Article — AI Water Footprint Calculator (ChatGPT)
AI Water Footprint Calculator
An AI water footprint counts the fresh water used to run an AI model: roughly 0.3 to 5 mL per ChatGPT-class query at a typical data center, depending on model size and cooling method. UC Riverside's 2023 paper estimated 519 mL per 100-token GPT-3 query. OpenAI's 2025 figure is 0.3 mL. The truth sits in between — between 0.3 and 25 mL per query, with most modern estimates at a few mL.
About 80% of that water is consumed off-site, at the power plants that generate electricity for the data center. The remaining 20% is direct cooling water that evaporates from the data center's cooling towers. Training a large model like GPT-3 used about 700,000 liters in 2020 — the annual water consumption of 28 average Americans.
What is the AI water footprint?
The AI water footprint is the total fresh water consumed by AI systems. It splits into two parts: the water used during model training (one-time, large) and the water used during inference (per query, ongoing). Both happen inside data centers that need cooling.
Cooling towers evaporate water to dump heat. Each evaporated liter takes about 2.3 megajoules of heat with it. A typical data center has a Water Usage Effectiveness (WUE) of 1.9 L per kWh of compute — meaning every kilowatt-hour of electricity translates to 1.9 liters of water gone. The 2023 industry average is from the Green Grid consortium.
The first major study to quantify AI water footprint was UC Riverside's 2023 paper "Making AI Less Thirsty". It calculated that GPT-3 training used 700,000 liters of fresh water at Microsoft's US data centers. The paper sparked global press coverage and led to data center water transparency demands from regulators in the EU and California.
Where data center water actually goes
Data center water consumption has three main destinations. First, on-site cooling — water that evaporates from cooling towers to reject heat from server racks. Second, off-site power generation — water consumed at the thermal power plants supplying electricity. Third, hardware manufacturing — water used to fabricate the chips, GPUs, and cooling equipment.
For a typical data center running on grid electricity, the split is roughly 20% on-site cooling and 80% off-site power generation. Renewable-energy-powered data centers shift the balance — solar and wind use almost no water (0.0 L/kWh), so the on-site cooling becomes the dominant component. Nuclear plants use 1.5–3 L/kWh, similar to coal and gas.
- On-site cooling = 20% of AI water (evaporative cooling towers)
- Off-site power generation = 80% of AI water (thermal power plants)
- Hardware manufacturing = 1–3 L per CPU/GPU (low fraction of total)
- WUE industry average = 1.9 L/kWh (Green Grid 2024)
- Google's best data centers = 0.07 L/kWh (Finland, free-air cooling)
- Microsoft fleet average WUE = 0.49 L/kWh (2023 report)
- Arizona/Nevada data centers = 4.0+ L/kWh (arid summer cooling)
- Northern Europe data centers = 0.3–0.5 L/kWh (free cooling available)
AI water footprint per ChatGPT query
A single ChatGPT query uses somewhere between 0.3 mL and 25 mL of water depending on whose number you believe. UC Riverside's 2023 paper measured GPT-3 at 519 mL per query (assuming 100 tokens and a US Iowa data center in July). OpenAI's CEO Sam Altman in 2025 cited 0.3 mL per query. Google's 2025 estimate for Gemini is 0.26 mL per query. Independent estimates from 2024 put the figure at 5 to 15 mL.
The wide range comes from different assumptions: which data center, which season, how many tokens per query, whether to include off-site power generation, which model class. Older models like GPT-3 used more energy per token. Newer models (GPT-4 Turbo, Claude 3.5, Gemini 1.5 Pro) are more efficient. Larger context windows mean more tokens per query, which scales water use up. Northern data centers in mild climates use less water than southern data centers in summer.
Training versus inference water cost
Training a large language model uses water once but at enormous scale. UC Riverside estimated 700,000 liters to train GPT-3 — roughly 28 average American households' annual water consumption. GPT-4 is estimated at 1.5 million liters, though OpenAI has not published exact numbers. LLaMA-65B (Meta, 2023) used about 80,000 liters because Meta's data centers have better cooling efficiency.
Inference is the per-query cost. It happens millions of times per day per model. At 5 mL per ChatGPT query and 200 million daily active users sending an average of 5 queries each, ChatGPT alone consumes about 5 million liters of water per day — 2 Olympic swimming pools. Multiply that by 365 days, multiply again by the number of AI services worldwide, and the global figure climbs into billions of liters per year.
Cooling methods and water tradeoffs
Three main cooling strategies dominate modern data centers. Water cooling (evaporative cooling towers) is the most common — moves heat efficiently but evaporates water. Air cooling (massive fans pushing outside air through server racks) uses minimal water but consumes more electricity for the fans, which indirectly uses more water at the power plant. Immersion cooling submerges chips in non-conductive fluid like 3M Novec — cuts on-site water by 95% but costs about 3x more upfront.
The choice depends on location. Hot, dry climates (Arizona, Nevada, Texas) tend toward water cooling because air cooling is too expensive when the outdoor temperature exceeds the server-rack temperature. Cool climates (Iceland, Norway, Finland) can use free-air cooling almost year-round, with WUE under 0.3 L/kWh. The push for AI compute in arid southwestern states puts pressure on water supplies that are already stressed by drought.
Tech companies have built data centers in Arizona, Texas, and Saudi Arabia — regions with cheap power and weak water regulations but the hottest summers. Cooling water demands in these locations exceed those in northern locations by 3 to 10 times. The 2022 Arizona drought forced one Microsoft data center to pause expansion plans, and Google paused construction near The Dalles, Oregon over water-permit conflicts.
Global scale of AI water consumption
US data center water consumption grew from 21.2 billion liters in 2014 to about 66 billion liters in 2023, an increase of 211% over nine years. Most of that growth is attributable to AI workloads, especially after the 2023 ChatGPT boom. The trend continues — 2024 estimates put US data center water at ~80 billion liters.
Projections by industry researchers (Trends Research & Advisory, 2024) put global AI water demand at 4.2 to 6.6 billion cubic meters per year by 2027. That equals the annual water consumption of Denmark or about half the United Kingdom. Even the lower estimate puts AI water demand at 1% of global freshwater extraction, comparable to the entire concrete industry.
How to reduce your AI water footprint
Individual users have limited direct control, but a few choices matter. First, use smaller models when the task allows — Gemini Flash, Claude Haiku, and GPT-3.5-Turbo use roughly 10–20% the water of large frontier models. Second, batch your queries — one well-formed query with full context uses less water than five back-and-forth exchanges. Third, choose cloud providers that publish their WUE — Google Cloud (0.95 L/kWh) and Microsoft Azure (0.49 L/kWh) report transparently; many smaller providers do not.
Institutional pressure works better. Microsoft and Google have committed to "water positive" operations by 2030, meaning they will return more water to local watersheds than they consume. OpenAI has not made an equivalent commitment as of 2025. The EU Sustainability Reporting Directive (CSRD) may require water disclosure by 2026; the US has no equivalent federal requirement yet.
If you run AI inference at scale (B2B software, research, or large-volume content generation), pick a region with low WUE. Google's Finland (Hamina) data center has WUE near 0.07 L/kWh. Microsoft's Sweden region is similar. By contrast, US-South-Central regions run 2–4 L/kWh in summer. The carbon-aware computing literature applies equally to water-aware computing.
Controversy over the per-query water number
The "AI water footprint" became a public flashpoint after UC Riverside's 2023 paper. Critics (notably the Sean Goedecke blog post and OpenAI's own statements) argued that the 519 mL figure assumed an inefficient data center, peak summer conditions, and outdated GPT-3 architecture — not representative of modern AI. OpenAI's 2025 claim of 0.3 mL per query is 1700 times lower, which is hard to verify because OpenAI has not published its methodology.
The truth is methodological. UC Riverside used a transparent accounting that included off-site power generation water. OpenAI's lower figure likely excludes off-site water (counting only direct data-center consumption) and uses newer, more efficient architectures. Both numbers are arguably correct under their respective accounting frameworks. For consumer-facing reporting, the mid-range estimate of 1–10 mL per query is more defensible than either extreme.
1 query (GPT-4) ≈ 5 mL water100 queries ≈ 0.5 L = 1 bottle10,000 queries ≈ 50 L = 0.5 shower1,000,000 queries ≈ 5000 L (about 50 bathtubs)Train GPT-3 = 700,000 L = 28 US households/yr