How this instrument works
Every time a large language model answers a prompt, the data center running it burns electricity and, in many facilities, evaporates water for cooling — either directly through on-site cooling towers or indirectly through the water used to generate that electricity at a power plant. That water use is invisible to the person typing the question, but it is real and measurable.
The most-cited figure comes from Li, Yang, Islam & Ren's 2023 paper 'Making AI Less Thirsty,' which found that training GPT-3 in a U.S. data center directly evaporated about 700,000 litres of freshwater, and that a typical exchange with a GPT-3-scale model consumes somewhere around 10 to 50 millilitres of water per query — the paper is explicit that this swings widely by model size, hardware, region, and even the season, since cooling systems work harder in hot, dry climates.
This calculator uses that published range as a selectable low/mid/high factor rather than inventing one fixed number: pick the estimate that matches how conservative or generous you want to be, and it converts your query volume into total litres and the equivalent number of 500 mL bottles — a scale that's easier to picture than an abstract litre count.
- Enter AI chatbot queries per day — how many prompts you typically send.
- Enter Number of days — the period you want to total up (a day, a month, a year).
- Choose a Water-per-query estimate — low, mid, or high end of the published 10-50 mL/query range.
- Read Estimated total water footprint — the total litres consumed over that period.
- Read Equivalent number of 500 mL bottles — the same total expressed as bottled water.
Worked example — a month of daily chatbot use
Someone sends 20 AI chatbot queries a day for 30 days, and picks the paper's mid-range estimate of 20 mL per query — the midpoint of the published 10-50 mL/query range. That's 20 × 30 × 20 ÷ 1000 = 12 litres of freshwater over the month.
12 litres is the equivalent of 24 standard 500 mL bottles of water (12 × 1000 ÷ 500 = 24) — roughly one bottle's worth of water for every 25 queries sent, at this mid-range estimate. Switching the selector to the paper's high-end 50 mL/query figure would push that same month's usage up to 30 litres, a reminder of how much this number depends on which model and data center answered the query.
Questions
Why does the water-per-query estimate range from 10 to 50 mL?
Because the underlying study found that water use per query depends heavily on factors outside a user's control: which model answers the prompt, how efficient that data center's cooling system is, which region it's located in, and even the season, since evaporative cooling towers work harder in hot, dry weather. Rather than pick one number and imply false precision, this calculator exposes the paper's full published range so you can choose a conservative, midpoint, or generous estimate.
Does this include the water used to train the AI model, not just to run it?
No — this figure covers only inference, the water consumed while the model is actively answering your queries (sometimes called the 'operational' or 'runtime' water footprint). Training a large model is a separate, one-time cost; the same 'Making AI Less Thirsty' paper found that training GPT-3 alone directly evaporated roughly 700,000 litres in a U.S. data center, spread across millions of future queries rather than any single user's usage.
Is this the same for every AI chatbot or model?
No. The published 10-50 mL/query range comes from analysis of a GPT-3-scale model; larger, more compute-intensive models (GPT-4-class and beyond) can use meaningfully more water and energy per query, and models run on more water-efficient hardware or in cooler climates can use less. Treat this calculator's output as an order-of-magnitude estimate for a typical mid-sized model, not a precise reading for any specific chatbot.
Why does AI water use matter if the water isn't destroyed?
Most data-center cooling water evaporates rather than being consumed and destroyed, but evaporation still removes it from the local water supply, competing with drinking water, agriculture, and ecosystems in that watershed — which matters more in already water-stressed regions where many data centers are built for cheap land and power. It's a real resource draw even though the water molecules themselves rejoin the broader water cycle eventually.
How can I reduce the water footprint of my AI use?
Sending fewer, more deliberate queries reduces the total directly, and where you have a choice, models and providers that publish efficiency or water-usage-effectiveness data tend to be more transparent about their footprint. Beyond that, the biggest lever is largely out of any individual user's hands — it's the data-center operator's choice of cooling technology, region, and grid mix that drives most of the variation this calculator's low/mid/high range reflects.