About the AI Water Footprint
Every question you ask a large language model has a hidden water cost. Data centers that run AI models consume water in two distinct ways: directly, through evaporative cooling towers that keep servers from overheating, and indirectly, through the water consumed by power plants generating the electricity those servers draw. This calculator estimates both, based on the peer-reviewed methodology popularized by University of California Riverside and University of Texas Arlington researchers in the 2023 paper "Making AI Less Thirsty."
The formula
Total water used (liters) = Queries × [Direct cooling water per query (mL) ÷ 1000] + Queries × [Electricity per query (Wh) ÷ 1000] × [Water consumed per kWh of grid electricity (L)]
The first term captures on-site evaporation at the data center itself. The second term captures the "embedded" water footprint of electricity generation — thermoelectric plants (coal, gas, and nuclear) withdraw and evaporate water to condense steam in their cooling cycles, so every kilowatt-hour drawn from the grid carries a water cost that varies by region and fuel mix.
Reference values used by default
- Direct cooling water per query: researchers estimate that roughly 500 mL of water is consumed by a data center for every 10–50 questions-and-answers exchanged with a model the size of GPT-3, depending on server location and cooling technology. This calculator defaults to 20 mL/query, near the middle of that published range — adjust it if you have a more specific figure for the provider or region you use.
- Electricity per query: independent measurements of modern chatbot inference commonly place a single typical response at roughly 0.3–3 Wh of electricity (some longer or more complex queries, or older/larger models, run higher). The default here, 3 Wh, is a conservative mid-to-upper estimate suitable for general text queries.
- Water per kWh of electricity: thermoelectric power generation in the U.S. consumes on the order of 1.8–2.0 liters of water per kWh on average (varies by cooling method — once-through, cooling tower, or dry cooling — and by fuel source). This calculator defaults to 1.8 L/kWh, a commonly cited average consumption figure; grids leaning heavily on hydro or dry-cooled natural gas can be lower, while some coal-heavy, tower-cooled grids run higher.
Putting the numbers in context
At the defaults (20 queries/day, 20 mL direct + 3 Wh indirect per query, 1.8 L/kWh), the calculator returns about 0.51 liters of water per day, or roughly 185 liters per year — comparable to a couple of showers' worth of water annually for a light daily chatbot user. Heavy users, or workloads that involve image or video generation (which use substantially more compute and therefore more water per request), can be several times higher; training a single large model can consume hundreds of thousands of liters in one data center alone.
What this estimate does not capture
This tool models per-query inference water use, not the one-time water cost of training a model (which is amortized across millions of users) and not manufacturing water embedded in the server hardware itself. It also assumes the figures you enter represent your actual provider and region — real values vary by data center location, time of year (cooling towers work harder in summer), and which AI provider and model size you use.