greenai.guide

Home › Articles › How much water does AI use?

Article

How much water does AI use?

Updated · 7 min read

In short

AI uses water in two main ways: directly, to cool data centres, and indirectly, at the power stations that generate their electricity. Per text prompt, published figures range from 0.26 mL (Google's figure for Gemini) to about 45 mL (Mistral AI's life-cycle figure), mainly because they count different things. In total, AI water use is large and rising: US data centres consumed about 66 billion litres directly in 2023, plus around 800 billion litres indirectly through their electricity, and where that water is taken from matters as much as how much.

Key points

  • Google reports 0.26 mL of water for a median Gemini text prompt; Mistral AI's life-cycle assessment puts a 400-token response at 45 mL.
  • Li and colleagues estimated that training GPT-3 in Microsoft's US data centres could consume 5.4 million litres of water, including 700,000 litres on site.
  • Lawrence Berkeley National Laboratory estimated US data centres consumed 66 billion litres of water directly in 2023, and nearly 800 billion litres indirectly through electricity generation.
  • Li and colleagues project global AI demand could account for 4.2–6.6 billion cubic metres of water withdrawal in 2027, of which 0.38–0.60 billion would be consumed.
  • Microsoft's water consumption roughly doubled from 3,990 to 8,170 megalitres between its 2020 and 2025 financial years; 48% of it in 2025 was in water-stressed areas.
  • AWS reports a global water usage effectiveness of 0.12 litres per kWh for 2025, while the US electricity it draws on consumes around 3–4.5 litres per kWh at the power station.

How much water does AI use? Less per prompt than many headlines suggest, but more in total, and more in some places, than most people realise. AI water use comes mostly from cooling data centres and from generating the electricity they run on. This page sets out the published figures for AI water consumption, explains why they range so widely, and says what is still unknown.

How much water does one AI prompt use?

Between a fraction of a millilitre and a few tens of millilitres per text prompt, depending on what is counted. The published figures are not comparable with one another, so read each with its boundary.

Figure What it covers Source
0.26 mL per median Gemini text prompt On-site data-centre water, May 2025 Google, 2025
45 mL per 400-token Le Chat response Life-cycle assessment, including upstream impacts such as manufacturing servers Mistral AI, 2025
500 mL per roughly 10–50 medium-length GPT-3 responses On-site plus power-station water, varying with where and when it runs Li, Yang, Islam and Ren, 2025

Google describes its figure as about five drops of water; it is calculated from the data centres' on-site water efficiency and does not include water used to generate the electricity. Mistral's is larger mainly because it is a life-cycle assessment, done with the consultancy Carbone 4 and the French environment agency ADEME, that counts upstream impacts such as manufacturing servers, not just on-site cooling. Li and colleagues' estimate is for an older model, but it showed something that still holds: the same AI answer can use several times more water depending on the data centre and the time of day.

How does AI use water?

Directly, by evaporating water to cool data centres, and indirectly, through the water used to generate electricity and make chips. Li and colleagues set out three scopes, mirroring carbon accounting.

  1. On-site (scope-1) water. Almost all the electricity a server uses becomes heat. Many data centres reject that heat through cooling towers or evaporation-assisted air cooling, which evaporate water. Liquid cooling inside the server, such as direct-to-chip cooling, does not itself consume water; what matters is how the facility finally releases the heat outdoors.
  2. Off-site (scope-2) water. Thermal power stations, which burn fuel or use nuclear heat, need water for cooling. Li and colleagues put the US average at about 43.8 litres withdrawn and 3.1 litres consumed per kilowatt-hour generated. Lawrence Berkeley National Laboratory estimated 4.52 litres consumed per kWh for US data centres, given the grids where they are located.
  3. Embodied (scope-3) water. Chip and server manufacturing needs large amounts of ultrapure water, and some of the discharged water may carry chemical waste.

What is the difference between withdrawal and consumption?

Withdrawal is water taken from a river, aquifer or supply; consumption is the part not returned, mostly through evaporation. Consumption is the better measure of water lost to a local area; withdrawal matters where supply is tight. Company reports and studies use both, so check which one a figure is before comparing.

How much water do data centres use in total?

Billions of litres a year on site, and many times more through their electricity. National and global figures are estimates; data centres do not all report water use.

  • United States: Lawrence Berkeley National Laboratory estimated that data centres consumed about 66 billion litres of water directly in 2023, up from 21.2 billion in 2014, with hyperscale and colocation sites accounting for 84%. It put their indirect water footprint through electricity at nearly 800 billion litres, about twelve times the direct figure, and expects hyperscale data centres alone to consume 60–124 billion litres directly in 2028.
  • Global AI: Li and colleagues projected that global AI demand could account for 4.2–6.6 billion cubic metres of water withdrawal in 2027, more than the annual withdrawal of half the United Kingdom, of which 0.38–0.60 billion cubic metres would be consumed.

The large difference between direct and indirect water is the main reason estimates diverge. A study that counts only cooling water will give a far lower figure than one that includes the power stations.

How much water does training an AI model use?

Hundreds of thousands to millions of litres for a large model, on the few published estimates.

  • GPT-3: Li and colleagues estimated that training it in Microsoft's US data centres could consume 5.4 million litres of water in total, including 700,000 litres evaporated on site.
  • Mistral Large 2: Mistral AI's life-cycle assessment put training plus 18 months of use at 281,000 cubic metres of water consumed.

Few model cards report water at all. Li and colleagues noted that while model cards routinely include carbon emissions, even on-site water use is usually missing.

What do AI companies report about water?

Their total water use is rising, and some now report efficiency per kilowatt-hour. Figures come from annual sustainability reports and are self-reported, though some are externally assured.

Company Figure Year
Microsoft Water consumption 8,170 megalitres, up from 3,990 in FY2020; withdrawal 13,266 megalitres Financial year 2025
Alphabet (Google) Water consumption 10,869 million gallons (about 41 billion litres), company-wide 2025
Google Replenished about 7.7 billion gallons, roughly 78% of its freshwater consumption 2025
AWS Global WUE of 0.12 litres per kWh of IT load, 20% better than 2024 2025

WUE (water usage effectiveness), defined by The Green Grid, is litres of water used on site per kilowatt-hour of IT energy. It is a useful efficiency measure, but it leaves out power-station water and says nothing about whether the local area is short of water. "Replenishment" means funding projects that return water to watersheds; it does not reduce what a data centre takes. See making sense of PUE and WUE.

Why does location matter so much?

Because a litre taken in a dry region costs far more, environmentally and socially, than one taken where water is plentiful. Climate and cooling design change water use per kilowatt-hour dramatically.

Microsoft's site-by-site figures show the spread. In its 2025 financial year, its data centres around Phoenix, Arizona, used about 954,000 MWh of electricity and withdrew 981 megalitres of water; those around Dublin, Ireland, used more electricity, about 1.3 million MWh, and withdrew 18 megalitres. Across the company, 48% of its 2025 water consumption was in areas of high or extremely high water stress, assessed with the World Resources Institute's Aqueduct tool.

Timing matters too, and it can pull against carbon. Li and colleagues found that water efficiency varies by hour as well as by place: on-site water use rises in the heat of the day, while the grid is often cleanest when solar power is plentiful. Cutting carbon may mean "following the sun" and cutting water "unfollowing" it, so the two footprints have to be weighed together rather than traded off blindly. If a data centre is planned near you, see have a say on a data centre near me.

What is uncertain about AI water use?

Most of it, in detail: few providers publish per-task water figures, methods differ, and local impacts are rarely reported. The main gaps:

  • Boundaries. On-site only, or power-station and manufacturing water too? This alone explains most of the spread between 0.26 mL and 45 mL per prompt.
  • Withdrawal or consumption. Some figures report one, some the other.
  • Location and timing. Global averages hide local stress, and few companies publish site-level data.
  • AI's share. Company totals cover all data centres and often offices too; AI is not reported separately.
  • Projections. Forecasts for 2027–2030 depend on assumptions about AI growth, cooling technology and the grid.

Reporting is improving in places. In the EU, operators of data centres with at least 500 kW of IT power demand must report energy and water use under Delegated Regulation 2024/1364. Elsewhere, disclosure is mostly voluntary.

How can AI use less water?

Use less computation, run it in the right places, and choose the right cooling and electricity. The same steps that cut energy also cut water.

What to do next

Frequently asked questions

How much water does ChatGPT use?

OpenAI has not published a water figure. Li and colleagues estimated that GPT-3 needed to consume a 500 mL bottle of water for roughly 10 to 50 medium-length responses, depending on where and when it ran. That older estimate includes power-station water; newer company figures for other models are lower.

How much water does one AI prompt use?

Published figures range from 0.26 mL for a median Gemini text prompt (Google, 2025) to 45 mL for a 400-token Mistral response counted over its full life cycle. The gap is mostly about what is counted: on-site cooling only, or a full life cycle including upstream impacts such as manufacturing servers.

Why do data centres use water?

Mainly to remove heat. Many data centres use cooling towers or evaporation-assisted cooling, which evaporate water. Power stations that generate their electricity also use water, and making chips needs ultrapure water.

What is the difference between water withdrawal and consumption?

Withdrawal is all the water taken from a source, much of which is returned. Consumption is water that is not returned, mostly because it evaporates. Consumption is the better measure of water lost to a local area.

What is WUE in a data centre?

Water usage effectiveness: litres of water used on site per kilowatt-hour of IT energy. Lower is better. AWS reports 0.12 L/kWh for 2025. WUE leaves out water used to generate the electricity, which is often much larger.

Is AI's water use a problem?

It depends on where it happens. Li and colleagues' projection for global AI water withdrawal in 2027 is more than the annual withdrawal of half the United Kingdom, and 48% of Microsoft's 2025 water consumption was in water-stressed areas. A litre taken from a stressed region matters far more than one from a water-rich one.

How can AI use less water?

Use less computation, run work in cooler or water-rich regions and at water-efficient times, and use cooling systems that do not evaporate water. For individuals, the same habits that save energy save water, because water use follows energy use.

Sources

  1. Making AI less 'thirsty': uncovering and addressing the secret water footprint of AI models, Li, Yang, Islam and Ren (arXiv 2304.03271, v5) (2025)
  2. Measuring the environmental impact of delivering AI at Google scale, Elsworth et al., Google (2025)
  3. Our contribution to a global environmental standard for AI, Mistral AI (2025)
  4. 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory (2024)
  5. 2026 environmental data fact sheet, Microsoft (2026)
  6. Alphabet FY2025 environmental indicators assurance letter (Schedule of select environmental metrics), Alphabet (2026)
  7. Read our 11th annual Environmental Report, Google (2026)
  8. 2025 AWS sustainability summary, Amazon Web Services (2026)
  9. Water Usage Effectiveness (WUE): a Green Grid data center sustainability metric (White Paper 35), The Green Grid (2011)
  10. Commission Delegated Regulation (EU) 2024/1364 on the first phase of the establishment of a common Union rating scheme for data centres, European Commission (2024)

Join us

Help cut AI's impact on the planet

Get new guides, funding calls and directory listings by email, about once a month. Or add your product, course, PhD or research to the directories.

No tracking. Unsubscribe in one click. Other ways to join · Privacy