Tools

Advertisement

AI water footprint calculator

Every AI reply is generated on a server that draws electricity and is cooled, usually with water. Enter how many queries you or your organisation send each day and the calculator returns the water and electricity per query, per day and per year.

OpenAI and Google publish average energy per prompt. Other vendors do not, so the estimate uses the mean of those two disclosures.

Water per query (mL)
0.34
Energy per query (Wh)
0.340
Water per day
0.01
Electricity per day
0.01
Water per year
3.7
Electricity per year
3.7
What this means
At 1 users this is 4 L of water and 4 kWh a year. The published figures cover the data centre, not the water used to generate the electricity, which some researchers put at two to three times more.

Estimates for information only. Not financial, medical, legal or engineering advice. See the Terms.

Daily water and electricity by number of users
UsersQueries per dayWater (L per day)Electricity (kWh per day)
1.0030.000.010.01
4.00120.000.040.04
10.00300.000.100.10
4,000.00120,000.0040.8040.80

Advertisement

How this calculator works

A language model answers by running billions of multiplications on specialised chips. The chips draw electricity, that electricity carries a carbon and water cost, and the servers need cooling, which usually means evaporating water. OpenAI published an average of 0.34 watt-hours per ChatGPT query and Google published 0.24 watt-hours, 0.26 millilitres of water and 0.03 grams of CO2e for a median Gemini prompt; both disclosures work out to roughly 1 millilitre of water per watt-hour, which is the rate used here. Multiplying the number of queries by the energy per query, the number of users and the days in a year gives the totals. The figures are averages for text prompts on each company's own infrastructure, so a long conversation, an image or a reasoning model with many internal steps can use several times more. The estimate is a fair way to compare habits and to size a sustainability report, not a metered measurement.

Formula

queries per day = queries per person x users. water per query (mL) = energy per query (Wh) x 1.0 mL/Wh. water per day (L) = queries per day x water per query / 1000. electricity per day (kWh) = queries per day x energy per query / 1000. Multiply by 365 for a year. The per-query energy is 0.34 Wh for ChatGPT, 0.24 Wh for Gemini, or the mean 0.29 Wh for another model without a disclosure.

Worked example

Thirty queries a day for one person on ChatGPT. Water is 30 x 0.34 x 1.0 = 10.2 mL a day, about 3.7 L a year. Electricity is 30 x 0.34 / 1000 = 0.0102 kWh a day, or 3.7 kWh a year. For a corporation with 4,000 users on the same habit, the daily figures are 4,000 times larger: 40.8 L of water and 40.8 kWh of electricity a day.

Advertisement

Frequently asked questions

Is the water really used up?

Most of it evaporates during cooling and leaves the local watershed, which is why data centres in dry regions attract attention. The published figures include on-site cooling; researchers such as Li and colleagues at UC Riverside argue that water used to generate the electricity should be added, which raises the total.

Why do the OpenAI and Google figures differ?

They cover different models, different hardware and different system boundaries. Google reports a median prompt with a full data centre allocation including idle capacity and power overhead, while OpenAI reports only an average. Neither figure includes the water used to produce the electricity.

Does a long chat use more energy than several short ones?

Yes. Each new message resends the earlier conversation as context, so a long thread costs more per reply. The per-query averages assume ordinary-length prompts, so treat a very long session as several queries.

Can I use this for a carbon report?

Use it as an estimate and cite the published per-query source. Google also publishes about 0.03 g CO2e per median prompt; multiply by your query count for a rough carbon figure, and use your own electricity contract for a defensible number.

Sources

Written byAmira MansourCalculator writer. Builds and tests Mathulate calculators against published sources.
Reviewed byMarc LefebvreMathematics teacher. Checks each formula and worked example by hand.

Last updated .

Advertisement