SOMA

    Does AI really drink a bottle of water per email, and what does the evidence say?

    For journalists, sustainability teams and anyone who has met the bottle-of-water claim and wants its source. Sources checked on 12 November 2026. SOMA figures are estimates for one exchange of 1,210 tokens on a mid-size model, from the open SAEF dataset.

    Page updated

    Does AI really drink a bottle of water per email, and what does the evidence say?

    No. The bottle figure is a 2024 estimate that has aged badly, and it counted the water behind the electricity as well as the water used for cooling. The Washington Post and researchers at UC Riverside put one 100-word GPT-4 email at about 519 mL in September 2024. SOMA AI estimates 0.15 mL to 0.94 mL for a whole mid-size chat exchange across 14 cloud regions, which is roughly 550 to 3,550 times lower. Google reports 0.26 mL for its median Gemini text prompt.

    Where did the bottle of water per email come from?

    It traces to a line of work by Shaolei Ren's group at UC Riverside. The first paper, by Li, Yang, Islam and Ren, was posted on arXiv in April 2023. Its current version (March 2025) estimates that GPT-3 needs to "drink" a 500 mL bottle of water for roughly 10 to 50 medium-length responses, depending on where and when it runs (arXiv 2304.03271). A response there means about 800 words in and 150 to 300 words out, at an assumed 0.004 kWh of server energy, and the authors call their inference estimate conservative. Ten to fifty responses per 500 mL is 10 to 50 mL per response.

    The email version came next. A Washington Post analysis with UC Riverside in September 2024 put one 100-word email written with GPT-4 at about 519 mL, counting direct cooling and the water embedded in the facilities that power the model (quoted by The American Prospect). We could not open the Post article itself, so the quote and scope come from that report. The estimate used the energy figure available for GPT-4 in mid 2024 and the cooling water ratios companies had disclosed by then.

    How do the published water figures per prompt compare?

    Water per prompt or response, as published, and the SOMA estimate. Boundaries differ between rows, so read the last column before comparing.
    Source and dateFigureModel and promptWhat it counts
    Li et al., UC Riverside, April 2023, revised March 202510 to 50 mL per responseGPT-3, medium responseCooling plus electricity water, 0.004 kWh per response. Inference assumed conservative.
    Washington Post with UC Riverside, September 2024, as reported by The American Prospect519 mLGPT-4, one 100-word emailCooling plus the water behind the electricity, mid-2024 inputs.
    Google (Elsworth et al.), 21 August 20250.26 mLMedian Gemini Apps text promptFull serving stack at the data-centre cooling ratio. The narrower active-chip method gives 0.12 mL.
    Sam Altman, June 20250.32 mL (0.000085 US gallons)Average ChatGPT queryNot stated in the post.
    SOMA AI estimate, November 20260.15 to 0.94 mL928 tokens in, 282 out, class B modelOn-site cooling (0.00 to 0.07 mL) plus electricity water, by cloud region.

    The lowest SOMA value is US South Central - Texas and the highest is US West - Oregon. The full ranking is on the water by cloud region page.

    Why do the figures differ by so much?

    • Energy per prompt fell. Google reports a 33 times drop in energy per median Gemini prompt between May 2024 and May 2025 for its own service. Estimates built on older energy-per-prompt inputs come out higher.
    • Which water is counted. On-site cooling is the small part. The water used to generate the electricity is the larger part in every SOMA region, and it depends on the grid. Counting both, as the 2024 estimate did, gives a bigger number than cooling alone.
    • The request itself. A 100-word email asks for generated text, and a chat exchange of 1210 tokens mostly processes input. Output length moves energy and water together.
    • Where it runs. The regional spread in our estimates is about 6.5 times for the same exchange, driven by the grid and the cooling design.
    • Method. Google measured its own fleet and publishes its method. SOMA publishes the method behind its estimates. Altman's post gives the figure without a method.

    Water is still a real issue at the level of the data centre and the local basin, and it grows with volume. The claim that fails is the per-email one. For the scarcity side, read the water guide, and for the same disagreement on energy and carbon, why AI footprint estimates differ.

    What are the limits of the SOMA estimate?

    • It is an estimate from published cooling ratios and grid water factors, not a provider reading. The equations are on the methodology page.
    • Water for making chips and servers is outside the boundary, as in every published per-prompt figure.
    • Longer prompts, reasoning modes and agents multiply the figures.
    • The method paper was accepted for publication in Sustainable Production and Consumption in November 2026.

    Keep reading