Public data on the environmental footprint of data centers: air sensor readings, noise measurements, water use, electricity use and carbon dioxide estimates, and updates on legal, regulatory and scientific developments.
Scope, years and methods differ between studies, so figures from different studies are not directly comparable.
Department of Energy-funded estimates from Lawrence Berkeley National Laboratory (LBNL), the longest-running series on United States data center electricity use (reports in 2008, 2016, 2024 and 2026). Covers all United States data centers. LBNL estimates electricity only, not carbon dioxide.
| Year | Electricity, terawatt-hours (TWh) per year | Share of United States electricity | Report |
|---|---|---|---|
| 2023 (estimate) | 176 | 4.4% | 2024 report (December 2024) |
| 2028 (projection) | 325-580 | 6.7-12.0% | 2024 report (December 2024) |
| 2030 (projection, reference case) | 649 (range 521-843) | 11.8% (range 9.5-15.3%) | 2025 update (June 2026) |
Method: a bottom-up model built from shipments of servers, storage and network equipment, per-device power and use, and facility cooling overhead. The 2023 figure excludes cryptocurrency mining. The 2030 range combines four sensitivity cases: fewer equipment installations (578 TWh), more graphics processing unit (GPU) shipments (664 TWh), shorter artificial intelligence (AI) chip lifetimes (590 TWh) and higher idle power and utilization of AI servers (782 TWh).
Guidi, Dominici and colleagues (Harvard T.H. Chan School of Public Health and others), Nature Sustainability, published November 16, 2026. Covers 403 hyperscale data centers (the largest class) operating May 2024 to April 2025. These are estimates, not measurements: see the notes below the table.
| Measure | United States | Virginia |
|---|---|---|
| Hyperscale data centers in the study | 403 | 142 |
| Electricity use, terawatt-hours (TWh) per year | 68-99 (central estimate about 82) | about 21 (central estimate) |
| Share of United States electricity use | about 1.8% | about one quarter of the study's hyperscale total |
| Attributed carbon dioxide (CO2), million metric tons per year | 21-31 (central estimate about 25.7) | about 7.0 (central estimate) |
| Carbon intensity, grams of CO2 per kilowatt-hour (g/kWh) | 314 (national average: 348) | 325 (PJM grid region) |
PJM is the regional grid operator whose territory includes Virginia. Of the electricity attributed to hyperscale data centers in PJM, 60.0% came from fossil fuels (45.5% natural gas, 13.8% coal), 33.1% from nuclear and 6.8% from renewables. Nationally, about 54% came from fossil fuels (38% natural gas, 15% coal), 21% from nuclear and 25% from renewables.
| State | CO2 (million metric tons per year) | Electricity (TWh per year) | Hyperscale data centers | Carbon intensity (g/kWh) |
|---|---|---|---|---|
| Virginia | 6.95 | 21.4 | 142 | 325 |
| Iowa | 3.19 | 7.2 | 26 | 445 |
| Ohio | 2.87 | 8.8 | 38 | 325 |
| Texas | 1.53 | 4.6 | 22 | 333 |
| Nebraska | 1.40 | 3.5 | 8 | 395 |
| Oregon | 1.33 | 11.1 | 56 | 119 |
| Arizona | 1.29 | 3.2 | 12 | 401 |
| Oklahoma | 1.14 | 2.9 | 8 | 395 |
| Wyoming | 1.04 | 1.2 | 2 | 848 |
| Utah | 0.81 | 1.3 | 5 | 640 |
Source: state-level data published with the paper (GitHub repository). The paper uses location-based accounting: emissions are assigned from the average mix of power plants in each grid region, using the Environmental Protection Agency's (EPA) 2023 power plant database (eGRID), so they occur at those plants and not necessarily in the same state. Renewable energy purchases by the companies are not counted. Electricity use is estimated from facility capacity times an assumed average utilization (58% in the central estimate, 48-70% across scenarios). Cryptocurrency mining and on-site backup generators are not included.
A June 2026 preprint of this study reported higher figures (37-54 million metric tons of CO2; 545 g/kWh). It divided power plant emissions by generation from fuel-burning plants only; the published version divides by all generation, including nuclear, hydro, wind and solar.
| Scope | About 2,800 United States data centers, 2025 |
| Reported figures | About 250 TWh per year (5-6% of United States generation). Air pollution and greenhouse gas damages about $25 billion per year (range $10-33 billion); Texas and Virginia account for 30% of the total |
| Type | Modeled damages; working paper |
Working paper; not peer reviewed. NBER Working Paper 35100.
| Scope | Virginia data center electricity load |
| Reported figures | About 31 TWh per year with no further growth; 147 TWh per year (moderate growth) to 263 TWh per year (unconstrained) in 2024-2040 projections. Without the Virginia Clean Economy Act, growth could raise electric sector greenhouse gas emissions by up to about 80% |
| Type | Projection |
Virginia legislative staff study of data center load and its effect on the state grid. Presentation (PDF).
Facility totals show scale, not efficiency. These figures express energy and carbon per unit of AI work: per prompt (one question and its answer) and per token (a word fragment; a typical answer is a few hundred tokens). They cover AI text generation only, not all data center work.
| Source | What it covers | Energy per prompt, watt-hours (Wh) | CO2 per prompt, grams | Per million output tokens (our arithmetic) | Boundary and method |
|---|---|---|---|---|---|
| Google, Elsworth et al. (August 2025) | Median Gemini Apps text prompt, May 2025 | 0.24 (0.10 counting AI chips only) | 0.03 CO2 equivalent (market-based) | not reported (token count not given) | Measured in Google's production fleet. Includes AI chips, host processors and memory, idle standby machines and data center overhead. Excludes networking, user devices and model training. |
| Microsoft, Oviedo et al., Joule (2026) | Frontier models (over 200 billion parameters) on NVIDIA H100 servers | Typical (about 300 output tokens): 0.31 (interquartile range 0.16-0.60). Long reasoning (about 5,000 output tokens): 3.91 (2.15-7.05) | not reported | about 1,000 Wh typical; about 780 Wh long reasoning | Modeled from measured token throughput under production-like batching, with data center overhead. Finds that combined model, software and hardware gains could cut energy per prompt 8 to 20 times. |
| OpenAI, Altman (June 2025) | Average ChatGPT query | 0.34 | not reported | not reported | Company statement; no method published. |
| Epoch AI (February 2025) | Typical ChatGPT query, GPT-4o | about 0.3 | not reported | not reported | Independent estimate from hardware and model assumptions; the authors describe it as erring high. |
| Mistral AI with Carbone 4 and ADEME (July 2025) | One 400-token response from Le Chat (Mistral Large 2) | not reported | 1.14 CO2 equivalent (life cycle) | about 2.85 kilograms CO2 equivalent | Life cycle assessment, so it includes a share of hardware manufacturing and model training, not just electricity. Reviewed by Resilio and Hubblo. Also reports 45 milliliters of water per response. |
Per-prompt efficiency and total demand move in opposite directions. Google reports a 33-fold drop in energy per median prompt between May 2024 and May 2025, while LBNL projects total data center electricity use to keep rising through 2030. Long reasoning and agent tasks use many more tokens per prompt, and a modest share of them can more than double total energy, according to the Microsoft study.