Data on Data Centers

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.

What's new

    Data center electricity use and carbon dioxide: published estimates

    Scope, years and methods differ between studies, so figures from different studies are not directly comparable.

    Lawrence Berkeley National Laboratory (2024 and 2026)

    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.

    YearElectricity, terawatt-hours (TWh) per yearShare of United States electricityReport
    2023 (estimate)1764.4%2024 report (December 2024)
    2028 (projection)325-5806.7-12.0%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).

    Energy and carbon per AI prompt and per token

    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.

    SourceWhat it coversEnergy per prompt, watt-hours (Wh)CO2 per prompt, gramsPer million output tokens (our arithmetic)Boundary and method
    Google, Elsworth et al. (August 2025)Median Gemini Apps text prompt, May 20250.24 (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 serversTypical (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 reportedabout 1,000 Wh typical; about 780 Wh long reasoningModeled 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 query0.34not reportednot reportedCompany statement; no method published.
    Epoch AI (February 2025)Typical ChatGPT query, GPT-4oabout 0.3not reportednot reportedIndependent 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 reported1.14 CO2 equivalent (life cycle)about 2.85 kilograms CO2 equivalentLife 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.

    One prompt as the baseline

    • A typical text prompt uses about 0.3 Wh. About 3,300 such prompts add up to 1 kilowatt-hour (kWh).
    • At the 314 g/kWh estimated for hyperscale data centers by Guidi et al. (location-based), one typical prompt corresponds to about 0.1 grams of CO2. About 10,600 prompts add up to 1 kilogram of CO2.
    • One long reasoning prompt (3.91 Wh) corresponds to about 1.2 grams of CO2 at the same rate. About 820 add up to 1 kilogram.
    • Google's 0.03 grams is market-based: it credits Google's clean energy purchases. On the local grid mix (location-based), the same prompt would be roughly three times higher, by outside estimates.

    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.

    Other efficiency measures in use

    • Power usage effectiveness (PUE): total facility electricity divided by electricity used by computing equipment. 1.0 would mean no cooling or other overhead.
    • Carbon usage effectiveness (CUE): kilograms of CO2 per kWh used by computing equipment; combines PUE with the grid mix.
    • Energy per token, or tokens per joule: used in research benchmarks of AI hardware and serving software. Results vary by 10 to 100 times with batch size, context length, numeric precision and chip generation.
    • AI Energy Score (Hugging Face, Salesforce, Cohere and Carnegie Mellon University): GPU energy per 1,000 queries and a one-to-five star rating for openly available models across common tasks.

    References