How much water does AI use?

How much water does AI use per prompt?

Four published figures are often quoted side by side, but they measure different things.

Google's figure is the lowest because it counts only water evaporated on site, for a median prompt, on recent hardware. The UC Riverside figure adds water from electricity generation, which makes up most of its total (14.7 of the 16.9 mL). Mistral's figure is the highest because it includes manufacturing. None of these is wrong. They answer different questions.

Why AI water usage estimates vary so widely

Direct water vs indirect water

Direct water is what a data centre uses on site, mostly to remove heat from servers. Indirect water is what power plants use to generate the electricity the data centre draws. Thermoelectric plants draw large volumes of water to carry away heat, and hydroelectric reservoirs lose water to evaporation.

Indirect water is usually the bigger share. The Lawrence Berkeley National Laboratory (LBNL) estimates that US data centres consumed 66 billion litres of water directly in 2023, and nearly 800 billion litres indirectly through their electricity use. That works out at 4.52 litres of indirect water for every kWh. The International Energy Agency reaches the same conclusion: the largest share of water use in the data centre value chain happens upstream, in power generation.

So any per-prompt figure that leaves out electricity generation, including Google's, covers the smaller part of the total water footprint. Our glossary entry on data centre water use explains both parts.

Water withdrawal vs water consumption

Water withdrawal is all the water taken from rivers, lakes or groundwater. Water consumption is withdrawal minus what is returned, so it is the water that evaporates or is otherwise lost to the local supply. In 2025, Google's data centres withdrew about 51 billion litres and consumed about 40 billion.

This matters when comparing providers. Water usage effectiveness (WUE) is the water used per kWh of electricity drawn by IT equipment, but providers calculate it differently:

• AWS reports a 2025 global WUE of 0.12 L/kWh, based on water withdrawn.
• Google uses 1.15 L/kWh for 2024, based on water consumed.
• Microsoft reports an average WUE of 0.27 L/kWh for its 2025 financial year, based on water used for humidification and cooling. Microsoft does not say whether that is water withdrawn or water consumed.

The figures look comparable but are not, in the same way power usage effectiveness (PUE) figures are published on different bases.

Model, hardware and time

Water estimates are built on electricity estimates, so anything that changes the power drawn per prompt changes water too: model size, prompt and answer length, hardware, batching and data centre overhead. Our guide to AI energy consumption covers those drivers. Figures also date quickly. Google reports that the electricity used by its median Gemini text prompt fell 33 times in the 12 months to May 2025. The UC Riverside estimate is based on GPT-3, a 2020 model.

Where does the water go in a data centre?

Most data centre water goes into cooling. In a cooling tower, part of the water evaporates to carry heat away, and the rest is discharged as wastewater. The UC Riverside paper estimates that around 80% of the water withdrawn for these towers is consumed, when water quality is good.

Evaporative systems also produce wastewater. As water evaporates, minerals concentrate in what is left, so operators drain part of it off (blowdown) and top up with fresh water. US federal guidance notes that typical water treatment programmes add corrosion, scaling and biological fouling inhibitors. A site therefore has two effects to manage: water loss through evaporation, and a chemically treated discharge that needs managing so it does not add to water pollution downstream.

The water usually has to be clean freshwater, often the same treated drinking water people use at home, to stop pipes clogging and bacteria growing. Google's 2025 data shows 13% of its freshwater withdrawals came from sources at high risk of depletion or scarcity, and AWS reports 22% of its 2025 withdrawals came from regions with high or extremely high water stress.

Is AI water recycled?

Partly. Water in a tower can only be cycled about 3 to 10 times before minerals build up and it has to be discharged, so fresh water keeps being added. Three other approaches cut freshwater use:

• Closed-loop cooling. Microsoft's data centre design for AI, introduced in 2024, uses a closed-loop, direct-to-chip system that consumes no water for heat removal during operations.
• Reclaimed and non-potable water. Microsoft's sites in Quincy (Washington), Singapore and San Antonio use recycled, reused or non-potable water for 74%, 99% and 79% of their needs. AWS has commitments to supply reclaimed water for heat removal at 130 data centre sites across six countries.
• Using air instead of water. Google says it chooses air cooling or recycled water where a local water source is at high risk.

Replenishment is different from recycling. When Google says it replenished 7.7 billion gallons in 2025, about 78% of its freshwater consumption, that water was returned through projects in local watersheds, not reused inside the data centre. Using air has a trade-off too: it increases electricity usage, which can raise indirect water use. The IEA flags this trade-off between saving water and saving electricity.

Location: why the same AI prompt uses different amounts of water

Outside temperature decides how often a data centre needs water to remove heat. Microsoft's outside-air system uses water only when it is hotter than 29.4°C outside. That means no water in parts of Northern Europe, less than 5% of the time in Dublin and Amsterdam, about 10% of the year in Virginia and up to 40% in Phoenix.

Google's site-level data for 2025 shows the same pattern. Its Council Bluffs campus in Iowa consumed 1,346 million gallons of water, while Hamina in Finland consumed 0.4 million and Dublin 0.1 million. Totals reflect the size of each site as well as local weather, but the cooler Nordic and Irish sites use almost none.

The grid matters too, because indirect water depends on how local electricity is generated. And the same litre has a bigger impact where water is scarce. The IEA finds that nearly 40% of identified capacity for AI data centres is in areas with high or extremely high baseline water stress.

Environmental impact: AI, water scarcity and local water systems

A data centre draws its water from one local water system, so the impact of AI water use depends on water availability there. During a drought, or where water scarcity is already a problem, it competes with people, farms and the rivers and ecosystems that rely on the same freshwater supply, often for potable water from the public water supply.

In England, the Environment Agency warns that the public water supply could be short by 5 billion litres a day by 2055, and names data centres among the emerging industries adding to water demand. With the UK government committed to a 10-year investment in AI infrastructure, its National Framework for Water Resources expects a large and rapid increase in data centre numbers and notes that some water companies have already refused them water. It also says climate change makes hot, dry summers more likely and drought quicker to set in, a risk for nature as well as people.

A survey of 73 commercial data centres by techUK with the agency found 64% used under 10,000 cubic metres a year and 51% used waterless systems, though the sample was voluntary and left out enterprise sites.

On water management, Google says it only considers water cooling where local resources are healthy and resilient.

Training vs inference water

Training a model uses a large amount of water once. Inference uses a small amount every time people send a prompt to a generative AI tool, which adds up as demand grows.

• The UC Riverside paper estimates that training GPT-3 in Microsoft's US data centres consumed about 700,000 litres of water on site, and about 5.4 million litres including electricity generation.
• Mistral AI reports that training Mistral Large 2 and 18 months of use consumed 281,000 cubic metres of water, as of January 2025.

For an organisation, inference is the part its people control, through the models they pick, the tokens they send and the regions they run in.

AI water use statistics

• Google consumed 10.9 billion gallons (41 billion litres) of water across its data centres and offices in 2025, up 34% on 2024 (Google, 2026).
• AWS data centres withdrew 9.4 billion litres of water in 2025 (AWS, 2026).
• US data centres consumed 66 billion litres (about 17 billion gallons) of water directly in 2023, up from 21.2 billion litres in 2014. Hyperscale data centres alone could consume 60 to 124 billion litres in 2028 (LBNL, 2024).
• Global AI demand could account for 4.2 to 6.6 billion cubic metres of water withdrawal in 2027 (Li et al.).

Company and national totals cover all data centre workloads, not only AI. Google links its 37% rise in freshwater use in 2025 to expanding its data centre fleet to meet demand for AI. The electricity side of that growth is covered in our article on data centre energy consumption.

How AI water use compares with everyday household water

In England, people use an average of 136.5 litres of water a day. Using the published figures for AI prompts, that one day of household water equals roughly:

• 525,000 median Gemini prompts, at Google's 0.26 mL
• 8,000 GPT-3 responses, at UC Riverside's US average of 16.9 mL
• 3,000 Mistral Large 2 responses, at Mistral's 45 mL

A standard bottle of water, the 500 mL measure UC Riverside uses, works out at roughly 1,900 Gemini prompts, 30 GPT-3 responses or 11 Mistral Large 2 responses on the same figures.

Google offers another comparison: the water it consumed in 2025 is about what it takes to irrigate 73 golf courses a year in the south-western United States.

These comparisons show that each prompt people send uses very little water. They miss the local picture: a data centre draws on one watershed, often in a dry region, and that is where the impact is felt.

The future of AI water use: innovation in cooling technology

The total amount of water AI uses will likely keep rising as demand grows and AI operations scale. LBNL projects that US hyperscale sites alone could consume 60 to 124 billion litres (about 16 to 33 billion gallons) in 2028. Most of the innovation that could slow this is in heat removal technology:

• Liquid cooling. Direct-to-chip liquid systems are gaining popularity, and closed-loop systems are more water efficient than open ones.
• Zero-water designs. Microsoft says its closed-loop design for AI will further reduce its water use intensity, towards a 40% improvement by 2030.
• Replenishment. Google aims to replenish more water than it consumes at its sites by 2030, more than 19 billion gallons (about 72 billion litres) a year.

Not all innovation is new technology. England's water regulator is collecting water data from the sector and working on data centre sustainability policy. For people using AI, insights into the water footprint of each model and region matter more than averages.

What organisations can actually measure

No organisation can meter the water its AI provider uses. What it can do is work out an estimate from usage data, in a consistent way:

• Collect usage for each AI provider, tool and agent: tokens by model, token type, region and time.
• Convert tokens to electricity with per-model kWh rates, including data centre overhead.
• Apply the provider's published WUE to the IT electricity to get on-site water.
• Apply a regional water intensity factor for electricity generation to get indirect water.
• Report the two separately, with a water stress rating for each region, so the people choosing models and regions can see where water is scarce.

Ask your AI and infrastructure providers which basis their WUE uses, and whether their per-prompt figures include electricity generation. Water belongs in a wider sustainable AI approach, next to electricity, carbon and cost.

How Greenpixie estimates AI water usage

GPX AI estimates the water, electricity and carbon of AI usage across over 500 models, from any usage report that shows tokens by model and token type. Water is reported in litres, in two parts:

• Cooling water: the provider's published WUE multiplied by the IT electricity used for inference.
• Electricity generation water: regional water intensity factors, adapted from World Resources Institute guidance, applied to total facility electricity.
• Water stress: a category for each region, from the World Resources Institute's Aqueduct 4.0.

The water approach was developed with Dr Shaolei Ren at the University of California, Riverside, and informed by Making AI Less Thirsty by Li et al. The same factors feed GPX Cloud, which adds water to every line of an AWS, Azure or GCP bill.

Using Greenpixie's AI methodology (version 6.4, November 2026), a chat-style request to GPT-6.1 Sol with 1,000 input and 500 output tokens uses about 7.8 mL of water. That covers both data centre cooling and the water used to generate its electricity. Almost all of it, about 7.1 mL, comes from the 500 output tokens. That puts it inside the published range above, well above on-site-only figures such as Google's 0.26 mL. A 500 mL bottle covers about 64 of these requests.

Carbon, electricity and water factors come from Greenpixie's cloud methodology. The methodology has been independently verified in accordance with ISO 14064-3, prepared with reference to ISO 14064-1, and designed to be in line with the GHG Protocol Corporate Standard. The verification covers greenhouse gas emissions. The water approach is informed by Dr Shaolei Ren at the University of California, Riverside. The AI per-token energy layer builds on these factors and is validated against independent research and metered electricity data.

Figures are estimates, given as a median with an interquartile range, so the uncertainty is visible. Try GPX AI in the app to compare models for your own usage, or book a call to work through your numbers.

Frequently asked questions

How much water does ChatGPT use?

OpenAI CEO Sam Altman has said an average ChatGPT query uses about 0.000085 gallons of water, roughly a fifteenth of a teaspoon or about 0.32 mL. OpenAI has not published how that figure was worked out, or whether it includes water used to generate electricity. Independent estimates for older OpenAI models, such as UC Riverside's for GPT-3, are higher, partly because they include electricity generation water.

What is data center water usage per query?

It depends on what is counted. On-site water alone is well under a millilitre for a median text prompt on Google's figures. Adding the water used to generate the electricity raises it to around 17 mL at the US average in UC Riverside's estimate for GPT-3. The location of the data centre and the size of the model also change the result.

Why does AI use so much water?

AI systems run on GPUs and other accelerators that draw a lot of power and produce a lot of heat. Many data centre operators remove that heat by evaporating water, which uses little electricity. Generating the electricity also uses water, which is usually the larger share.

Does AI use drinking water?

Often, yes. Evaporative systems need clean freshwater to avoid clogging and bacteria, so many data centre operators draw potable water from the local supply. Some sites now use reclaimed wastewater or other non-potable sources instead.

Is air cooling better than using water?

Not always. It uses little or no water on site but needs more electricity, which can raise carbon emissions and the indirect water used in generating it. The better choice depends on local water stress and how clean the grid is.

How much water do AI data centers use per day?

Published totals cover all workloads, not only AI. Google's data centres and offices consumed 41 billion litres in 2025, about 112 million litres of water per day (30 million gallons of water). Most sites are far smaller: in a techUK survey, 64% of commercial sites in England used under 10,000 cubic metres of water per year, or under 27,500 litres a day.

Sources

1. Li, Yang, Islam and Ren, Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, arXiv v5 (26 March 2025); published in Communications of the ACM 68(7), 2025
2. Google Cloud, Measuring the environmental impact of AI inference (21 August 2025)
3. Elsworth et al. (Google), Measuring the environmental impact of delivering AI at Google scale, arXiv (21 August 2025)
4. Google, 2026 Environmental Report, covering 2025 (June 2026)
5. Google, Water stewardship commitments for local communities (3 June 2026)
6. Sam Altman, The Gentle Singularity (June 2025)
7. Mistral AI, Our contribution to a global environmental standard for AI (22 July 2025)
8. Amazon Web Services, 2025 AWS Summary (July 2026)
9. Microsoft, Inside Microsoft's two-decade push to cut water intensity while scaling for growth (24 June 2026)
10. Shehabi et al., Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (December 2024)
11. IEA, Powering Artificial Intelligence, executive summary (14 October 2026)
12. Environment Agency, Water resources 2024 to 2025 (18 November 2025)
13. Greenpixie AI methodology: 7. Assessment and Calculation
14. Greenpixie Cloud methodology: 7. Assessment and Calculation
15. Environment Agency, England faces 5 billion litre public water shortage by 2055 without urgent action (17 June 2025)
16. Environment Agency, National Framework for Water Resources 2025, chapter 2: Current and future pressures on water resources (2025)
17. Environment Agency, National Framework for Water Resources 2025, chapter 9: Taking action on other significant water using sectors and emerging demands (updated 27 August 2026)
18. techUK, New report sheds light on commercial data centres water usage in England (18 August 2025)
19. Microsoft, Measuring energy and water efficiency for Microsoft datacenters (undated, FY25 data)
20. US Department of Energy, Best Management Practice #10: Cooling Tower Management (undated)