How Much Water Does AI Use? And Which AI Uses the Least?
How much water does AI use? ChatGPT, Gemini and Mistral per-prompt figures, the total for AI data centers, why the numbers differ so much, and which AI uses the least water.
No AI uses zero water, but the amounts per answer are small and vary widely. The lowest published figures are Google’s 0.26 mL for a median Gemini text prompt, about five drops, and OpenAI’s about 0.32 mL for an average ChatGPT query. Both count only cooling water at the data center.
What decides how much water an AI uses is how its data center is cooled, the local climate, how its electricity is made, and how big the model is.
Why AI uses water at all
AI runs on chips that get hot. Water comes into it in three places:
- Cooling at the data center. Many data centers use evaporative cooling towers. They work well, but the water that evaporates is gone from the local supply.
- Making the electricity. Coal, gas and nuclear plants use water for cooling too. So an AI running on that grid uses water indirectly, even if its data center uses none on site.
- Making the hardware. Chip factories use large amounts of ultra-pure water.
Researchers also separate withdrawal, water taken from a river, lake or aquifer, from consumption, water that is evaporated or otherwise not returned. A closed cooling loop that returns water warm, rather than evaporating it, withdraws water but consumes little.
How much water does AI use per prompt?
| Source | Water per text answer | What it counts |
|---|---|---|
| Google (Gemini) | 0.26 mL, about five drops | Cooling water consumed at Google’s data centers, for a median prompt (May 2025 data) |
| OpenAI (ChatGPT) | About 0.32 mL, one fifteenth of a teaspoon | A single figure from OpenAI’s CEO, with no method published |
| Mistral (Le Chat) | 45 mL for a 400-token answer | A full life-cycle study that also covers upstream impacts like making the hardware |
The gap between 0.26 mL and 45 mL is mostly about what gets counted, not about one AI being 170 times thirstier. On-site cooling water alone is small. Add the water used to generate the electricity and make the hardware, and the number grows a lot. When you compare AI apps, check that you are comparing like with like.
For scale: a 250 mL glass of water equals about 960 median Gemini prompts using Google’s on-site figure, or about five answers using Mistral’s full life-cycle figure.
How much water does AI use in total?
Per answer the amounts are tiny. In total they add up. Researchers at UC Riverside and UT Arlington estimated that training GPT-3 in Microsoft’s US data centers could directly evaporate 700,000 liters of clean freshwater. They project that global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027, comparable to the annual water withdrawal of four to six countries the size of Denmark.
Water use also matters more in some places than others. A data center evaporating water in a dry, hot region puts more pressure on local supplies than one in a cool, wet one.
What makes an AI use less water
- Cooling that doesn’t evaporate water. Closed-loop liquid cooling recirculates the same water. Microsoft says its new data-center designs, in use since August 2024, avoid more than 125 million liters of water a year per data center this way, though they use slightly more electricity.
- A cool climate. In cold places, outside air does much of the cooling for free, so less water and less electricity go to it.
- Electricity that needs little water. Wind and solar farms need very little water to run, unlike coal, gas and nuclear plants.
- An efficient data center. Water usage effectiveness (WUE) measures liters of water per kilowatt-hour. Microsoft reported 0.30 L/kWh across its data centers in fiscal 2024, down from 0.49 in 2021.
- A smaller model and a shorter job. Less computing means less heat, which means less cooling. Water use tracks energy use.
How Viro handles this. Viro’s own model, Viro AI, uses far less water than typical AI. Data centers usually consume water through evaporative cooling; Viro’s sites sit in cold northern climates where outside air and closed-loop liquid cooling carry the load instead. Our Norwegian site draws fjord water at 6–9°C in a closed loop and returns it warm, on to nearby fish farms, not evaporated away. When you choose ChatGPT, Claude or Gemini inside Viro, those answers run in their makers’ data centers. See the data centers.
How to use AI with less water
- Use a smaller model for everyday questions; water use tracks energy use.
- Avoid regenerating the same answer or image again and again.
- Send only the part of a document you need.
- Choose an AI provider that says where its data centers are and how they are cooled.
For the energy side of the same question, read how much energy ChatGPT uses per query, or the wider picture in what is eco-friendly AI?
Try Viro AI free, with its own model running on renewable power in cold northern data centers.
Frequently asked questions
Is there an AI that doesn’t use water?
No. Every AI answer needs electricity and cooling, and making the chips uses water too. Some AI uses very little: data centers in cold climates with closed-loop cooling, run on renewable power, use far less water per answer than average.
How much water does ChatGPT use per question?
OpenAI’s CEO has said an average ChatGPT query uses about 0.000085 gallons of water, roughly one fifteenth of a teaspoon or about 0.32 mL. OpenAI has not published how that figure was measured.
How much water does Gemini use per prompt?
Google measured 0.26 mL, about five drops, for its median Gemini Apps text prompt, using May 2025 data. That counts water consumed for cooling at Google’s data centers.
Why do AI water estimates vary so much?
Because they count different things. Google’s 0.26 mL counts only cooling water at its data centers. Mistral’s 45 mL is a full life-cycle figure that also covers upstream impacts like making the hardware. Older estimates also assumed much higher energy per answer than today’s models use.
Does Viro AI use water?
Far less than typical AI. Data centers usually consume water through evaporative cooling; Viro’s sites sit in cold northern climates where outside air and closed-loop liquid cooling carry the load instead. Its Norwegian site draws fjord water at 6–9°C in a closed loop and returns it warm, on to nearby fish farms, not evaporated away.
Sources
- How much energy does Google’s AI use? We did the math, Google Cloud, August 2025
- The Gentle Singularity, Sam Altman, OpenAI, June 2025
- Our contribution to a global environmental standard for AI, Mistral AI, July 2025
- Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models, Li, Yang, Islam and Ren, Communications of the ACM, 2025
- Sustainable by design: Next-generation datacenters consume zero water for cooling, Microsoft, December 2024
Cover photo by César Couto on Unsplash
