MTS

Wiki

AI Water Consumption: The Numbers

"Each ChatGPT query uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon." — Sam Altman, The Gentle Singularity (Jun 2025)

"It is not possible to look at the numbers involved without coming to the conclusion that this is a fake problem." — Andy Masley, The AI water issue is fake

The headline numbers

Quantity Value Source
Water per ChatGPT query (OpenAI, Jun 2025) 0.000085 gal / 0.32 mL Altman, Gentle Singularity
Water per conversation (Goedecke, Oct 2024) ~5 mL (per conversation) seangoedecke.com
Water per query (Ren 2023, GPT-3 + Azure 2022) 10–50 mL arXiv:2304.03271
All US data centers — daily freshwater (2023) 200–250 million gal Masley / EESI
Total US daily freshwater consumption ~132 billion gal USGS / Masley
Data centers as share of US freshwater (2023) ~0.2% derived from EESI / USGS
AI-specific share of US freshwater (2023) ~0.008% Masley
Projected AI share by 2030 ~0.08% Masley
US agriculture share of US freshwater ~70–80% USGS

The story in five beats

  1. The viral number is wrong. "ChatGPT uses 500 mL per query" — the figure that launched a thousand op-eds — comes from a 2023 paper (Ren et al.) that measured 500 mL per 10–50 prompts on GPT-3 running on Microsoft Azure data centers in 2022 using evaporative cooling on a water-stressed grid. Stripping out the misreading, that's 10–25 mL per query, not 500. Two model generations and one efficiency cycle later, the realistic figure is between 0.3 mL (OpenAI / Altman) and 5 mL (Goedecke). See the_500ml_myth.md for the full chain of misreadings.

  2. A hamburger is two million prompts. A single quarter-pound burger has a water footprint of ~660 gallons (Water Footprint Calculator). At 5 mL per prompt, that buys you ~500,000 prompts. At Altman's 0.32 mL figure, ~7.8 million. A pair of jeans: ~1,800 gallons, several million prompts. One almond: roughly one gallon, ~750–3,000 prompts. Full table in comparisons.md.

  3. Most "AI water" isn't water AI uses — it's water power plants use. Of an estimated 16.9 mL per GPT-3 query in 2023 (Ren et al.), only 2.2 mL was on-site cooling. The other 14.7 mL was indirect — water consumed at coal and gas plants generating the electricity. That water is going to be consumed whether the load on the grid is AI, Bitcoin, lighting, or air conditioning. Decarbonising the grid eliminates ~80% of the "AI water footprint" automatically. Detail: scope_2_water.md.

  4. The cooling trade-off goes the other way. Critics frame evaporative cooling as a sin. But the alternative — closed-loop / air cooling — uses roughly 10% more electricity, which (today) means more fossil generation, more CO₂, and more water consumed at the power plant. (OECD.AI on this trade-off.) Operators in arid regions are already migrating to air cooling on cost grounds.

  5. Local stress is real; national alarm is not. US-wide, AI is a rounding error on freshwater. But ~40% of US data centers sit in high or extreme water-stress regions (Lincoln Institute) — Phoenix (Stanford & The West), the Atlanta exurbs, parts of Texas. The legitimate policy conversation is about siting, permitting, and revenue-sharing in those communities — not about your ChatGPT habit. See local_vs_national.md.

What the editorial should not claim

  • "AI uses zero water." It uses water. Just not interestingly more than other comparable industrial loads.
  • "There are no legitimate concerns." Construction-phase impacts (e.g. the Meta site in Newton County, GA), local groundwater stress in arid siting, and transparency gaps in scope-2 reporting are all real.
  • "All numbers in the discourse are equally bad." The Ren et al. paper itself is reasonable scholarship; its viral misrepresentation is the problem.

Where to go next