A Harvard-led study estimates that 403 hyperscale U.S. data center sites consumed about 82 TWh of electricity in a year, 1.8% of the country’s total, with more than half coming from gas and coal sources, Wired reports. This compares to the entire state of Michigan selling 99 TWh in 2024. The data centers’ pull also put out an estimated 25.7 million metric tons (Mt) of CO2, which is two to three times a 2018 estimate.

Hyperscale data centers, in this case, refer to sites run by a hyperscale cloud or content provider, or a colocation operator who serves one, with a capacity in the tens of megawatts. The 675 candidate sites came from data vendor Baxtel, with 403 confirmed with satellite imagery and building-footprint maps, at an average capacity of 39.9 MW and a median of 36 MW.

Nobody officially counts what the data centers actually draw from the grid. Instead, the electricity is estimated by multiplying each site’s maximum capacity by the hours in a year and an estimate of how hard the data centers run. For CO2 and the source mix, each site is placed in its grid region, of 23 total in the study, with a regional average mix from the Environmental Protection Agency (EPA). These are then multiplied and added up for the total number. Such estimation methods may reduce the strength of arguments against data center buildouts.

The power is also not evenly laid out, with 142 of the sites located in Virginia alone, taking 21.4 TWh, or about a quarter. PJM, a grid operator for 13 states, including Virginia, and D.C., or America’s largest grid, covers 32.8 TWh, or about 40%, more than any other region. Notably for Virginia, its power draw is actually just a bit below the 22.85 TWh estimate for all U.S. hyperscale sites from 2018. See the table for more details.

Top 10 states by CO2 attributed to hyperscale data centers

State

Sites

Electricity (TWh)

CO2 (million metric tons)

CO2 per kWh (g)

Virginia

142

21.4

7.0

325

Iowa

26

7.2

3.2

445

Ohio

38

8.8

2.9

325

Texas

22

4.6

1.5

333

Nebraska

8

3.5

1.4

395

Oregon

56

11.1

1.3

119

Arizona

12

3.2

1.3

401

Oklahoma

8

2.9

1.1

395

Wyoming

2

1.2

1.0

848

Utah

5

1.3

0.8

640

Not all power is equal — renewables, for example, can be perceived more favorably — and 53.9% of the estimated draw comes from fossil fuels, 38.2% of the total from gas and 15.2% from coal. Renewables cover 25.3% while nuclear handles the remaining 20.9%. Estimates put this in the 21–31 Mt range for CO2 output, versus the 10.5 Mt estimate from 2018.

Stacked bar chart showing the share of coal, gas, nuclear, hydro, solar and wind in the electricity attributed to US hyperscale data centers, in total and for seven grid regions.

Estimated fuel mix of the electricity attributed to 403 U.S. hyperscale data centers (“US Total”) and to the seven grid regions with the most data center load. (Image credit: Guidi et al. / arXiv (CC BY 4.0))

Estimated carbon intensity, though, is lower than the average, at 314 g of CO2 per kWh versus 348 for the entire U.S. grid, or about 10% lower. Admittedly, almost a quarter of the load is in regions dirtier than the U.S. average. The likely reason it’s lower is that the biggest loads lie in PJM territory, which is nuclear-heavy, while Oregon, for its part, is hydro-heavy. The estimates, in this case, use 2023 plant data, and the method does not trace which plant feeds which site and also leaves out clean-power contracts and credits.

The study’s results also lag reality as the window ended in April 2025 and, with how much development there has been for AI, among other things, the footprint “could already be substantially larger now in 2026,” lead author Gianluca Guidi said in a statement. Another issue is that the median site capacity in the study is relatively small compared to upcoming buildouts, with many campuses now planned in the hundreds or thousands of megawatts. On-site power, such as gas plants, also may not show up in grid data.

Official numbers may be forthcoming, though, as the Energy Information Administration’s pilot surveys of data center energy use are complete, and the agency is using them to design a survey of the sector, a spokesperson told Wired. The first pilots, which launched in March, relied on web surveys and in-person interviews, with 196 companies asked about energy sources, electricity use, plus servers and cooling, although not covering all states. Gauging the direct human impact is next in the chain, with Francesca Dominici, the study’s corresponding author, saying that quantifying health effects is an important next step.