“AI Uses X Litres of Water per Question”

Figures for the water cost of a single AI query circulate widely and differ from each other by more than an order of magnitude. That spread is not evidence that some of them are dishonest. It is what happens when a number is the product of four independent estimates, each of which can reasonably be made several ways.

Why published figures differ so much

There is no measurement of the water used by one query, and there cannot be. A datacenter does not meter per request. Every per-query figure in circulation is a division: some facility-level or grid-level quantity, divided by some estimate of how many requests that quantity supported. Both the numerator and the denominator are estimates, and both are contested.

This is the moving-denominator problem in its purest form, and it explains the whole spread without anyone having to be wrong. The useful response is not to pick a figure but to learn to read the assumptions behind one.

The calculation behind the number

Every such figure has this structure, whether or not it is shown:

energy_per_query   = energy_per_token x tokens_per_query x PUE
water_per_query    = energy_per_query x (WUE_onsite + WUE_offsite)

PUE   power usage effectiveness — facility overhead multiplier
WUE   water usage effectiveness — litres per kWh
      onsite:  evaporative cooling at the datacenter
      offsite: water consumed generating the electricity

Nothing here is exotic. PUE and WUE are standard industry metrics with published definitions, and large operators report them at facility or fleet level. The uncertainty is entirely in the inputs, and the two WUE terms behave very differently from each other.

Four choices that move it by orders of magnitude

• Which query. Energy per query is dominated by model size, active parameters and output length. A short answer from a small model and a long answer from a large reasoning model that spends thousands of tokens thinking before it replies differ by orders of magnitude. A figure quoted without saying which model and how many tokens is an average over a mixture nobody has described.
• Which datacenter. WUE varies enormously by cooling design and climate. Evaporative cooling consumes a great deal of water and comparatively little electricity; closed-loop and air-cooled designs consume very little water and more electricity. The same workload in two facilities can differ by more than the entire disputed range of published figures. The trade is real: cooling design is a choice between water and power, not a choice between good and bad.
• Which grid. Off-site water is consumed generating electricity, and that quantity depends on the generation mix. Thermoelectric generation consumes water for cooling; some renewables consume almost none; hydroelectric accounting is contested because reservoir evaporation can be attributed several ways. Two identical workloads on two grids differ substantially, and some published figures include this term while others count only on-site water.
• Training or inference, and how amortised. Training a model is a large one-off cost. Whether it is included, and over how many queries it is spread, changes the answer by a factor that depends entirely on a lifetime query count nobody knows in advance. The two compute budgets are different in size and in shape.

A figure that does not state all four is not comparable to another figure that does not state all four, even if both are honestly derived. This is the single most useful thing to know about the genre.

Withdrawal is not consumption

One distinction accounts for a large share of the confusion and it is not specific to AI. Water withdrawal is water taken from a source; water consumption is water that does not return to it, typically because it evaporated. A once-through cooling system may withdraw a great deal and consume very little. An evaporative system withdraws less and consumes most of what it takes.

Reporting standards differ on which is quoted, and a withdrawal figure compared against a consumption figure will show a difference that is purely definitional. Whenever two sources disagree by roughly an order of magnitude, this is worth checking before anything else.

The local dimension matters too, and it cuts against dismissiveness. Water is not fungible across regions. A quantity that is trivial against national supply can be significant against a specific catchment during a specific drought, which is why the meaningful questions are local — what does this facility consume, from which source, under what conditions — rather than global per-query averages.

What is genuinely true here

The dismissive reading is as wrong as the alarming one, so it is worth being explicit about what survives the scrutiny above.

• Datacenter water consumption is real, is growing, and is concentrated in specific places where it interacts with local supply. None of that depends on the per-query figure being right.
• Per-query figures are estimates built on estimates, and the honest form of one is a range with its assumptions stated.
• The comparison that actually informs a personal decision — this versus other daily activities — is available and is usually omitted from coverage in both directions, because it makes the number less striking.
• The lever that matters is not individual query volume. It is siting, cooling design and generation mix, all of which are decisions made by operators and regulators rather than by users.

If you want to hold a defensible position on this, hold the calculation rather than the number. The calculation stays correct when the inputs change, and the inputs change every year.