Per-query water estimates range from a fraction of a milliliter to hundreds, depending on the model, workload, and measurement method
Efficiency per query can improve while total water demand still rises, since AI computing volume grows faster than efficiency gains
The sharper concern is local: concentrated demand on specific water systems near new data centers
An AI query can use a fraction of a milliliter of water. Or it can use several hundred. The gap comes down to the model running it, the location powering it, and the cooling system pulling heat away from the servers.
In 2024, a Washington Post analysis with UC Riverside researchers estimated that a 100-word GPT-4 email required about 519 milliliters of water. Roughly one bottle. That number combined two separate things: water evaporated directly at the data center for cooling and water used further upstream to generate the electricity that powered the query. It did not measure water consumed inside the facility alone.
The researcher behind the estimate has since placed a GPT-4-class prompt closer to 15 milliliters on that same full-scope basis and near 5 milliliters when counting only onsite cooling.
Google states its median Gemini text query uses about 0.26 millilitres onsite, roughly five drops. OpenAI has cited a similar figure near 0.3 millilitres. Both sit far below the 2024 bottle estimate. Part of the gap comes from a narrower measurement boundary. Part comes from newer, more efficient systems.
Model design changes the outcome too. One 2025 benchmarking study found efficient models staying under 2 milliliters per query across several test lengths. Reasoning-heavy models, which generate long internal steps before answering, used well over 150 milliliters on the same tests. Longer reasoning means more computation.
More computation means more heat to remove. These figures are not directly comparable unless the measurement boundary is stated clearly, and that detail often goes missing in casual coverage.
| Model type | Water use per query |
|---|---|
| Efficient text models (e.g. GPT-4.1 nano class) | Under 2 mL |
| Typical short text query (Google, OpenAI disclosures) | 0.26 to 0.32 mL |
| Full-scope GPT-4-class estimate (2026 update) | ~15 mL |
| Reasoning-heavy models (e.g. DeepSeek-R1 class) | 150+ mL |
| Standard-resolution AI image | ~28.6 mL |
Per-query numbers help compare model efficiency, but they can pull attention away from a larger pattern. A data center can use less water for every unit of computing it performs, while its total water draw still climbs.
That happens when the volume of AI computing grows faster than the efficiency gains. Efficiency improves. Scale outpaces it anyway. This pattern deserves more attention than the average cost of a single reply.
Two terms get used loosely in most coverage of this topic. Water withdrawal means water pulled from a river, aquifer, or utility supply. Water consumption means water that does not return to that source, usually lost through evaporation during cooling.
A facility can withdraw a large volume and return most of it. Or it can consume most of what it withdraws through evaporative cooling towers. A report that quotes only one of these figures, without naming which, is easy to misread.
Generating an image costs more than generating a paragraph of text. A mid-2026 assessment placed a standard-resolution AI image at roughly 28.6 milliliters, over 100 times a short text query. That figure uses a broader footprint method, so it should not sit side by side with narrower, on-site-only numbers.
Video generation has no comparably audited figure yet, so specific claims online deserve some skepticism. A smaller, less visible layer sits underneath all of this: the water used to manufacture the chips and servers themselves, rarely counted in public estimates at all.
A UC Riverside and Caltech study estimated that U.S. community water systems could need between 10 billion and 58 billion dollars in new infrastructure by 2030 to keep pace with data center growth. That figure ties to the pace of new facility construction, not to any single company's footprint.
Utilities plan for peak demand, not averages, and cooling load peaks on the same hot days that strain local water use. This explains the pushback against new data center campuses proposed in parts of Arizona, Georgia and Texas, even where national per-query numbers look modest.
Also Read: Why Iceland is Attracting AI Data Centers
Two engineering paths trade off against each other. Evaporative cooling saves electricity but consumes water. Air cooling and mechanical chillers save water but need more electricity, shifting the cost back to the power plant. Some operators now use direct-to-chip and immersion cooling to target heat more precisely.
Others rely on recycled or non-potable water for cooling loops or choose cooler climates for new sites. These steps help at the facility level, but rising AI demand overall can offset the gains industry-wide.
Also Read: Dell’s AI Flywheel is Taking Shape: From AI PCs to Data-Center Infrastructure
The most useful habit for reading any AI water claim is checking what boundary the number represents before repeating it. As disclosure improves, the sharper signals will come from facility-level reporting and local utility data, not global averages that flatten very different realities into one figure.
1. How much water does a single AI query actually use?
It depends on the model. Simple text queries can use as little as 0.26 to 0.32 milliliters, while reasoning-heavy models can exceed 150 milliliters. There is no single fixed number.
2. Is the viral "one bottle of water per AI email" claim accurate?
The 519 milliliter figure came from a 2024 estimate for a 100 word GPT-4 email. It combined onsite cooling water with water used to generate electricity, not water consumed at the data center alone.
3. What is the difference between water withdrawal and water consumption?
Withdrawal refers to water pulled from a source like a river or utility supply. Consumption refers to water that does not return to that source, usually lost through evaporation during cooling.
4. Why does AI image generation use more water than text generation?
Images require more computation than short text replies. A standard-resolution AI image has been estimated at roughly 28.6 milliliters, over 100 times the water used by a short text query.
5. Why is local water infrastructure a bigger concern than average water use per query?
Utilities plan for peak demand, not national averages. Data center growth in specific regions, especially water-stressed areas, can strain community water systems even when per-query figures nationally appear modest.