What the study says
Large US data centers consume electricity at a scale comparable to a medium-sized country and are still powered predominantly by fossil fuels. That is the conclusion of an article published in Nature Sustainability and covered by The Verge, according to which hyperscale facilities—enormous plants spanning millions of square feet, built largely to support generative AI—have an energy and climate impact that is anything but negligible.
Led by Gianluca Guidi, a postdoctoral researcher at the Harvard T.H. Chan School of Public Health, the authors identified 403 hyperscale data centers in the United States using satellite imagery. For each, they estimated consumption based on electrical capacity and, using data from the Environmental Protection Agency, calculated the CO2 emissions that would have resulted from connection to the local grid between May 2024 and April 2025.
The result: about 82 terawatt-hours consumed in that year, equal to 1.8% of total US electricity. Associated emissions are estimated at 25.7 million tonnes of CO2. According to the research, about 54% of the electricity used came from gas and coal.
Comparison with countries and the weight of the local grid
To give a sense of scale, John Sterman of MIT—not involved in the study—told The Verge that the amount of electricity exceeds that consumed individually by about 148 nations, including Austria, Israel and Venezuela, excluding large countries such as the United States and China. The emissions, again according to Sterman, would exceed the total emissions of about 113 countries and would be comparable to Denmark's annual footprint. Sterman specified that he used Claude's assistance for the international comparisons.
The most significant finding concerns geographic variability. In the study, producing electricity in Wyoming, which relies heavily on coal, entails emissions per kilowatt-hour about seven times higher than in Oregon, which is rich in hydropower. Guidi stresses that two data centers with the same consumption can have very different carbon footprints depending on the grid that powers them.
Consequences and open questions
Fengqi You, a professor of energy systems engineering at Cornell and not involved in the research, speaks of a long-term 'carbon lock-in' risk: plants and infrastructure built for the growing demand of data centers can remain in operation for decades, shaping future emissions. Sterman calls for policymakers to require operators to add new clean energy.
Several limitations remain. The study does not consider local air pollution linked to new gas infrastructure or the turbines installed by some AI companies to power themselves off-grid. Hyperscale data centers do not coincide with AI-related computing alone and represent only a portion of the thousands of facilities present in the United States. The authors themselves highlight the difficulty of estimating consumption because operators do not disclose it; for this reason they are building a monitoring website and cite a bipartisan bill in Congress that could impose transparency requirements.
