A new Harvard University study published in Nature Sustainability reveals that the 403 largest hyperscale data centers in the United States consumed roughly 82 terawatt-hours (range 68–99 TWh)—about 1.8 percent of total U.S. electricity—between May 2024 and April 2025, releasing an estimated 25.7 million metric tons (28.3 million short tons) of carbon dioxide (CO2) in a single year, an annual greenhouse gas footprint comparable to the entire country of Denmark or the state of Michigan.[1][3][4] Released alongside international climate negotiations at the United Nations COP31 summit in Antalya, Turkey, the findings arrived just as negotiators unveiled a voluntary "Clean AI Pledge" aimed at curbing the energy, water, and carbon footprint of rapid artificial-intelligence expansion.[3]
Crucially, the Harvard researchers built their baseline from satellite imagery and U.S. Environmental Protection Agency (EPA) grid records captured from May 2024 to April 2025, before the newest wave of ultra-large AI training clusters came fully online.[1][4] That means current and near-term electricity demand—and associated smokestack pollution—is poised to climb substantially unless data-center operators shift to zero-carbon power.[3][4]
What Satellite and Grid Data Reveal About Hyperscale Pollution
Because technology companies rarely publish facility-by-facility power and emissions breakdowns, tracking data-center pollution has long relied on rough industry estimates.[3][4] To overcome that transparency gap, the Nature Sustainability team combined high-resolution satellite imagery of physical data-center campuses with local electrical-grid fuel-mix data from the EPA.[1][4]
Across the 403 major U.S. hyperscale facilities analyzed, roughly 54 percent of the electricity powering their servers and cooling systems came from fossil-fuel power plants, primarily natural gas (~38 percent) and coal (~15 percent), with four states—Virginia, Oregon, Ohio, and Iowa—accounting for more than half of the studied electricity use.[1][4] A separate 2026 National Bureau of Economic Research (NBER) working paper by Carnegie Mellon University economist Nicholas Z. Muller evaluated roughly 2,800 operational U.S. data centers and estimated about $25 billion in gross external damages from localized air pollution and greenhouse gases, with impacts heavily concentrated in high-growth hubs such as Texas and Virginia.[2]
| Environmental Metric | Observed U.S. Baseline (2026 Studies) | Policy / Grid Implication |
|---|---|---|
| Hyperscale Facilities Analyzed | 403 major U.S. campuses (Nature Sustainability) | May 2024–April 2025 baseline predates newest AI scaling |
| National Electricity Share | ~82 TWh (range 68–99 TWh; ~1.8% of U.S. electricity) | Roughly 54% generated from fossil fuels (38% gas, 15% coal) |
| Annual Carbon Emissions | ~25.7 million metric tons (28.3 million short tons) CO2/year | Comparable to annual emissions of Denmark or Michigan |
| Broader Air & Climate Damages | ~$25 billion across ~2,800 facilities (NBER) | Concentrated in grid hubs like Texas and Virginia |
Inside the COP31 Clean AI Pledge
Against that backdrop of rising power demand, delegates at the COP31 climate talks introduced the voluntary Clean AI Pledge.[3] The framework invites national governments and technology firms to commit to powering AI and cloud data centers with verifiable clean electricity—such as wind, solar, or other low-emission generation—while tightening disclosure standards for both electricity and cooling-water consumption.[3]
A central focus of the pledge is emissions accounting integrity. Climate negotiators and independent researchers have warned that long-term "net-zero" pledges or unbundled carbon offsets can obscure real-time smokestack pollution when fossil-fuel plants are kept running or restarted to meet round-the-clock data-center loads.[3] The Clean AI Pledge pushes instead for transparent, facility-level reporting of present-day emissions.[3]
Why Grid Cleanliness Determines AI's Climate Footprint
Data centers require continuous, uninterrupted electricity 24 hours a day to run high-density computer chips and remove waste heat. When new campuses plug into regional grids powered largely by fossil fuels, every additional megawatt of computing capacity translates directly into higher carbon dioxide and local air-pollutant emissions—including fine particulate matter, sulfur dioxide, and nitrogen oxides that affect nearby and downwind communities.[2][4]
Pairing the new empirical satellite baselines with transparent reporting under initiatives like the Clean AI Pledge gives grid planners and policymakers clearer visibility into where computing growth is straining fossil-heavy power networks—and where clean-energy siting can decouple artificial intelligence from rising climate pollution.[1][3][4]
Updates… (Nov 16, 2026): Corrected the U.S. hyperscale data-center electricity and emissions figures to match the published November 16, 2026 Nature Sustainability study (Guidi et al., 2026): central-scenario electricity consumption of ~82 TWh (range 68–99 TWh, ~1.8% of total U.S. electricity), ~25.7 million metric tons (28.3 million short tons) of CO2 emissions, and ~54% fossil-fuel generation (~38% natural gas, ~15% coal), replacing overstated figures of 52 million tons, 2.1%, and 60%; added full author metadata for Ref. [1], replaced the generic LAist homepage with the direct report URL, and normalized title typography.
References & Primary Sources
- [1] Guidi, G., Dominici, F., Squartini, T., Sprinkle, B., Gilmour, K., Butler, J., Bell, B., Delaney, K., & Bargagli-Stoffi, F. J. (2026). Assessing the carbon emissions and energy consumption of US hyperscale data centres. Nature Sustainability. — https://doi.org/10.1038/s41893-026-01985-7
- [2] Muller, N. Z. (2026). Measuring the Impact of Data Centers in the United States Economy: Monetary Damage from Air Pollution and Greenhouse Gas Emissions. NBER Working Paper No. 35100. — https://doi.org/10.3386/w35100
- [3] Borenstein, S., & McDermott, J. / LAist & Associated Press (2026). Study estimates magnitude of US data center pollution as climate talks unveil clean AI pledge. — https://laist.com/brief/news/climate-environment/study-estimates-magnitude-of-us-data-center-pollution-as-climate-talks-unveil-clean-ai-pledge
- [4] Harvard T.H. Chan School of Public Health (2026). Hyperscale data centers’ energy consumption, carbon emissions newly estimated by researchers. — https://hsph.harvard.edu/news/hyperscale-data-centers-energy-consumption-carbon-emissions-newly-estimated-by-researchers/