—per request
—on-site cooling
—off-site generation
What that volume of water looks like
Equivalents use 500 ml per bottle, 65 L for an eight-minute shower, 2 L of drinking water per person per day and 2,500,000 L for an Olympic pool. They are there for scale, not to imply the water is interchangeable — cooling water and drinking water are drawn from different systems and returned differently.
Same workload, different region
Cooling design and climate move the on-site half by more than an order of magnitude. The grid behind the plug moves the off-site half. Neither alone tells the story.
| Region / cooling | On-site | Off-site | Total / month | ml / request |
|---|
Reference WUE and grid water-intensity ranges from public disclosures; individual sites vary by season and methodology. Click a row's values into the fields above by editing them directly.
Energy is the input; water is one of the outputs
Every litre on this page starts as a kilowatt-hour, which is why the fastest way to cut a water footprint is to cut tokens rather than to argue about cooling. A shorter system prompt, a smaller model for the easy 80% of traffic, and a cache in front of repeated questions each remove energy before it is ever drawn. Price the same traffic in kilowatt-hours and CO₂e on the AI inference energy & carbon calculator, see what the tokens themselves cost on the LLM token cost calculator, model routing the cheap traffic away on the LLM cascade savings calculator, and check how much repeat traffic a cache would absorb with the semantic cache ROI calculator.
How this calculator works
The AI Water Footprint Calculator works forward from energy. Your requests per month multiplied by energy per request gives the IT load in kilowatt-hours; multiplying by PUE adds the cooling, power conversion and lighting overhead to get facility energy. That facility energy is then converted into water twice. On-site water is facility kilowatt-hours times WUE, the litres evaporated per kilowatt-hour in the building's cooling system. Off-site water is the same facility energy times the grid water intensity, the litres consumed per kilowatt-hour at the power stations that supplied it. Adding the two gives the all-in figure, and dividing by requests gives the millilitres-per-prompt number the headlines quote.
The input that moves the result most is energy per request, which varies by more than a hundredfold between a small model answering in a sentence and a reasoning model thinking for thousands of tokens — so the model-class preset is the first thing to get right. After that, region matters: free-cooled northern sites report WUE an order of magnitude below evaporatively cooled sites in hot dry climates, which is what the region table quantifies. Off-site water usually exceeds on-site water on a thermal-heavy grid, which is why reporting only one of them can make the same workload look either negligible or alarming. All values are reference estimates from public disclosures and vary by site, season and methodology.
How much water does one AI prompt actually use?
Published figures for a single prompt range from a fraction of a millilitre to more than half a litre, and almost all of the disagreement comes from what is being counted rather than from the underlying hardware. The smallest numbers count only water evaporated in the data centre's own cooling loop, on an efficient modern site, for a small model answering a short question. The largest numbers count a long answer from a frontier model plus the water consumed generating the electricity at power stations, which is usually several times bigger than the cooling water itself. Both can be defensible and neither is the whole picture on its own, which is why this calculator splits the two and lets you set each independently. For a mainstream model at typical answer lengths on a reasonably efficient site, the all-in figure usually lands in the low single-digit millilitres per request — meaningful at the scale of millions of requests a month, and small compared with almost anything else a person does in a day.
What is the difference between on-site and off-site water?
On-site water is what the building itself consumes, mostly evaporated in cooling towers to reject the heat that servers produce. It is measured as water usage effectiveness, or WUE, in litres per kilowatt-hour of facility energy, and it varies enormously with climate and cooling design: a free-cooled site in a cold climate can be near zero, an evaporatively cooled site in a hot, dry region can be well above one litre per kilowatt-hour, and fully closed-loop or air-cooled designs consume essentially none on site while using more electricity. Off-site water is consumed somewhere else entirely, at the power stations that produced the electricity, through evaporation in thermoelectric cooling and from reservoirs. It scales with how the local grid generates power and is often the larger of the two. Reporting only on-site water makes an air-cooled data centre look free; reporting only off-site water ignores what the building does to local supply. The honest answer needs both, and the two respond to completely different decisions.
Does running AI in a colder region really cut the water footprint?
It cuts the on-site half dramatically and the off-site half only if the grid also happens to be different. Cool, humid climates allow free cooling for most of the year, which is why sites in Ireland and the Nordics report water usage effectiveness an order of magnitude below the global average, while sites in hot dry regions lean on evaporative cooling precisely when local water is scarcest. That is a real and large difference, and it is the single biggest lever a provider controls. But the electricity behind the workload still carries its own water cost wherever it is generated, and that depends on the generating mix rather than the weather: a grid heavy in thermal generation consumes several litres per kilowatt-hour regardless of how cold the data centre is. So region choice can plausibly cut the total by half or more, but it cannot take it to zero, and the region table on this page is there to show the size of the effect rather than to suggest the problem disappears.
How accurate are these numbers?
They are a transparent model, not a measurement, and every input is exposed precisely so you can see what drives the answer. Providers publish very little per-request energy data, so the energy-per-request figure is the weakest link and the one worth stress-testing: a reasoning model that thinks for thousands of tokens can use an order of magnitude more energy than a small model answering in one line, and the default presets here reflect broad published ranges rather than any specific vendor's measurement. Water usage effectiveness and grid water intensity are better documented but still vary by site, by season and by methodology, and providers report them inconsistently. Use this to understand the shape of the problem — which term dominates, how much region choice matters, what happens when volume grows tenfold — and treat any single headline figure, including the ones this page produces, with suspicion. If you need a number for a report, state your assumptions alongside it.