AI uses water in two ways: directly, to cool the servers that run it, and indirectly, through the water needed to make the electricity that powers them. How much is harder to pin down than headlines suggest. One chatbot prompt uses a small amount, the total across all data centers is large and growing, and the number that matters most is often local.
The Short Answer
- U.S. data centers directly consumed about 17 billion gallons of water in 2023, according to a 2024 Lawrence Berkeley National Laboratory (LBNL) report. Water used to generate their electricity added roughly 211 billion gallons.
- Company figures put a single AI prompt at a fraction of a milliliter, about a few drops. Estimates that count power-plant water run higher.
- These figures cover data centers as a whole, so they include email, streaming, and cloud storage along with AI.
Why AI Uses Water
Cooling. Servers produce a lot of heat. Many data centers cool them with water that evaporates, which removes the heat but also uses the water up.
Electricity. Power plants, especially those that boil water to make steam, use water too. How much depends on the fuel mix of the local grid.
Chip manufacturing. Making chips also takes water, but most data center estimates leave it out.
Studies also use different measures. Consumption is water lost, usually through evaporation, and not returned to its source. Withdrawal is water taken from a source, some of which may go back. The two aren’t interchangeable, which is one reason published numbers don’t always match.
How Much Water Do U.S. Data Centers Use?
The most cited government estimate comes from LBNL. As explained in this analysis in The Conversation, the lab estimated that U.S. data centers consumed 17 billion gallons directly through cooling in 2023 and another 211 billion gallons indirectly through electricity, about 12 times as much. The report projects the direct figure could double or even quadruple by 2028.
USAFacts compared the 17 billion gallons to the annual water use of about 155,000 homes.
Company reporting doesn’t fill the gaps. Sustainability reports are voluntary, companies measure water differently, and the disclosures often leave out indirect use from electricity.
How Much Water Does One AI Prompt Use?
| Source | Estimate | What it covers |
|---|---|---|
| Google (August 2025), median Gemini text prompt | 0.26 mL, about five drops | Water used at its data centers, not power-plant water |
| OpenAI CEO Sam Altman (June 2025), average ChatGPT query | About 0.32 mL (0.000085 gallons, roughly one-fifteenth of a teaspoon) | Company statement with no published method |
| Independent estimates that include electricity | Higher, and widely varying | Adds water used at power plants |
Estimates differ because of what they count, the size of the model, the length of the response, the cooling method, and the local electricity mix.
The widely repeated “bottle of water” claim traces to University of California, Riverside researchers, who estimated about 500 milliliters for a batch of roughly 10 to 50 medium-length responses. Many viral posts turned that into one bottle per question, which is not what the study says.
Newer systems such as agentic AI can make many model calls to finish one task, so they probably use more computing than a single prompt. We found no reliable water figures for them yet.
Training vs. Everyday Use
One UC Riverside estimate put the direct evaporative water used to train GPT-3 at about 700,000 liters, or roughly 185,000 gallons. Training happens once per model version, while everyday use repeats billions of times. Researchers who study AI energy use say everyday use can outweigh training once a model has many users, and water use follows energy use to a large extent.
Why Local Impact Matters More Than the National Total
Whether a national total sounds large or small depends on what you compare it to. The comparison that matters most is local. A data center in a dry region draws on the same supply as nearby homes and farms, and even a modest increase in withdrawals can matter for the resilience of a particular river basin, as water researchers have pointed out.
How Data Centers Can Use Less Water
- Different cooling methods. Closed-loop liquid cooling and air cooling use less water, but they can need more electricity, which carries its own water cost.
- Siting and sources. Using reclaimed or non-drinking water, and building where supplies are less strained.
- More efficient hardware and models. Less energy means less heat to remove.
- Consistent disclosure. Reporting both direct and indirect water use in a comparable way.
Should You Worry About Your Own AI Use?
Based on company figures, one prompt uses a few drops, so skipping a question won’t change much. The bigger levers are where data centers are built, how they are cooled, and whether companies are open about what they use.
Conclusion
AI’s water use is real, growing, and badly measured. The best federal estimate covers all data centers, company disclosures are inconsistent, and per-prompt figures vary depending on what is counted. The clearest takeaway is that local water supply, not the national total, is where the impact is most likely to matter.
At Future AI Wire, we’ll keep following the research as better data becomes available.
Frequently Asked Questions
How much water does ChatGPT use per question?
OpenAI’s CEO said an average query uses about 0.32 milliliters of water. OpenAI hasn’t published its method, and estimates that include power-plant water are higher.
How much water do U.S. data centers use?
A 2024 LBNL report estimated 17 billion gallons consumed directly in 2023, plus about 211 billion gallons used indirectly to generate their electricity.
Is AI water use a problem?
It depends on where the data centers are. In water-stressed regions, even small additional demand can matter, while the national total can look modest in comparison.
What is the difference between water consumption and withdrawal?
Consumption is water lost, usually through evaporation, and not returned to its source. Withdrawal is water taken from a source, part of which may be returned.
The Future AI Wire Editorial Team covers artificial intelligence, AI agents, agentic AI, and emerging technologies. We publish clear, research-informed explanations and practical insights to help readers understand AI developments, real-world applications, and the technologies shaping the future.