center boom has not slowed since the study period ended in April 2025. New facilities have been built, existing campuses have expanded and demand for AI computing keeps growing — so the electricity use and carbon footprint reported could already be substantially larger in 2026.

The study did not directly measure health effects, but the authors plan to investigate the topic in future work: beyond CO2, the fossil-fuel plants associated with this demand release pollutants such as fine particulate matter, nitrogen oxides and sulfur dioxide, linked to heart and lung disease.

Read also: thousands marched in Edinburgh against AI data centers in Scotland.

Image: server racks in a data center. Photo: Victor Grigas, Wikimedia Commons, CC BY-SA.

Fontes

Transparency note: this article was produced with the assistance of artificial intelligence, based on public publications on X and internet sources, and reviewed editorially. Consult the original sources to confirm the information.

Por Muse