The water footprint of AI: cooling water, the water behind the electricity, and scarcity

Published answers run from about a tenth of a millilitre to over 200 mL per query, and most of that spread is about what is counted rather than about the models. Google reports 0.26 mL for its median Gemini Apps text prompt, counting the water consumed at its data centres. Li and co-authors, counting the water used to generate the electricity as well, estimated 16.9 mL for a medium-length GPT-3 response at the US average. The SOMA AI (somaai.earth) factors give about 0.24 mL for a mid-size exchange of 1210 tokens served from US East - N. Virginia, on-site cooling plus generation water. These are estimates. Before comparing any two figures, check which of the three kinds of water below each one includes.

• On-site cooling water. Water evaporated or otherwise consumed at the data centre, mostly in cooling. Expressed through water usage effectiveness (WUE), in litres per kWh. Li and co-authors call it scope-1 water.
• Water for the electricity. Water consumed at power plants to generate the electricity the data centre draws, expressed as an energy water intensity factor (EWIF), in litres per kWh. Li and co-authors call it scope-2 water. It depends on the grid mix, not on the data centre.
• Scarcity. Not a third volume but a weight. A litre taken where water is scarce matters more than a litre taken where it is plentiful. The AWARE 2.0 factor expresses that, with 1 as the world average.

Two counting choices sit across all three. Withdrawal is water taken from a source, consumption is the part that is not returned, so the two cannot be added. And water evaporated from hydropower reservoirs can be allocated to the electricity or left out, which changes the answer a lot on hydro-heavy grids. The manufacturing of chips and servers also uses water, and none of the per-query figures on this page include it.

Read the boundary before the number. The table sorts the figures checked against their primary sources by the kind of water they include. The figures guide has the energy and carbon values from the same sources.

In Li's US-average figure, 87% of the water is off-site, at the power plants. That is why a figure counting only on-site water is not comparable with one that counts both, even before model, hardware and year differ. The same split appears in training. Li and co-authors estimate that training GPT-3 in Microsoft's US data centres could directly evaporate 700,000 litres on-site, out of a total of 5.4 million litres including off-site water.

By more than an order of magnitude between sites. Among the regions SAEF prints, published WUE runs from 0.02 L per kWh in EU North - Sweden to 1.57 in Asia Pacific - Singapore. Google's fleet average, the basis of its 0.26 mL, is 1.15 L per kWh. Many providers publish WUE as withdrawal. SAEF converts those figures to consumption with one documented factor, 0.67, so on-site and electricity water are on the same basis before they are added.

The denominator also differs. Most cloud WUE figures divide by the IT load, the servers' electricity, while others divide by the whole site's electricity. SAEF applies each WUE to the energy it was published against, and the region pages show which one was used.

Because in most regions the grid's water factor per kWh is far larger than the data centre's own. For a class B model in EU West - Ireland, on-site cooling is 1% of the SAEF water total and the electricity is the rest. That is also why the water map follows the generation mix more than the climate, and why low carbon and low water do not go together. EU North - Sweden has the lowest carbon per token among the regions with a full water factor, yet its water per token is 0.43 litres per million tokens, against 0.11 in US South Central - Texas, the lowest.

The hydropower convention drives much of that. With reservoir evaporation allocated to electricity, US West - Oregon is the highest-water region in the dataset at 0.68 litres per million class B tokens. Leave hydro out and it drops to 0.046. SAEF publishes both totals side by side and prefers neither, because they are two conventions for one quantity, not an uncertainty range. A company that needs one volumetric number reports the total without hydro and discloses the allocated one beside it.

By weighting the volume with a local scarcity factor rather than adding anything. SAEF applies AWARE 2.0 to the on-site water only, because the data centre is the one location known. The power plants behind a grid are spread across many basins. The result can reverse a ranking. US South Central - Texas has one of the lowest water volumes, 0.11 litres per million class B tokens, but an AWARE factor of 13.56, so its scarcity-weighted on-site water is the highest in the dataset. EU West - Ireland sits at 1.17.

Scarcity-weighted litres are not litres. Report them as a separate indicator with the factor named, never added to a volume.

Asia Pacific - Singapore has no published water factor for electricity, so only on-site water is shown and no total is printed. EU South - Spain has no water factor in this version at all. A missing value reads "not available", never zero. Each value applies to any provider serving from that region, which is a proxy, and the confidence of each WUE source is in the full factor table.

A team sends 500 million billed tokens a year to a mid-size (class B) model. The arithmetic follows the SAEF method, and every factor comes from the dataset.

1. Facility electricity. 500 million tokens at 0.0713 kWh per million is 35.6 kWh, of which the servers draw 29.7 kWh.
2. On-site water, EU West - Ireland. Server electricity times a WUE of 0.02 L per kWh, published as withdrawal, times 0.67 to convert to consumption, is 0.40 litres.
3. Electricity water. Facility electricity times an EWIF of 1.476 L per kWh is 52.6 litres, or 40.3 litres with the 1.132 L per kWh factor that leaves hydro out.
4. Total. About 53.0 litres a year with hydro allocated, 40.7 litres without. Scarcity-weighted on-site water is 0.46 litres at an AWARE factor of 1.17.

Served from EU North - Sweden instead, the same tokens use 214.9 litres with hydro allocated and 49.8 without, while carbon falls from 12.3 to 3.6 kg CO₂e. The move cuts carbon and raises water on both conventions, about 4 times with hydro allocated, which is the trade-off a carbon-only figure hides. Neither total is large in absolute terms, and the arithmetic is the same at any volume, so the useful output is the comparison and the stated convention, not the litres alone.

Totals exist but are not per-query figures, and they mix boundaries. Google reports consuming 10.9 billion US gallons (41 billion litres) of water in 2025 across its data centres and offices, not just AI, in its environmental report published in 2026. Li and co-authors project global AI demand at 4.2 to 6.6 billion cubic metres of water withdrawal in 2027, which is withdrawal, not consumption, and a projection. Mistral reports 281,000 cubic metres of water consumed in the footprint of training Mistral Large 2, as of January 2025 after 18 months of use, without itemising the boundary. None of these can be divided down into a per-query value without assumptions the publisher did not make.

Water is not part of a GHG inventory, so report it beside the carbon line, never converted into it. A defensible line names the volume basis (consumption), the hydropower convention and both totals, the regions used, and the dataset version. If scarcity matters to your readers, add the AWARE-weighted on-site figure as a separate indicator. The per-service steps for collecting tokens are in the Scope 3 guide, and the same token export gives water with no extra data.

• WUE values are per region or per site, and some are continental aggregates or older site series, marked in the dataset.
• A region's value applies whichever provider serves from it. Providers' own sites can differ.
• Water used to manufacture chips and servers is outside the boundary, as in every published per-query figure above.
• AWARE weighting covers on-site water only, because the power plants behind a grid sit in many basins.
• The method paper is under peer review. Corrections from providers and researchers go into the next version.

The water equations are on the methodology page and every factor is open under CC BY 4.0. Anyone who can correct a WUE, a grid water factor or a source can write to guillermollopis@somaai.earth. Accepted corrections go into the next dataset version and its changelog.

Figures per query are one plain exchange, before the Tier 2b reasoning uplift. Electricity term. GPU energy per token from the ML.ENERGY Leaderboard v3, a facility overhead of 1.20, and grid carbon intensity from EPA eGRID 2023 and Ember 2023. Embodied hardware, data-centre construction and amortised training terms are documented on the methodology page. Water combines on-site cooling water (Li et al. 2025) with the water embedded in grid electricity (Reig et al. 2020). The modelled range of the models inside a class, around the class value, is −41% to +48% for class A, −39% to +24% for class B and −67% to +55% for class C.

Dataset: AI Inference Emission and Resource Factors for Corporate GHG Inventories (SOMA AI Emission Factors (SAEF)), Llopis-Garcia, G., SOMA AI, 19 October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447, earlier version 2 values at 10.5281/zenodo.22767475). Method: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, arXiv:2606.10660, under peer review.

Partners can fetch any row from the AI factors API, and the whole table is a CSV download.

• Why AI footprint estimates differ: published figures, dated. Energy, carbon and water figures with their boundaries
• How SOMA calculates the carbon and water footprint of AI usage. The equations behind every factor
• SOMA AI Emission Factors table and download. kg CO₂e, water and uncertainty by class and region