SOMA

    Does an AI query use more energy than a Google search?

    For analysts, journalists and sustainability teams who have met the "ten times a search" claim and want its source. Sources checked on 31 October 2026. Factor version soma-ai-ef-2026.10-dataset-v3.0.0. Last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.

    Part of the SOMA AI Emission Factors cluster.

    Does an AI query use more energy than a Google search?

    On the 2024 numbers, yes, by about ten times. On the 2025 numbers for a short text query, the gap is much smaller. The "ten times" claim comes from a 2024 EPRI white paper, which put a ChatGPT request at 2.9 Wh against 0.3 Wh for a traditional Google search, both as estimates with no stated method. In 2025 Google reported 0.24 Wh for its median Gemini prompt, and Epoch AI and Microsoft researchers modelled 0.3 Wh and 0.31 Wh for a short query, which is about the same as the 0.3 Wh search figure. The comparison depends on which query and which boundary.

    Where does the 2.9 Wh against 0.3 Wh claim come from?

    It traces to EPRI's 2024 white paper on data-centre demand, reported by Utility Dive on 30 May 2024. On page 16 EPRI writes that at 2.9 Wh per ChatGPT request, AI queries are estimated to need ten times the electricity of a traditional Google search, at 0.3 Wh. EPRI gives both numbers as estimates and cites one reference for them, a 2023 Joule paper by Alex de Vries. The white paper states no method and no boundary, so it is not clear whether 2.9 Wh counts only the GPUs or the whole system.

    Two things follow. Both numbers are second-hand estimates, and the reference EPRI cites is itself a 2023 estimate and not a measurement of ChatGPT or of search. And the 0.3 Wh search figure has not been remeasured in any of the sources checked here, so every comparison in this page uses it as given.

    What do the 2025 figures say?

    Three groups published newer figures for a short text query, and all three are lower than 2.9 Wh by roughly a factor of ten. Epoch AI wrote in February 2025 that a typical GPT-4o query likely uses about 0.3 Wh, "ten times less than the older estimate". Google's paper of August 2025 reports 0.24 Wh for a median Gemini Apps text prompt, counting chips, host machines, idle capacity and data-centre overhead. Microsoft's Oviedo et al. estimate a median 0.31 Wh for frontier-scale models and say that widely cited estimates are overstated by 4 to 20 times.

    Energy per AI query against the search figure, as published. Estimates, not meter readings.
    Publisher and dateFigureWhat it isWhat it counts
    EPRI (reported by Utility Dive), 30 May 20242.9 Wh against 0.3 WhChatGPT request against a traditional Google searchNot stated. Estimates citing a 2023 paper.
    Epoch AI, 7 February 20250.3 WhGPT-4o, 500 output tokensModelled, inference only, with server and data-centre overhead.
    Google (Elsworth et al.), 21 August 20250.24 WhMedian Gemini Apps text promptActive chips, host CPU and memory, idle machines and overhead. Training excluded.
    Oviedo et al. (Microsoft), September 20250.31 Wh (IQR 0.16 to 0.60)Models above 200 billion parameters, 300 median output tokensModelled bottom-up for H100 nodes, with overhead.
    Sam Altman, June 20250.34 WhAverage ChatGPT queryNot stated. No method is given.

    None of these sources measured a Google search, so the 2025 comparison still uses 0.3 Wh for search. Treat the result as a comparison of two estimates, not a measured ratio.

    Does the answer change for long or reasoning queries?

    Yes, a great deal. Epoch AI estimates around 2.5 Wh for GPT-4o with a 10,000-token input and almost 40 Wh with a 100,000-token input. Oviedo et al. estimate a median 3.91 Wh, with an interquartile range of 2.15 to 7.05 Wh, for test-time-scaling queries with 5,000 median output tokens. On those figures a long query uses several to more than a hundred times the energy of a search. A person asking a long question of a reasoning model is not doing the same thing as a person typing a few keywords. Reasoning is covered on its own page, how much more energy reasoning models use.

    What if every search used AI?

    Alex de Vries estimated in Joule in October 2023 that if every Google search used generative AI, the electricity would be about 29.2 TWh a year. He built it as a scenario from SemiAnalysis's server count, not from a measurement, and New Street Research gives a lower figure of 22.8 TWh. It is a hypothetical from 2023, not a forecast and not a measurement.

    How does the SOMA estimate compare?

    SOMA AI (somaai.earth) estimates one exchange of 1210 tokens in the SOMA AI Emission Factors (SAEF) dataset. The table sets each class against the 0.3 Wh search figure. The comparison is only indicative, because the search figure carries no boundary and a chat exchange of this length is not one query in the EPRI sense.

    Estimated electricity for one exchange, by SAEF model class, against 0.3 Wh. Estimates.
    Model classWh per exchangeAgainst 0.3 Wh
    Class A0.03 Wh0.1 times
    Class B0.099 Wh0.3 times
    Class C0.37 Wh1.2 times

    Classes A and B lie below the search figure and class C, the largest, lies above it. For the models most companies use, the useful answer is the class and the volume, not a single ratio.

    What should a company report?

    Do not report AI as a multiple of a search. Report the activity (tokens, messages or spend), the model class, the region and the factor version, and keep the search comparison out of an inventory, where it has no standing. The Scope 3 method is in the Scope 3 guide, and the free calculator gives a first estimate without an account. The equations are on the methodology page.

    Keep reading