SOMA

    How much CO2 does artificial intelligence produce?

    For readers who want one honest answer on AI and CO₂. Sources checked on 12 November 2026. SOMA figures are estimates from the open SAEF dataset. Factor version soma-ai-ef-2026.11-dataset-v3.1.0. Factors last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.

    Page updated

    How much CO2 does artificial intelligence produce?

    There is no single figure, because the answer depends on whether you mean one query or the whole sector. For one text exchange of 1210 tokens, the SOMA AI estimate runs from 0.016 g CO₂e for a small model to 0.2 g for a frontier model at the world-average grid, and a mid-size model emits between 0.0082 g and 0.095 g depending on the grid of the data centre. For the sector, the International Energy Agency estimates that all data centres, AI included, used about 415 TWh of electricity in 2024, and put the CO₂ from that electricity at about 180 million tonnes. Nobody has published a reliable CO₂ total for AI alone.

    How much CO2 does one AI query produce?

    One text exchange of 928 input and 282 output tokens, at the world-average grid. Estimates from the open dataset.
    Model classEnergy, WhCO₂e, g
    Class A, small0.030.016
    Class B, mid-size0.0990.053
    Class C, frontier0.370.2

    The grid moves the number. A mid-size exchange comes to 0.0082 g on the grid of Quebec (Canada) and 0.095 g on the grid of Bahrain, the cleanest and most carbon-intensive grids that host a major cloud region in the dataset. Providers publish figures too. Google reports 0.03 g CO₂e for a median Gemini Apps text prompt, using a market-based method. Mistral AI, with Carbone 4 and ADEME, reports 1.14 g for a 400-token Le Chat response, including upstream impacts such as server manufacturing. The two differ in boundary and method, so they are not a like-for-like range. For the reasons, see why AI footprint estimates differ.

    How much CO2 does AI produce in total?

    Sector-level figures from the International Energy Agency's Energy and AI report, April 2025. Data centres of all kinds, not AI alone.
    FigureValueStatus
    Data-centre electricity, 2024About 415 TWh, 1.5% of world electricityEstimate
    Data-centre electricity, 2030About 945 TWh, just under 3% of world electricityBase Case projection
    CO₂ from data-centre electricity, baselineAbout 180 million tonnesEstimate
    CO₂ from data-centre electricity, 2035About 300 million tonnes in the Base Case, up to 500 million in the Lift-Off CaseProjections

    The IEA report does not split these totals into AI and other computing, so a statement that AI produces a given number of tonnes of CO₂ needs its own source and boundary. Our site collects the sourced figures on the AI environmental impact statistics page.

    Why are totals so uncertain?

    • Providers do not publish energy per model, per customer or per region, so totals are built from server counts and assumptions.
    • Data-centre totals mix AI with search, storage, streaming and other cloud work.
    • Carbon depends on the grid and on whether the method is location-based or market-based.
    • Training, hardware manufacturing and construction are counted by some studies and not by others.

    A company that wants its own number should estimate from its usage, as set out on the company footprint how-to. The estimate method is on the methodology page, and the paper was accepted for publication in Sustainable Production and Consumption in November 2026.

    Keep reading