The Hidden Water Cost of Every AI Prompt
Every answer from an AI chatbot is generated on a server in a data centre, and data centres get hot. Shedding that heat is one of the largest ongoing costs of running AI at scale, and a good deal of it is done with water.
Where the water goes
Many large data centres use evaporative cooling: water absorbs heat and carries it away as it evaporates, the same mechanism by which sweating cools a body. It is efficient in energy terms, and it consumes water rather than merely circulating it.
There is a second, less visible cost: the electricity a data centre draws was generated somewhere, and thermal power generation uses water for cooling too. So a query has a direct footprint and an indirect one.
What the research actually says
This is where honesty matters more than a memorable number.
Published estimates vary by more than an order of magnitude, and they are not measuring the same thing. The figures depend on the model's size, the hardware, the cooling design, the local climate, and whether the estimate counts on-site water or the power generation too.
Widely cited work from the University of California, Riverside put a short exchange with a large model in the region of half a litre; some operators have since published far lower per-query figures for their own systems. Both can be defensible, because they describe different systems and different accounting.
The honest summary is that a single query's water cost is small and real, that published figures disagree, and that the aggregate across billions of queries is what makes it worth discussing at all. Anyone quoting one precise number with confidence is overstating what is known.
Where the local part matters most
A litre of water where there is plenty is not the same as a litre in a drought-prone region, and data centres are sited for cheap power and cool air rather than water availability. Several have drawn genuine local objection on that basis.
Which is the part that generalises badly. The global average matters far less than where a particular facility sits.
Small tasks do not need a data centre
Here is the practical point, and it is the one worth acting on.
Counting words, converting a measurement, generating a password, formatting some JSON - none of these needs a large language model. They have one correct answer, and a purpose-built tool does them with a fraction of the computation.
Most tools here run entirely in your browser, on your own device, with no server doing the work at all. The Word Counter counts on your machine; so does the Unix Timestamp Converter and nearly everything else.
None of which is an argument against AI for what it is genuinely good at - open-ended writing, research, judgement. It is an argument for matching the tool to the task, which happens to be the greener option and the faster one.
Common questions
Is it worth worrying about one prompt?
No single prompt matters. The aggregate does, which is a different claim.
Does this apply equally to all AI?
No. Larger models cost more per query than small ones, and generating an image or a video costs considerably more than generating text.
Are the companies doing anything?
Several have published water targets and are moving toward air cooling and recycled water. Progress is real and uneven.
Want the answer without a data centre involved?