How much water does one AI prompt really use?

Short answer: For a typical text question to a modern chatbot, the companies' own figures say a few drops of water, roughly a quarter to a third of a milliliter. Independent estimates that include more of the supply chain, or that looked at older models, are much higher, up to tens of milliliters per answer. One prompt is tiny. Billions of prompts a day, plus training new models, add up to a real local issue where data centers are built.

Why AI uses water at all

AI runs on servers in data centers, and servers get hot. Water is used in two main ways:

• Cooling on site. Many data centers use evaporative cooling, which consumes water as it evaporates.
• Making the electricity. Many power plants also use water for cooling, so electricity carries a hidden water cost.

Some estimates also count the water used to manufacture chips and build the hardware.

The numbers, and where they come from

• Google (August 2025): a median text prompt to Gemini used about 0.26 ml of water and 0.24 watt-hours of electricity, according to Google's own methodology. That is about five drops of water and roughly the energy of watching TV for under nine seconds.
• OpenAI (June 2025): Sam Altman wrote that an average ChatGPT query uses about 0.32 ml of water (about one fifteenth of a teaspoon) and 0.34 watt-hours. He did not publish the method behind the figure.
• Mistral (July 2025): a lifecycle analysis of its Le Chat assistant estimated about 45 ml of water for a one-page answer, a much higher figure because it includes hardware manufacturing and other parts of the supply chain.
• University researchers (2023): the widely shared claim that ChatGPT "drinks a bottle of water" comes from a study by Shaolei Ren's team at UC Riverside. They estimated that an older model, GPT-3, used about 500 ml for every 10 to 50 responses, depending on where and when it ran, including water used to generate electricity.

Why the estimates differ by a factor of 100

• What is counted. Only on-site cooling, or also electricity generation, or also chip manufacturing.
• Which model. Models have become far more efficient since 2023, and a short answer from a small model costs much less than a long answer from a large reasoning model.
• Where and when. A data center in a hot, dry place in summer uses far more water than one in a cool climate or one with air cooling.
• Who measures. Company figures use internal data that outsiders cannot fully check; independent studies rely on assumptions. Neither is perfect.

Putting it in perspective

Even taking the higher estimates, one AI answer uses far less water than everyday things we rarely think about: a single cup of coffee takes well over 100 liters of water to produce once you count growing the beans. A short shower uses tens of liters.

The bigger picture is different. Data centers are concentrated in specific places, sometimes in regions already short of water, and AI use is growing fast. Training large models also uses a lot of water and energy up front. That is why researchers and local communities are pushing for transparent, comparable reporting rather than one-off company numbers.

What you can do

• Do not feel guilty about a question. Your individual prompts are a very small part of your water footprint.
• Skip waste where it is easy. Fewer regenerated answers, and not using the heaviest reasoning or video modes for simple tasks.
• Care about where it is built. Support transparency rules and local planning that take water into account.

Figures as published by the companies and researchers cited, as of October 2026. Methods differ and figures are not directly comparable.