AI Water Footprint Calculator
Estimate the freshwater consumed by your use of large language models and generative AI, per prompt, per image, per year and per training run.
💧 What is an AI Water Footprint?
The AI water footprint is the volume of freshwater consumed, meaning evaporated and lost to the local basin, in order to run a generative AI workload. It has two distinct parts that are easy to confuse. The first is onsite water, the water that boils away in a data centre's cooling towers and evaporative chillers while it dumps server heat into the air. The second is offsite water, the far larger volume that evaporates at the power stations feeding electricity into that data centre, from thermoelectric cooling at gas, coal and nuclear plants and from the surface of hydroelectric reservoirs. A single chat prompt touches both, and for most grids the offsite share dominates.
The reason this became a public issue is scale. Individually the numbers are trivial: a typical answer from a frontier model consumes something on the order of a teaspoon of water. But a service handling a billion prompts a day multiplies that teaspoon into thousands of cubic metres daily, and those cubic metres are drawn from specific aquifers and rivers next to specific towns. Data centre clusters concentrate in places like central Oregon, northern Virginia, Arizona and Ireland, so a globally small number becomes a locally significant one. Journalists, regulators, municipal water utilities and corporate sustainability teams all now need a defensible way to convert AI usage into litres.
Three misconceptions get in the way. The first is confusing withdrawal with consumption. A once-through cooled power plant withdraws enormous volumes and returns almost all of it warmer; consumption counts only what evaporates, and it is the consumption figure that stresses a basin. The second is assuming the data centre is where the water goes. On a typical grid, offsite generation water is roughly five to ten times the onsite cooling water, so switching a workload to a wind and solar heavy region does more than any cooling upgrade. The third is treating a single published number as authoritative. Energy per query varies by more than a factor of ten between a small quantised model on efficient hardware and a large frontier model answering at length, and operators publish very little.
This calculator makes those variables explicit rather than hiding them behind one headline number. You set the usage, the response length, the model size class, the site cooling efficiency, the power overhead and the grid water intensity, and it reports water per prompt, per image or per training run, together with the onsite and offsite split and equivalents in bottles, showers and cups of coffee. Every output is an estimate with wide uncertainty, and the tool is designed so you can see exactly which assumption is driving your answer.
📐 Formula
📖 How to Use This Calculator
Steps
💡 Example Calculations
Example 1 — A Daily Chatbot Habit on a Frontier Model
20 prompts per day, 800 token answers, large frontier model at 2.9 Wh per 1000 tokens, global average site
Example 2 — A Heavy User on a Small Self-Hosted Model
500 short prompts per day, 300 token answers, 7B model at 0.13 Wh per 1000 tokens, low water site on a clean grid
Example 3 — A Designer Generating Images in an Arid Region
50 images per day at 2.9 Wh each, US West site with evaporative cooling (WUE 0.55, EWIF 2.40)
Example 4 — Training a GPT-3 Class Model
1287 MWh of training electricity, United States average grid (WUE 0.35, PUE 1.20, EWIF 3.14)
Example 5 — A Monthly Work Allowance
600 prompts per month, 800 token answers, frontier model, global average site
❓ Frequently Asked Questions
🔗 Related Calculators
How much water does one ChatGPT prompt use?
On this calculator's defaults, a large frontier model producing an 800 token answer at 2.9 Wh per 1000 tokens on a global average data centre consumes about 7.10 mL of freshwater per prompt. Published estimates range from under 1 mL for small efficient models to roughly 50 mL for long answers on inefficient hardware in a hot, dry region.
What is WUE in a data centre?
Water use effectiveness is the litres of water a facility consumes per kilowatt hour of IT energy. Hyperscale operators report annual figures between about 0.1 and 0.5 L/kWh. A site with closed loop or free air cooling can be near zero, while an evaporative cooled site in an arid climate can exceed 1.5 L/kWh in summer.
What is the difference between onsite and offsite AI water use?
Onsite water is what evaporates in the cooling towers and humidifiers of the data centre itself, set by WUE. Offsite water is consumed at the power stations that generate the electricity, set by the grid electricity water intensity factor. Offsite water is usually the larger share, often 85 to 95 percent of the total.
What is EWIF and what value should I use?
The electricity water intensity factor is the litres of freshwater consumed per kilowatt hour of grid electricity, counting evaporation from thermoelectric cooling and hydro reservoirs. National averages sit between about 1.2 and 3.2 L/kWh. Use around 3.1 for the United States average, 2.3 for a global average and under 1.5 for a wind and solar heavy grid.
How much water did it take to train GPT-3?
Li et al. estimated roughly 700,000 litres of onsite water for GPT-3 training in Microsoft's US data centres. At 1287 MWh of training energy on the United States grid, the same model gives about 450 cubic metres onsite and 5.30 megalitres once the offsite generation water is included, which shows how much the answer depends on where you draw the boundary.
Does asking for shorter answers actually save water?
Yes. Inference energy scales close to linearly with the number of tokens the model generates, so halving the response length roughly halves the water. Going from a 2000 token essay to a 300 token summary on a large model cuts the per prompt water by about 85 percent.
How much water does generating one AI image use?
At 2.9 Wh per image on an arid region site with a water intensity of 3.43 L/kWh, one generated image consumes about 9.95 mL. Lighter diffusion models running at 0.5 Wh per image land nearer 1.7 mL, and high resolution or multi-step pipelines at 6 Wh per image reach about 20 mL.
Is the water consumed or just borrowed?
This calculator reports consumptive use, water that evaporates and leaves the local basin rather than returning to it. Withdrawal, the water drawn in and largely returned, is far larger. The consumption figure is the one that matters for local water stress, which is why it is the standard reported quantity.
Why is PUE multiplied by EWIF but not by WUE?
WUE is defined per kilowatt hour of IT equipment energy, so it already accounts for the whole facility's cooling water against the IT load. EWIF is defined per kilowatt hour drawn from the grid, and the grid supplies the full facility load, which is PUE times the IT load. Multiplying only the offsite term by PUE keeps both definitions consistent.
How accurate are these AI water estimates?
Treat them as order of magnitude. The three biggest uncertainties are energy per query, which varies by more than tenfold between models and batching strategies, the site WUE, which operators publish only as annual averages, and EWIF, which changes hour by hour with the generation mix. A result of 7 mL should be read as roughly 2 to 25 mL.
Does running AI locally on my own machine use less water?
Usually far less, because a laptop or desktop uses no evaporative cooling and consumes only the offsite generation water for a few watt hours. A 7 billion parameter model producing 300 tokens on a laptop uses roughly 0.06 mL of water at a clean grid intensity, over a hundred times less than a frontier model on an arid region data centre.
Can data centres run without consuming water?
Some can. Air cooled and closed loop liquid cooled designs trade water for electricity, pushing WUE close to zero while raising PUE slightly. That is a net win for water in a water stressed basin, but if the extra electricity comes from a thermoelectric grid the offsite water can rise by more than the onsite saving.