How much water did AI use in 2025?
AI's water use is not metered separately, so there is no audited 2025 total. Our best-anchored estimate: US data centers, where most AI runs, directly consumed roughly 23–33 billion gallons of water in 2025 (Low, cooling-only boundary; site inference from LBNL's 2023 measurement and 2028 projection). A single chat query used 0.3–45 mL, depending on what you count.
Why the answer is a range, not a number
Three things keep 2025 from having a single figure.
First, nobody bills "AI" for water. Utilities meter data centers, and data centers run everything: streaming, email, payroll, and AI. The Lawrence Berkeley National Laboratory's 2024 report, still the most rigorous public accounting, measured US data centers at 176 TWh of electricity and about 17 billion gallons of directly consumed water in 2023, and attributed most of the growth since 2017 to AI-accelerated servers (LBNL 2024). Everything after 2023 is projection.
Second, the boundary you draw changes the answer by two orders of magnitude. Count only the cooling water evaporated on site (Low, cooling only) and a chat query costs well under a milliliter. Add the water consumed generating the electricity (Mid, + electricity) and it's a few milliliters. Add training runs and chip manufacturing (High, full lifecycle) and Mistral's lifecycle estimate lands at 45 mL per 400-token response (Mistral LCA, 2025).
Third, 2025 was, somewhat ironically, the first year companies published official per-prompt numbers, which narrowed the per-query debate while leaving the annual total as foggy as ever.
What your own 2025 looked like
A daily chatbot habit over all of 2025 (ten ChatGPT queries a day) works out to roughly 26 liters at Mid (+ electricity): about 40% of one shower. The gap between that and "billions of gallons" is the story of this page: the totals are an infrastructure question, not a personal one.
frontier class · 2 Wh per 1,000 tokens (mid benchmark)
The same prompt can be “five drops” or “a bottle of water” depending on what you count: just data-center cooling (Low), the water behind the electricity (Mid), or training and hardware too (High).
Read the methodology16.9gallons
= 63.9 liters
ChatGPT (standard), 25 prompts/day · per year, Mid scenario
Energy
12.8 kWh
Carbon
4.5 kg CO₂e
What was actually disclosed in 2025
2025 was the year the per-query figures stopped being purely academic:
- Google published a technical disclosure: a median Gemini Apps prompt uses 0.24 Wh of energy and 0.26 mL of water (Low, cooling only; Google 2025).
- Sam Altman wrote in June 2025 that an average ChatGPT query uses 0.34 Wh and about 0.32 mL of water (Low, cooling only; Altman 2025).
- Mistral released the first externally reviewed lifecycle assessment from a model developer: 45 mL of water and 1.14 gCO₂e per 400-token response (High, full lifecycle; Mistral LCA 2025).
Note what happened there: the two operator disclosures and the one lifecycle study differ by a factor of roughly 150, and none of them is wrong. They are answering differently sized questions. Google and OpenAI counted the water their cooling systems consumed per query; Mistral also counted the power plants, the training runs, and the factories that made the chips. The widely cited ~519 mL per 100-word email comes from a 2024 Washington Post analysis with UC Riverside's Shaolei Ren; it counts cooling plus electricity water at a high 2024 energy estimate for GPT-4, not the full lifecycle. The UC Riverside paper's own estimate, ~500 mL per 10–50 medium-length responses on GPT-3-era infrastructure, uses the same cooling-plus-electricity boundary (Li & Ren, CACM 2025).
Bottom-up: queries × water per query
One way to build a 2025 total is from the queries up. OpenAI disclosed roughly 2.5 billion ChatGPT queries per day in 2025; scaling for Gemini, Claude, Meta AI, and the rest, we estimate ~5 billion AI chat queries per day globally (estimated; this site's ticker assumption). At a blended ~0.8 Wh per query (a site assumption between our standard-class ~0.56 Wh and frontier-class ~1.4 Wh per exchange; operator disclosures run 0.24–0.34 Wh), that's about 1.5 TWh per year of chat inference. Applying each scenario's water factor:
- Low (cooling only): ~0.24 mL/query → roughly 0.4 billion liters (~0.1 billion gallons) per year
- Mid (+ electricity): ~4 mL/query → roughly 7.3 billion liters (~1.9 billion gallons) per year
- High (full lifecycle): ~24 mL/query → roughly 44 billion liters (~11.6 billion gallons) per year
All three are site inferences from sourced coefficients (Siddik 2021 for the Mid factor; Mistral LCA 2025 for the High factor). And all three are floors, not totals: they cover chat queries only. Training runs, image and video generation, coding agents, and enterprise API traffic (the workloads that actually fill data centers) are not in these numbers.
Top-down: the data-center view
The other way is from the buildings down. LBNL measured 176 TWh and ~17 billion gallons of direct water consumption for US data centers in 2023, and projected 325–580 TWh by 2028 (LBNL 2024). Interpolating that trajectory puts 2025 at roughly 240–340 TWh. If water scaled proportionally (a simplification that assumes cooling efficiency held steady), direct consumption in 2025 was roughly 23–33 billion gallons (Low, cooling-only boundary; site inference, since the LBNL baseline counts on-site water).
Include the water consumed generating the electricity (~4.5 L/kWh on the US grid, Siddik 2021) and the all-in figure for US data centers in 2025 (direct plus electricity water) lands in the neighborhood of 310–440 billion gallons (Mid, + electricity; site inference), though again, that is all data-center workloads, of which AI is the fastest-growing slice, not the whole.
For scale: 23–33 billion gallons of direct consumption is about what 0.8–1.1 million Americans use at home in a year (82 gal/person/day, EPA WaterSense). Significant, and growing fast. But the more useful 2025 takeaway is where it's growing, since a gallon evaporated in a water-stressed county is not the same as a gallon in a rainy one.
What companies themselves reported for the prior year supports the trajectory: Google disclosed about 7.7 billion gallons consumed across its data centers in 2024 (Google Environmental Report 2025; the widely quoted 6.1 billion was 2023), and Equinix 1.2 billion gallons (Equinix 2025): two companies, already more than half of LBNL's entire 2023 US figure, though their fleets are partly outside the US.
What 2025 didn't tell us
Three gaps remain. No one published an AI-specific total: operator disclosures cover whole fleets, and LBNL covers whole buildings. Withdrawal versus consumption is still routinely confused: a plant that borrows river water and returns it is doing something different from a cooling tower that evaporates it, and headlines rarely distinguish. And location is everything: per-query milliliters aggregate into local, not global, stress. Those gaps are why this site reports ranges with labeled boundaries rather than one confident number.
This page stays up as the historical record; the 2026 page tracks the current year.