How Much Water Does AI Use? Why the Numbers Disagree

Quick answer

Google measured 0.26 mL of water per median Gemini text prompt, counting cooling in its data centers. University researchers estimated 10-50 GPT-3 responses per 500ml bottle, counting cooling and electricity. Scope, model, year and method differ, so the figures are not comparable and there is no single number to quote.

Key takeaways

• AI uses water in two places: on-site at the data center (cooling) and off-site at the power plants that supply its electricity.
• Published figures measure different things (per answer, per training run, per data center or per country), so treat each as a number for its own scope, never as a rival to another.
• Google, a vendor, published a measurement of its own system; OpenAI's CEO published one figure without a stated method; university researchers published an estimate for an older model.
• None of these figures lets you compute your own business's water use; what you can act on is what your vendor discloses and stating the scope whenever you repeat a number.

How does artificial intelligence use water?

Google measured 0.26 mL of water for a median Gemini text prompt, counting cooling in its data centers; university researchers estimated that GPT-3 uses a 500ml bottle for 10-50 responses, counting cooling plus electricity. The figures measure different things and are not comparable.

AI uses water in two places: on-site, to cool the computers in a data center, and off-site, at the power plants that make the electricity those computers draw. The Li paper described below defines the first as scope 1 and the second as scope 2. It names a third place, making the chips and servers (scope 3), and says that "Although largely obscure, scope-3 water usage is likely significant". This page leaves scope 3 out.

On-site, Li and colleagues say server heat is moved out with air or liquid cooling methods that "do not evaporate or consume water", then rejected from the building. For that second step they name "water-intensive cooling towers and water evaporation-assisted air cooling" as two common approaches in many data centers. Off-site, they write that "Different power plants use different amounts of water for each kWh generation, depending on the cooling techniques", so the water behind a kilowatt-hour depends on how the local grid generates power.

Why do the published numbers disagree?

They measure different things, so they should not be set side by side. Five differences matter when you read a figure.

• Scope. On-site cooling only, or cooling plus the water behind the electricity.
• Unit of account. Per answer, per training run, per data center, or per country per year. For any total you meet, ask which water it counts (direct or indirect), which year, which facilities, whether it is withdrawal or consumption, and whether it is a measurement or a model.
• Withdrawn or consumed. The Li paper defines consumption as "water withdrawal minus water discharge"; withdrawal is freshwater taken from ground or surface sources.
• Who measured. A company measuring its own system, a company executive stating a number, or outside researchers estimating from public data.
• Year, model and place. A figure for one model, region and year does not carry over to another.

The table lists each figure with its own scope. Read across a row, not down a column.

How much water does one AI answer use?

It depends on whose figure you pick and what it counts, and the sources do not agree. Here are the three, each in its own terms.

Google reports that a median Gemini Apps text prompt consumed 0.26 mL of water, using its comprehensive measurement method on prompts from May 2025. This is Google's measurement of its own serving system, and the paper equates 0.26 mL to five drops of water.

For "a methodology more similar to a narrow, existing approach", the paper's prose gives 0.15 mL of water consumption, while the water row of its Table 2 (mL/prompt) lists 0.12 under the Existing Approach column and 0.26 under the Comprehensive Approach column. The paper's two values for the narrower approach differ. The paper computes water as energy per prompt, less overhead energy, times Google's Water Usage Effectiveness, in the consumptive variant (ISO WUE Category 2).

OpenAI's CEO wrote that an average ChatGPT query uses about 0.000085 gallons of water. The post gives that number with a comparison to a fraction of a teaspoon and states no method or measurement boundary.

The university estimate is of a different kind. Li and colleagues estimated that GPT-3 consumes a 500ml bottle of water for roughly 10-50 medium-length responses, depending on when and where it is deployed. Their count includes off-site electricity water, and it is for an older model. The authors also say their inference estimate is conservative and that the actual consumption could be several times higher. So two published sources, one a vendor's own measurement and one an outside estimate, give answers that cannot be reconciled without knowing the model, the year, the site and the boundary.

How much water does training a model use?

One published estimate: training GPT-3 in Microsoft's US data centers could consume 5.4 million liters of water in total, of which 700,000 liters was on-site cooling. That is an outside estimate for a single older model's training run, from the same Li paper. It says nothing about newer models or about the cost of running a model for users afterward.

Do not add this to the per-answer figures above. A training-run number and a per-answer number answer different questions.

Is AI a real threat to water supplies?

It depends on where the data center sits, so a national total cannot answer it for your town. Water stress is local, and the Li paper reports that Water Usage Effectiveness "varies both spatially and temporally", so the same computing can carry a different water cost in a different place or at a different time.

Companies know this. Google says all new data centers it develops in high-stress locations will use air-cooled technology during normal operations. That is a vendor commitment about new sites, not a measurement of existing ones, and it is not evidence about any other company.

This page gives no verdict on whether the total is large or small. The sources here do not state one you can lift, and no figure above is a forecast.

Does it matter for my business, and what can I do?

None of the figures on this page lets you compute your own business's water use. They cover one prompt on one system, one older model, or one training run, and no source here gives a total you could take a share of. What an owner can act on is what a vendor discloses, and the scope note when you repeat a number.

1. Ask a vendor for its own environmental report. Look for water, not only carbon, and for whether the figure is consumption or withdrawal.
2. Check the scope. Is the figure on-site cooling, or does it include electricity? Is it per prompt, per year or per data center? Skip any number that does not say.
3. Check the year and the model. A figure for an older model or an earlier year does not carry over.
4. Prefer a stated method to a bare number. A figure that says how it was measured can be checked. One that does not cannot.
5. Match the tool to the job. See the practical guide to AI automation for small business and AI agents for small business for what each job needs.

This article was drafted with Claude and checked against each publisher's own material. Next step: before you quote any AI water figure to a customer, a board or a post, write down its scope in one line using the table headings.

FAQ

How much water does ChatGPT use per question?

OpenAI's CEO wrote in June 2025 that an average ChatGPT query uses about 0.000085 gallons of water. The post does not say what that covers or how it was measured, so it cannot be set against figures with a stated scope. Researchers estimate other, older models differently.

Does AI use more water than other things?

This page does not rank it. The sources here measure different scopes, and none of them states a comparison you can lift.

What is the difference between water withdrawn and water consumed?

The Li paper defines water consumption as water withdrawal minus water discharge, meaning water evaporated, transpired, incorporated into products or crops, or otherwise removed from the immediate water environment. Withdrawal is freshwater taken from ground or surface sources. Check which one a figure means.

Do all data centers use water for cooling?

Not in the same way. The Li paper says server-level air or liquid cooling methods do not evaporate or consume water, and that water-intensive cooling towers and water evaporation-assisted air cooling are two common facility-level approaches in many data centers. Google says new data centers in high-stress locations will use air-cooled technology during normal operations.