403 hyperscale US data centers used staggering 82 terawatts of electricity in a year, more than 80% of Michigan’s entire

AI & Compute

403 US Hyperscale Data Centers Burned 82 TWh in a Year, Study Finds

A Harvard-led study estimates 403 hyperscale U.S. data centers drew 82 TWh of electricity in a year and emitted 25.7 million metric tons of CO2, two to three times 2018 estimates.

By
Grace Kim
Filed
Channel
AI & Compute
Read
4 min read

Four hundred three hyperscale U.S. data center sites consumed roughly 82 TWh of electricity in a single year, according to a Harvard-led study — about 1.8% of total U.S. generation and more than 80% of the electricity Michigan sold statewide in 2024, which was 99 TWh.

The study, reported by Wired, estimates the sites emitted 25.7 million metric tons of CO2 over the year. That is two to three times a 2018 estimate of 10.5 Mt, with the study's uncertainty range running from 21 to 31 Mt. Fossil fuels supplied 53.9% of the power: 38.2% from gas and 15.2% from coal. Renewables covered 25.3% and nuclear 20.9%.

How were the numbers built?

No agency officially tracks what data centers draw from the grid. The researchers instead multiplied each site's maximum capacity by the hours in a year and an estimate of how hard the facility runs. For emissions, they assigned each site to one of 23 grid regions and applied regional average generation mixes from the Environmental Protection Agency.

The site list came from data vendor Baxtel, which supplied 675 candidate locations. The team confirmed 403 with satellite imagery and building-footprint maps. Average site capacity is 39.9 MW; the median is 36 MW. Hyperscale, in this definition, means a site run by a hyperscale cloud or content provider or a colocation operator serving one, with capacity in the tens of megawatts.

The estimation method carries a caveat the authors acknowledge: imprecise numbers may weaken arguments made against data center buildouts.

Where does the load sit?

Virginia alone hosts 142 of the 403 sites, drawing 21.4 TWh — about a quarter of the total. That single state's draw nearly matches the 22.85 TWh estimate for all U.S. hyperscale sites in 2018. PJM, the grid operator covering 13 states plus Washington, D.C., and America's largest grid, accounts for 32.8 TWh, or roughly 40%.

The top 10 states by attributed CO2:

  • Virginia: 142 sites, 21.4 TWh, 7.0 Mt CO2, 325 g/kWh
  • Iowa: 26 sites, 7.2 TWh, 3.2 Mt, 445 g/kWh
  • Ohio: 38 sites, 8.8 TWh, 2.9 Mt, 325 g/kWh
  • Texas: 22 sites, 4.6 TWh, 1.5 Mt, 333 g/kWh
  • Nebraska: 8 sites, 3.5 TWh, 1.4 Mt, 395 g/kWh
  • Oregon: 56 sites, 11.1 TWh, 1.3 Mt, 119 g/kWh
  • Arizona: 12 sites, 3.2 TWh, 1.3 Mt, 401 g/kWh
  • Oklahoma: 8 sites, 2.9 TWh, 1.1 Mt, 395 g/kWh
  • Wyoming: 2 sites, 1.2 TWh, 1.0 Mt, 848 g/kWh
  • Utah: 5 sites, 1.3 TWh, 0.8 Mt, 640 g/kWh

What about carbon intensity?

The load's estimated carbon intensity is 314 g of CO2 per kWh. That is about 10% below the U.S. grid average of 348 g. The likely reason: the largest loads sit in PJM territory, which is nuclear-heavy, while Oregon leans on hydropower. Almost a quarter of the load, however, sits in regions dirtier than the national average.

The emissions figures rely on 2023 plant data. The method does not trace which plant feeds which site, and it excludes clean-power contracts and credits.

Is the footprint already bigger?

Almost certainly. The study window closed in April 2025. Given the pace of AI-driven development, the footprint "could already be substantially larger now in 2026," lead author Gianluca Guidi said in a statement. The median site capacity in the study also looks small next to planned buildouts, with many new campuses sized in the hundreds or thousands of megawatts. On-site generation, such as gas plants, may not appear in grid data at all.

Official figures are coming. The Energy Information Administration has completed pilot surveys of data center energy use and is using them to design a sector-wide survey, an agency spokesperson told Wired. The first pilots launched in March, using web surveys and in-person interviews; 196 companies answered questions on energy sources, electricity use, servers and cooling, though not all states were covered.

Corresponding author Francesca Dominici said quantifying the health effects of the emissions is an important next step for the research. With AI capacity forecasts still climbing and EIA survey data pending, the measured gap between estimated and reported consumption is set to narrow — while the absolute numbers continue to rise.

Source: Tom's Hardware

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,055 articles

Related articles

« Previous article