403 US hyperscale data centers used 68–99 TWh, study estimates


The researchers attribute 21–31 million metric tons of carbon dioxide to electricity used by 403 hyperscale facilities between May 2024 and April 2025, using regional power mixes and assumptions about operating loads.
A Harvard-led study estimates that 403 US hyperscale data centers operating between May 2024 and April 2025 consumed approximately 68–99 TWh of electricity across different operating-load scenarios. Using location-based regional electricity-generation accounting, the study attributes approximately 21–31 million metric tons of CO2 to that consumption.
The central estimate puts the facilities' electricity use at about 82 terawatt-hours (TWh), or 1.8% of US consumption, and attributes 25.7 million metric tons of carbon dioxide (CO2) to that use. A terawatt-hour is a measure of electricity consumed over time. The study appeared in Nature Sustainability on 16 November 2026.
Earlier estimates put electricity use by US hyperscale facilities in 2018 at 22.85 TWh and associated emissions at 10.5 million metric tons of CO2. The comparison is between estimates, not a consistent series of meter readings.
The study defines hyperscale facilities as sites serving large cloud or content providers, either directly or through operators that host their equipment, with electrical capacity in the tens of megawatts. The team started with 675 candidates from data provider Baxtel and screened them for capacity and suitable buildings, using satellite images and OpenStreetMap building footprints.
Instead of meter readings, researchers multiplied each site’s electrical capacity by the hours in a year and an assumed operating load. The range reflects different assumptions about how heavily facilities operated. Available information did not cover every US hyperscale site, and the data could not separate electricity used for superintelligence (SI) from general computing at individual campuses.
The authors disclosed that one researcher works for Baxtel, which supplied commercial facility data, and another works for Environmental Systems Research Institute. The other authors declared no competing interests.

To estimate emissions, the team applied regional average electricity mixes using plant-generation and emissions data from the US Environmental Protection Agency’s eGRID2023 dataset. The method does not trace the plants supplying each facility or measure additional emissions caused by its demand. It also leaves out on-site generation and clean-power contracts and credits.
Roughly 54% of the electricity generation attributed to the sites came from fossil fuels, mainly natural gas and coal; 25% came from renewables and 21% from nuclear power. The study estimated that the sample’s electricity was less carbon-intensive than the US average: 314 grams of CO2 per kilowatt-hour, compared with 348 grams.
Lead author Gianluca Guidi, a Harvard biostatistics postdoctoral fellow, said facilities using the same amount of electricity can have very different carbon footprints depending on their location and power supply. Virginia led the study’s central-scenario state estimates for both electricity consumption and attributed emissions, at 21.4 TWh and 7 million metric tons of CO2; Oregon ranked second for electricity use at 11.1 TWh and Iowa second for emissions at 3.2 million metric tons.
Guidi said new facilities, expanded campuses and growing SI demand mean the footprint could already be substantially larger in 2026. He and his colleagues are building a website to track individual data centers' environmental impacts, The Verge said. Francesca Dominici, the study’s corresponding author, said quantifying health effects is an important next step.
Guidi contributed to the study’s conceptualization, methodology, data collection and investigation. Dominici contributed to its conceptualization and methodology and shared responsibility for supervision and project administration.
Separately, the US Energy Information Administration is using completed pilot surveys to design a survey of data-center energy use, an agency spokesperson told Wired.

The Gemini app’s task feature became available to free users in the US, Japan, Korea, Canada, Australia and India on 17 November, after previously requiring Google AI Pro.
2 min

Microsoft says the programming model nearly matches rivals at a fraction of the cost. The company also released image-generation and transcription models and plans limited access to a cybersecurity model.
6 min

Google says future chats will be kept for around two months rather than the previous 72 hours when history is off. Separate controls will let users save chats without allowing model training.
2 min