SOMA

    What explains the range between 0.3 Wh and 3 Wh per AI query?

    For analysts and sustainability teams who see 0.3 Wh in one place and 3 Wh in another. Sources checked on 12 November 2026. SOMA figures are estimates from the open SAEF dataset.

    Page updated

    What explains the range between 0.3 Wh and 3 Wh per AI query?

    Mostly the age of the estimate and the length of the query. The 3 Wh end comes from a 2024 EPRI figure of 2.9 Wh per ChatGPT request, which traces to a 2023 estimate built from server counts. The 0.3 Wh end comes from 2025 estimates and provider figures for a short text query. Within 2025 numbers, the boundary (chips only or the whole facility), the model size and the number of output and reasoning tokens decide the rest, and a long or reasoning query can exceed 3 Wh on current hardware.

    Which published figures sit at each end?

    Energy per query, as published. Wh is watt-hours. Training is excluded wherever a boundary is stated.
    Source and dateFigureQueryBoundary
    EPRI white paper, reported 30 May 20242.9 WhOne ChatGPT request, against 0.3 Wh for a Google searchTraces to de Vries (2023), servers and their power rating divided by daily requests. Boundary not stated.
    Sam Altman, June 20250.34 WhAverage ChatGPT queryNot stated.
    Google (Elsworth et al.), 21 August 20250.24 WhMedian Gemini Apps text promptActive chips, host CPU and memory, idle machines, data-centre overhead.
    Google, 21 August 20250.10 WhSame promptActive TPU and GPU consumption only.
    Epoch AI, 7 February 20250.3 Wh, then 2.5 Wh and almost 40 WhGPT-4o, 500 output tokens, then 10,000 and 100,000 input tokensModelled, with server and data-centre overhead.
    Oviedo et al. (Microsoft), September 20250.31 Wh, then 3.91 WhModels above 200 billion parameters, 300 median output tokens, then reasoning-style queries with 5,000Modelled for H100 nodes, with overhead.

    Which factors move the number, and by how much?

    Drivers of energy per query. SOMA values are from the open dataset and are estimates.
    DriverWhat changesSize of the effect
    Model size classParameters served per tokenSOMA estimates 0.03, 0.099 and 0.37 Wh for the same 1210-token exchange in classes A, B and C, a factor of 12 from smallest to largest.
    Output lengthTokens generated and tokens readEpoch AI goes from 0.3 Wh to about 2.5 Wh when the input grows to 10,000 tokens. The SOMA exchange is 928 tokens in and 282 out.
    Reasoning tokensHidden thinking before the answerMicrosoft's test-time-scaling queries come to a median of 3.91 Wh, about 13 times its short-query median. SOMA multiplies estimated output tokens for reasoning and declares it as an assumption.
    Hardware and batchingChip generation and how full the batches areSOMA assumes a utilisation of 0.75 of provisioned capacity, with 0.40 to 0.80 as its range. A lightly loaded server spends the same idle power on fewer queries.
    BoundaryChips only, whole server, or whole facilitySOMA multiplies GPU energy by 1.50 for the host server and 1.10 for idle capacity. Google's own two methods differ by a factor of 2.4, from 0.10 to 0.24 Wh.
    Facility overhead (PUE)Cooling and power conversionSOMA uses a PUE of 1.20, with 1.09 to 1.45 as its range. Google reports a fleet average of 1.09.
    Age of the estimate2023 and 2024 inputs against 2025 inputsThe 2.9 Wh figure predates the 2025 measurements. Google reports a 33 times fall in energy per median prompt from May 2024 to May 2025 for its own service.

    The first four rows change the work done per query. The last three change what is counted around it. A fair comparison needs the same row settings on both sides.

    Why does the 2.9 Wh figure look so different from 0.3 Wh?

    EPRI gave 2.9 Wh for a ChatGPT request and 0.3 Wh for a Google search, a search figure Google published in 2009.Epoch AI traces the 2.9 Wh to a 2023 Joule paper by Alex de Vries, which built on a February 2023 SemiAnalysis estimate of the servers running ChatGPT, assumed 2,000 output tokens per query and older A100 hardware, and did not measure queries. The 2025 numbers come from three different routes. Google measured its own serving fleet. Epoch AI and the Microsoft researchers built bottom-up models for short text queries. Altman gave a figure without a method. They agree within a factor of two for a short query, between 0.24 and 0.34 Wh, while 2.9 Wh sits about ten times higher. A long input, a reasoning mode or the largest model class reaches several Wh in 2025 data as well, so a single number for the average query hides the mix.

    What should a company use?

    Pick the model class and the token counts of your own workload, state the boundary, and report a range. SOMA's class estimates for a 1210-token exchange are in the factor table, and the terms above are defined on the methodology page. The method paper was accepted for publication in Sustainable Production and Consumption in November 2026. For all published figures side by side with carbon and water, see why AI footprint estimates differ.

    Keep reading