How much water does SI (AI) use? Ask the question online and you will meet two very different answers. One says a chatbot "drinks" a bottle of water for every short conversation. Another says a single prompt uses about five drops. Both figures come from serious sources, and both are real. They just measure different things, at different times, with different assumptions. This guide lays the main numbers side by side, explains why they disagree, and shows what they can and cannot tell you.
Short answer. Per prompt, the water directly used to cool the servers is tiny: Google measured 0.26 mL for a median Gemini text prompt in 2025. Older estimates that also count the water used to produce the electricity, and that assumed much more energy per answer, landed between about 10 and 50 mL per response. The bigger story is the total: data centers are growing fast, and their water use grows with them.
Where the water goes
A data center turns electricity into computation, and almost all of that electricity ends up as heat. Removing the heat is where water comes in. Many large sites use evaporative cooling towers: warm water gives up its heat by evaporating, and fresh water has to be added to replace what leaves as vapour. That is the water people usually picture.
There is a second, less visible share. Power plants also use water, mostly for their own cooling. When researchers count the water behind every kilowatt-hour a data center buys, the total rises. The paper that started much of the public debate, "Making AI Less Thirsty" by Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren (UC Riverside and UT Arlington), uses the same vocabulary as carbon accounting:
- Scope-1 water: water used on site, mainly for cooling the servers.
- Scope-2 water: water used off site to generate the electricity the data center consumes.
- Scope-3 water: water embedded in the supply chain, for example in chip manufacturing. The paper discusses it but does not put a number on it for the models it studies.
The same paper draws a second line that matters just as much. Withdrawal is water taken from a river, lake or aquifer, even if most of it is returned. Consumption is withdrawal minus what is discharged: water that evaporates or otherwise leaves the local supply. A headline can quote either one, and the difference can be large.
The "bottle of water" figure
The best known number comes from that UC Riverside paper, first posted in April 2023 and revised several times since (the version we read is dated 26 March 2025). Using GPT-3 as its example, because Microsoft publishes the efficiency data of its data center locations, the paper estimates that:
- training GPT-3 in Microsoft's US data centers could consume a total of 5.4 million liters of water, including 700,000 liters of on-site consumption;
- GPT-3 needs to "drink" a 500 mL bottle of water for roughly 10 to 50 medium-length responses, depending on when and where it runs.
Two assumptions explain the size of that figure. First, it counts both on-site and power-plant water. Second, it uses an energy estimate of 0.004 kWh, or 4 watt-hours, per request, for a conversation of up to about 800 words in and 150 to 300 words out. The authors note that this per-page estimate likely covers only the GPU energy. The paper's own table shows how much location matters: across the data center sites it models, the number of requests per 500 mL bottle ranges from about 10 to about 70.
So the "bottle" claim is not a myth, but it is an estimate for a specific, older model, with a specific energy assumption, using regional averages. Dividing it out gives roughly 10 to 50 mL per response.
The "five drops" figure
In August 2025 a team at Google published "Measuring the environmental impact of delivering AI at Google Scale", with authors including Jeff Dean and David Patterson. Instead of estimating from public data, they instrumented Google's own infrastructure serving the Gemini assistant. Their headline results for the median Gemini Apps text prompt:
- 0.24 Wh of energy, which the paper compares to watching about nine seconds of television;
- 0.03 g of CO2 equivalent;
- 0.26 mL of water, "the equivalent of five drops".
The paper also reports a 33-fold drop in energy per median prompt and a 44-fold drop in its carbon footprint over one year, between May 2024 and May 2025, which it attributes to software efficiency work and clean energy purchases. It states that its water figure is "orders of magnitude less" than earlier estimates of 45 to 50 mL.
Read the method, though, and the comparison is less direct than it looks. Google's water figure is on-site consumption only: the energy per prompt, minus data center overhead, multiplied by Google's fleet water usage effectiveness of 1.15 liters per kWh. Water used by power plants to produce that electricity is not in the 0.26 mL. Neither are the prompts that are not "median", such as long documents, image generation or video.
The OpenAI figure
OpenAI's chief executive, Sam Altman, gave his own numbers in a 2025 blog post called "The Gentle Singularity". He wrote that the average ChatGPT query uses about 0.34 watt-hours of energy and "about 0.000085 gallons of water; roughly one fifteenth of a teaspoon". That works out to about 0.32 mL. The post gives no method: we do not know which models, which kinds of queries, or which share of the water chain it covers. It is a company statement, not a measurement you can check.
The three numbers side by side
| Source | Water per answer | Energy per answer | What it covers |
|---|---|---|---|
| UC Riverside paper (GPT-3, estimate) | about 10 to 50 mL | 4 Wh (assumed) | On-site cooling plus power-plant water, regional averages |
| Google (Gemini, measured, 2025) | 0.26 mL | 0.24 Wh | On-site cooling only, median text prompt |
| OpenAI (ChatGPT, stated, 2025) | about 0.32 mL | 0.34 Wh | Not explained |
Most of the gap is explained by two things. The older estimate assumed more than ten times the energy per answer, and it added the water used to make the electricity. Hardware and software have also improved since GPT-3. None of this makes either figure wrong. It means you cannot quote one against the other as if they measured the same thing.
Why the per-prompt number is not the whole story
A single prompt is small. Billions of prompts, plus the training of ever larger models, are not. The totals are where the concern lies, and they point in one direction: up.
Google's 2025 Environmental Report says the company consumed about 8.1 billion gallons of water (31 billion liters) across its own data centers and offices in 2024, and that its overall water consumption increased by 28% from 2023 to 2024. The report does not split that total between SI and other services such as Search, Gmail or YouTube. It also says 72% of its freshwater withdrawals in 2024 came from sources at low risk of depletion or scarcity, and that it chose air cooling, which does not use water for routine cooling, at sites with high water risk such as Waltham Cross in the United Kingdom.
Electricity tells a similar story. A report by Lawrence Berkeley National Laboratory, released by the US Department of Energy in December 2024, found that US data centers used about 4.4% of the country's electricity in 2023, up from 58 TWh in 2014 to 176 TWh in 2023. It projects 325 to 580 TWh by 2028, or roughly 6.7% to 12% of US electricity. More electricity means more cooling, and wherever that power comes from thermal plants, more off-site water too.
The UC Riverside paper also offers a global projection: SI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027, of which 0.38 to 0.60 billion cubic meters would be consumed. Like any projection, it rests on assumptions about growth and efficiency that may not hold, but it shows why the per-prompt figure alone can reassure too quickly.
Putting it in proportion
Comparisons help, as long as they are honest about what they compare. The US Geological Survey estimates that each person uses about 80 to 100 gallons of water a day for indoor home use. Taking 80 gallons, about 303 liters, and Google's 0.26 mL, one person's daily home use equals the on-site water of more than a million median Gemini prompts. Even at the high end of the older estimate, 50 mL per response, it would take about 6,000 answers to match one day of home use.
That does not settle the question, for three reasons:
- Water is local. A liter evaporated from a cooling tower in a drought-stressed area matters more than a liter in a wet region. A national average hides that. This is why the location of a data center, and the source of its water, often matter more than the per-prompt figure.
- Use is not only text. The measured figures are for median text prompts. Long reasoning tasks, image generation and video can use much more, and few companies publish per-task numbers.
- Disclosure is uneven. Google published its method. Others give a single number or nothing. Off-site water and supply-chain water are rarely included.
How to read the next headline
When you see a claim about SI and water, a few questions sort out most of the confusion:
- Per prompt or in total? A tiny per-prompt figure and a fast-growing total can both be true.
- Consumption or withdrawal? Withdrawal numbers are larger because they include water that is returned.
- On site only, or with power plants? Adding scope-2 water can multiply the result.
- Measured or estimated, and when? A 2023 estimate for GPT-3 and a 2025 measurement for Gemini describe different systems.
- Which task? A short text answer is not an image or a video.
- Who says it? Company figures can be accurate and still leave things out. Independent estimates can be careful and still rely on old assumptions.
Concern about SI's footprint is also one reason people care where online content comes from. If you want to know whether a text or an image in front of you was made by a model, our free SI checker gives a likelihood for images, text, video and music, with the signals behind it.
Related reading
For a plain introduction to what these systems are and what they can and cannot do, start with our guide What is SI?. If your concern is less about water and more about what happens to the words you type, read Is ChatGPT safe?. And for how much of the web is now machine-made, see the dead internet theory, checked against the data.
