Research

Fossil Fuels Supplied 54% of Power to US Hyperscale Data Centers

A Harvard-led study in Nature Sustainability attributes 82 TWh and 25.7 million metric tons of CO2 to 403 US hyperscale data centers, with fossil fuels supplying 54% of their power.

Harvard Study Finds Fossil Fuels Power Most US Hyperscale Data Centers
Harvard Study Finds Fossil Fuels Power Most US Hyperscale Data Centerspixel2013 / pixabay
By Marcus Bennett5 min read

Why it matters

  • Fossil fuels supplied roughly 54% of electricity attributed to 403 US hyperscale data centers, per a Harvard-led study published November 16, 2026 in Nature Sustainability.
  • Central scenario: ~82 TWh consumed (about 1.8% of US electricity) and ~25.7 million metric tons of CO2, within a range of 21–31 million metric tons.
  • Natural gas supplied 38% of attributed generation; coal supplied 15%.
  • Virginia, Oregon, Ohio, and Iowa together accounted for more than half of the electricity use examined.
  • Lead author Gianluca Guidi said the footprint "could already be substantially larger now in 2026."

Fossil fuels supplied roughly 54% of the electricity attributed to US hyperscale data centers, according to a Harvard-led study published November 16, 2026 in Nature Sustainability. Natural gas accounted for 38% of attributed generation and coal for 15%. The 403 facilities in the study's dataset consumed approximately 82 terawatt-hours (TWh) under the central scenario, or about 1.8% of total US electricity consumption.

The central estimate for attributable carbon dioxide emissions was approximately 25.7 million metric tons. Across facility-load scenarios, the emissions range ran from 21 to 31 million metric tons, and estimated electricity consumption spanned approximately 68 to 99 TWh. Harvard noted that the upper consumption estimate approaches the annual electricity use of states such as Indiana, Michigan, and Tennessee.

The numbers matter for three constituencies at once. Hyperscale operators (companies running massive cloud campuses for services such as cloud computing, large language model training, and global deployment of AI tools) face mounting pressure to disclose facility-level emissions. Policymakers weighing grid expansion need to know where demand actually lands. And investors tracking AI's physical footprint now have a facility-resolved method rather than a single national aggregate.

How did researchers measure 403 data centers?

The estimates cover May 2024 through April 2025. The research team used satellite imagery to compile the facility list, then modeled annual electricity consumption for each site under a range of operating scenarios. They drew on Environmental Protection Agency data to attribute emissions according to the mix of power plants in each facility's regional grid.

The paper describes its headline accounting as a location-based, total-output balancing-authority attribution method. In practice, the approach allocates generation and emissions from plants within the relevant regional grid to the facilities drawing on it. The result is a facility-resolved estimate of electricity sources and emissions, not direct meter readings from every site.

That distinction carries weight. Utility-level and corporate sustainability reports often rely on contractual instruments such as renewable energy credits; this study instead ties each campus to the physical generation mix of its grid region. The method produces a different, and frequently higher, emissions picture than market-based accounting.

Why does the regional grid determine the footprint?

On the paper's electricity-weighted, total-output measure, carbon intensity at the studied facilities was approximately 314 grams of CO2 per kilowatt-hour. The US total-output average was approximately 348 grams. Data centers as a group ran slightly cleaner than the national grid average. The aggregate figure, however, conceals sharp regional variation.

About a quarter of the electricity was consumed in grid regions with carbon intensity above the national average. Many facilities were concentrated in areas where coal-fired generation contributes heavily to supply. Virginia, Oregon, Ohio, and Iowa together accounted for more than half of the electricity use examined.

The implication is structural. A national average cannot represent every campus. The emissions attributed to a facility depend on the generation mix of its regional grid, which means two identical data centers can carry materially different carbon footprints based solely on location. For hyperscalers siting new capacity, grid composition becomes a measurable design variable rather than an externality.

Can the totals be read as AI's electricity bill?

No, and the researchers were explicit about the limitation. Available data did not cover every US hyperscale data center, and the team could not separate electricity used for AI-specialized operations from general-purpose computing at individual facilities. The studied campuses support services including cloud computing, large language model training, and global deployment of AI tools. The totals therefore cannot be interpreted as AI-only consumption.

That gap matters for the ongoing debate over AI's energy demand. Industry forecasts and policy arguments frequently cite AI-specific consumption figures; this study's scope is broader and more conservative. It measures hyperscale facilities as a class, whatever workloads run inside them.

What did the lead authors say?

Lead author Gianluca Guidi said in Harvard's announcement that the reported electricity use and carbon footprint "could already be substantially larger now in 2026." The statement is a projection, not a new measurement. It rests on facilities built and campuses expanded after the May 2024 to April 2025 study period, together with continued growth in AI computing demand.

Corresponding author Francesca Dominici outlined the health dimension. The study did not directly measure health effects, which the researchers plan to investigate in future work. Dominici described two pathways: carbon dioxide drives climate change, affecting health through heat waves, wildfire smoke, flooding, and other pathways; fossil-fuel plants also release fine particulate matter, nitrogen oxides, and sulfur dioxide linked to heart and lung disease and premature death.

The second pathway has an equity dimension. People living near and downwind of those plants may bear health costs even when they are far from the data center itself. Emissions attribution, in other words, does not stop at the campus fence line.

What is the framework for?

Dominici said the goal is to give the public, policymakers, investors, and developers a more concrete basis for decisions about where data centers go and how they are powered. The framework moves the accounting beyond a single national total down to regional grids and their power plants, which is the level at which hyperscale siting, electricity supply, and attributable emissions actually meet.

The study arrives amid intensifying scrutiny of data center energy demand, with utilities, regulators, and grid operators debating how to serve new AI-driven load. Its facility-resolved method gives those disputes a common evidentiary basis. If subsequent work quantifies health effects as planned, the siting conversation may shift from megawatt-hours and tons of CO2 to downstream medical and mortality costs, a change that would raise the stakes for every pending campus approval.

Source: unite.ai

Share this article:

More from Marcus Bennett

Marcus Bennett

Show full bio

Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

718 articles

Related articles

  1. 403 U.S. Data Centers Used as Much Electricity as Michigan
  2. 403 largest U.S. data centers rivaled Denmark’s emissions
  3. Hyperscale data centers used 2% of US electricity in 2024-25
  4. Google's Emissions Jumped 50% in 2024 to Feed Generative AI
  5. Voters Approved 40 of 41 Ballot Measures Restricting AI Data Centers

« Previous articleNext article »