Proceedings · Session S-902 · filed November 16, 2026
Research InfrastructureSession paper
U.S. hyperscale data centers used an estimated 68–99 TWh in 2024–25
Harvard-led researchers estimate 403 U.S. hyperscale data centers used 68–99 TWh in the year ending April 2025, with 25.7 million metric tons of associated CO2.
By Sophie Lindqvist4 min read893 words
Summary
- A Harvard-led team estimated that 403 U.S. hyperscale data centers used 68–99 TWh in the year ending April 2025, or 1.8% of national electricity consumption.
- The estimated fossil-fuel share was about 54%, including around 38% natural gas and 15% coal.
- Associated emissions reached an estimated 25.7 million metric tons of carbon dioxide, more than Denmark's total for the same period.
- Lawrence Berkeley National Laboratory estimated 176 TWh for all U.S. data centers in 2023 and projected 325–580 TWh in 2028.
- The reported U.S. inventory included 1,156 operating data centers in 2025 and 1,860 running or under construction as 2026 came to a close.

By Sophie Lindqvist
An international, Harvard-led team estimated that 403 U.S. hyperscale data centers consumed 68–99 terawatt-hours in the year ending April 2025, equal to 1.8% of total U.S. electricity use.
At the high end of that estimate, their demand approached the annual electricity use of Michigan or Tennessee. The associated power generation would have added an estimated 25.7 million metric tons of carbon dioxide, more than Denmark’s total for the same period.
These are model-based estimates, not sums of facility meters. The analysis cannot separate power used for AI services, including therapy sessions, from storage such as vacation photos. Nor did it quantify the broader claim that data-center demand is erasing electricity savings elsewhere.
“Data centers are growing incredibly fast, driven largely by AI, and decisions about where to build them and how to power them are being made right now,” said Gianluca Guidi, lead author and a biostatistician at Harvard University. Hyperscale facilities, the largest data-center class, handle cloud storage, online services and, increasingly, large-language-model programs.
How did researchers estimate power use?
Detailed figures for individual facilities rarely reach the public, so the team used an indirect method:
- It checked commercial facility records against building footprints and satellite imagery to verify 403 locations.
- It estimated each site’s load from assumptions about capacity, including cooling and continuous 24-hour operation.
- It matched every facility to the grid region responsible for balancing supply and demand.
- It applied 2023 U.S. Environmental Protection Agency power-plant and generation records to assign each facility a share of regional emissions.
These steps produced a geographic estimate of emissions rather than direct facility measurements.
The study summary does not explain how representative the 403-site sample was and names no funding source. Its main methodological limits are the absence of facility-level metering and the use of 2023 regional generation averages for a load year ending April 2025.
How much power came from fossil fuels?
About 54% of the attributed electricity came from fossil fuels, with natural gas supplying around 38% and coal 15%. Virginia accounted for roughly one-quarter of the electricity; Oregon, Ohio and Iowa, added to Virginia, made up more than half.
Across the sample, each kilowatt-hour carried about 314 grams of carbon dioxide, below the comparable U.S. average of 348 grams. Yet nearly one-quarter of demand sat in grid regions with above-average carbon intensity.
For R&D and procurement teams, the calculation makes location a factor in both electricity budgets and emissions. The same digital workload has a different estimated grid footprint depending on the balancing region, so power availability, fossil-fuel exposure and emissions should be assessed together in site and cloud-capacity decisions.
How fast is data-center demand changing?
An earlier study put hyperscale electricity use at about 23 terawatt-hours in 2018, less than one-third of the new central estimate. Different facility inventories and methods restrict the comparison to broad scale rather than a clean growth rate.
Lawrence Berkeley National Laboratory estimated that U.S. data centers of all sizes used 176 terawatt-hours in 2023 and projected 325–580 terawatt-hours in 2028. That range is a forecast, not a measured outcome, and its all-size scope differs from the 403-site hyperscale estimate.
For R&D managers, the figures set two planning horizons: near-term requirements for site power and cooling, and longer-term exposure to grid capacity, emissions and health constraints as AI workloads expand.
What are the public-health implications?
The emissions estimate addresses climate change, but fossil generation also creates a separate air-quality exposure. Fossil-fuel combustion can affect people who live nowhere near a data center.
A separate University of California, Riverside analysis modeled California’s data-center air-pollution health burden at roughly three times its 2019 level by 2023, as the sector’s electricity use nearly doubled.
“Decisions about where data centers go and how they’re powered are critical environmental and public health decisions, and they should be made with good data,” said study co-author Francesca Dominici, also a biostatistician at Harvard.
The available summary gives no model specification, uncertainty range, sample size or funding source for that California analysis, so its threefold result remains a modeled estimate rather than a count of observed health outcomes.
What should R&D managers check now?
The United States had 1,156 data centers in operation in 2025. As 2026 comes to a close, the reported total reaches 1,860 facilities running or under construction, making the new electricity estimates partly obsolete already.
Those totals cover different categories, so they do not by themselves establish a one-year growth rate. For R&D portfolio reviews, four checks matter:
- Reconcile proposed capacity with commercial records, satellite-verified footprints and metered data where available.
- Price electricity by balancing region and generation mix, tracking natural gas and coal exposure separately.
- Test cooling and 24-hour operating assumptions separately from computing workload.
- Update carbon and public-health estimates before construction, not after switch-on.
The Harvard-led work appeared in Nature Sustainability. For hyperscale facilities not yet built, the research leaves a narrow timetable: decisions on location and power made now will shape electricity demand, emissions and public-health costs long after the sites switch on.
via newatlas.com (Source)
Filed under
- hyperscale-data-centers
- data-center-energy-use
- ai-infrastructure
- grid-carbon-intensity
- sustainable-computing
More from Sophie Lindqvist
Show full bio
Correspondent covering money, funding and market shifts around research and development at Hypothesis Wire.
242 articles
References
- Google signs 50 MW deal for superhot geothermal power
- AI data centers drive $26B climate-tech surge, carbon removal left behind
- AI Data Center Buildout Runs on Debt: Nearly $500 Billion Issued
- ANU Atlas Counts 820,000 Pumped Hydro Sites, 200 Times Global Need
- One in Four Data Centers Sits in a Water-Stressed Basin, Study Finds