An ai water footprint calculator estimates how much fresh water is consumed when you ask an AI model a question. It cools the data centre and generates the electricity that the model uses. Pick a model and a query size from a published study, say how often you ask, and see the water per query, per day and over a period.
How to use the ai water footprint calculator
- Choose a model from the study, or your own figures.
- Choose the size of the query.
- Enter the queries per day and the number of days.
- Press Calculate.
Where the figures come from
The method and the numbers come from a research paper by Jegham and others, titled How Hungry is AI, in its sixth version of 24 November 2025 [1]. We built this tool from the paper, which we read in full. Our numbers will therefore differ from any other calculator that uses newer models or different data.
The formula
The paper's equation 4 gives the water for one query: the energy of the query divided by the PUE, times the on-site water usage effectiveness, plus the energy times the off-site water usage effectiveness [1]. The energy is in kilowatt hours and the effectiveness values are in litres per kilowatt hour. The first term is the water used to cool the data centre. The second is the water used by the power stations that make the electricity. We multiply the water per query by your queries and days, which is our own working.
The multipliers
The paper's first table gives the multipliers by host [1]. For models on Microsoft Azure, the PUE is 1.12, the on-site effectiveness is 0.30 and the off-site effectiveness is 4.35 litres per kilowatt hour. For models on AWS, they are 1.14, 0.18 and 5.11. For DeepSeek's own data centres in China, they are 1.27, 1.20 and 6.016. The calculator picks the right set for the model you choose.
The energy per query
The paper measures the energy of each model for three sizes of query: 100 tokens in and 300 out, 1,000 in and 1,000 out, and 10,000 in and 1,500 out [1]. For example, GPT-4o uses 0.423 watt hours for the short query, and Claude-3.7 Sonnet uses 0.950, 2.989 and 5.671 watt hours for the three sizes. The study left out the longest size for GPT-4 and for two LLaMA-3 models, so the calculator says so if you choose them.
A worked example
For GPT-4o and the short query, the energy is 0.000423 kilowatt hours. The on-site term is 0.000423 ÷ 1.12 × 0.30 = 0.1134 millilitres, and the off-site term is 0.000423 × 4.35 = 1.840 millilitres. Together they give 1.953 millilitres per query. At 8 queries a day for 365 days, that is about 5.7 litres a year. These figures are our own arithmetic on the paper's tables.
Checking against the paper
The paper says that a long query to DeepSeek-R1 on DeepSeek's own host uses over 200 millilitres, and that on Azure the same query needs only 34 millilitres [1]. This calculator gives 202.4 and 34.22 millilitres for those two cases, so it reproduces the statements in the paper. The gap shows how much the host matters: the same model uses six times as much water in a data centre with a higher PUE and higher water effectiveness.
Your own figures
The models in the study are not the newest ones, and newer models are not in the paper. If you have the energy per query and the PUE and water effectiveness of a data centre for another model, choose your own figures and enter them. The same equation is applied. You must source those values yourself. We make no claim about models that the paper did not measure.
What the figure leaves out
The paper focuses on operational resource use during the inference phase, and it excludes the water and emissions of hardware manufacturing and supply chains, called Scope 3 [1]. The figure is therefore the water of running queries only, and the true footprint of a model is larger. The paper does not cover the training of the model in its method, and neither does this tool. We carry this limit over.
How reliable is it
The numbers are estimates. They rest on public performance data, inferred hardware and company-level multipliers [1]. As a check, the paper notes that the chief executive of OpenAI said in June 2025 that a ChatGPT query used about 0.34 watt hours, against the paper's estimate of 0.42 for GPT-4o [1]. That is close, but it shows the uncertainty. The water follows the energy, so a lower energy would give less water.
Reading the result
The result shows the water per query in millilitres and in US fluid ounces, the water per day and over your period, the electricity, and a split between cooling and power generation. We also give the total as a number of bottles of 500 millilitres, which is our own choice of a familiar size. Compare models and query sizes to see the spread, and remember that a short question uses far less than a long document.
Common mistakes
- Treating the result as a measurement. It is an estimate from a paper.
- Mixing up query sizes. The three sizes differ a lot in energy.
- Comparing with a newer model. The study does not include it.
- Forgetting the host. The same model uses different amounts of water on different hosts.
What next
Use this often? Add Ai Water Foo to your Home screen.
Frequently asked questions
What is an ai water footprint calculator?
A tool that estimates the fresh water consumed by the cooling and the electricity behind an AI query.
How is the water per query worked out?
Energy divided by PUE times on-site water effectiveness, plus energy times off-site water effectiveness.
Which models does the calculator cover?
The 30 models measured in the 2025 study, from GPT-4.1 to LLaMA-3.3, and your own figures.
Does it include the water used to train a model?
No. The study counts inference only and leaves out the water used to make the hardware.
Why does my result differ from other calculators?
They may use newer models or different multipliers. This one follows the paper.
Was this tool helpful?
What went wrong?
We only receive the reason you pick, never what you typed into the tool.
Thanks for your feedback.