How much water does an AI query use, by cloud region?
For sustainability teams, journalists and anyone asked where an AI query's water goes. Every figure is a SAEF estimate for one exchange of 1,210 tokens (928 in, 282 out, the dataset's default short chat) on a class B (mid-size) model, the same exchange as the water guide and the carbon ranking by country, computed on this page from the open dataset. Factor version soma-ai-ef-2026.11-dataset-v3.1.0. Location-based, not provider readings. Spanish version: ¿Cuánta agua consume una consulta de IA según la región de nube?
Page updated
How much water does an AI query use in each cloud region?
About 0.146 mL to 0.944 mL per mid-size exchange across the 14 cloud regions with a complete water factor, a spread of about 6.5 times. The lowest is US South Central - Texas and the highest is US West - Oregon. These are SOMA AI estimates, not provider readings. The total adds two kinds of water. On-site cooling water runs from 0.001 mL in EU North - Finland to 0.067 mL in Asia Pacific - Japan. The water used to generate the electricity is the larger part everywhere, between 77.3% and 99.8% of the total, and it depends on the grid rather than on the data centre.
Which cloud regions use the least and the most water per query?
| Rank | Region | Total (mL) | On-site cooling (mL) | Electricity (mL) | Electricity share of total | Total without hydro (mL) |
|---|---|---|---|---|---|---|
| 1 | US South Central - Texas | 0.146 | 0.019 | 0.127 | 86.9% | 0.139 |
| 2 | EU West - Ireland | 0.147 | 0.001 | 0.146 | 99.3% | 0.113 |
| 3 | EU Central - Germany | 0.200 | 0.009 | 0.190 | 95.3% | 0.175 |
| 4 | US South - Southeast | 0.236 | 0.008 | 0.228 | 96.6% | 0.180 |
| 5 | US East - N. Virginia | 0.239 | 0.003 | 0.235 | 98.6% | 0.199 |
| 6 | EU West - UK (London) | 0.239 | 0.008 | 0.232 | 96.8% | 0.190 |
| 7 | EU West - Belgium | 0.260 | 0.003 | 0.258 | 98.9% | 0.219 |
| 8 | Asia Pacific - Japan | 0.295 | 0.067 | 0.228 | 77.3% | 0.214 |
| 9 | US Central - Iowa | 0.314 | 0.008 | 0.306 | 97.4% | 0.174 |
| 10 | Netherlands | 0.342 | 0.002 | 0.340 | 99.3% | 0.342 |
| 11 | EU West - France | 0.365 | 0.003 | 0.362 | 99.2% | 0.206 |
| 12 | EU North - Finland | 0.449 | 0.001 | 0.448 | 99.8% | 0.170 |
| 13 | EU North - Sweden | 0.595 | 0.001 | 0.594 | 99.8% | 0.138 |
| 14 | US West - Oregon | 0.944 | 0.007 | 0.937 | 99.3% | 0.064 |
Other model classes scale these figures with their energy per token. The table is for a class B model only.
What drives the spread between regions?
- The grid's water per kWh. Generating electricity consumes water, mostly in thermal plants and hydropower reservoirs. The factor runs from 1.29 litres per kWh in US South Central - Texas to 9.50 in US West - Oregon. It is the main reason the regions differ.
- The hydropower convention. Counting reservoir evaporation as electricity water lifts hydro-heavy grids a lot. US West - Oregon falls from 0.944 mL to 0.064 mL when hydro is left out, from 9.50 to 0.58 litres per kWh. Without hydro, the lowest region is US West - Oregon and the highest is Netherlands. The two totals are two conventions for one quantity, not an uncertainty range.
- The data centre's own cooling. Water usage effectiveness (WUE) differs by site, from 0.01 to 1.22 litres per kWh among these regions as published, before conversion to consumption. It is still the smaller term in every region listed.
The consequence is that a low-carbon region is not automatically a low-water one. The water guide shows the scarcity weighting, which can reverse a ranking because a litre taken in a water-stressed basin counts for more.
Which regions have no published water factor?
Asia Pacific - Singapore has a cooling-water factor but no grid water factor for the electricity in this dataset version, so the page prints no total. The on-site part alone would be 0.086 mL for Asia Pacific - Singapore, which leaves out the larger share and is not comparable with the totals above. EU South - Spain has no water factor at all in this version, neither for cooling nor for the electricity, because SAEF has no source it can cite for that region. A missing value is not zero and is never filled by copying a neighbouring region. Corrections and sources are welcome and go into the next dataset version.
How are these estimates made, and what are their limits?
- Each figure is an estimate. It multiplies the facility energy of a class B model by the region's water factors, with on-site water converted to consumption on one documented basis. Cooling factors come from provider site disclosures (AWS, Microsoft, Meta) and Li et al. 2025, grid water factors from Reig et al. 2020. The equations are on the methodology page.
- A region's factor applies to any provider serving from it. A provider's own sites can differ, and the dataset marks each source's confidence in the factor table.
- The exchange is the dataset's default short chat. Longer prompts, reasoning modes and agents multiply the figures.
- Water used to make chips and servers is outside the boundary, as in every published per-query figure.
- Grid factors are averages, not renewable contracts or the mix of one data centre.
- The method paper was accepted for publication in Sustainable Production and Consumption in November 2026.
Keep reading
- How much water does an AI query use? Which water each published figure counts, scarcity and a worked example
- Where does an AI query emit the least? The carbon footprint of AI by country. The same exchange ranked by carbon on 210 national grids
- Does AI really drink a bottle of water per email, and what does the evidence say? Where the bottle-per-email claim came from and why it is far lower now
- Why should AI providers say where your prompt runs? The policy case for disclosing the serving region
- How do I calculate the carbon footprint of my company's AI usage? Seven steps from AI tools to a Scope 3 line, with water from the same token export