Every AI answer has a water bill. Here is who pays it.

Ask an AI model a question and the reply lands in under a second, apparently out of thin air. But the servers that produced it ran hot, and cooling them took water. Generating their electricity took more. And the chips inside were rinsed in ultrapure water before they ever shipped. The water cost of AI is real, it is measurable, and for a water-stressed country like the UK it is arriving faster than the plans to manage it.

The drop you cannot see

In August 2025, Google published the first detailed figure from a major AI provider: the median Gemini text prompt consumes about 0.26 millilitres of water, or roughly five drops, alongside 0.24 watt-hours of energy. Five drops. You could run a thousand prompts and not fill a shot glass.

That number is real, but it is also the narrowest honest way to state it, and two years earlier researchers at UC Riverside had arrived somewhere very different. Their study, Making AI Less Thirsty, estimated that an older model, GPT-3, “drinks” a 500 ml bottle of water for every 10 to 50 responses. Same technology, wildly different headline. So which is right?

*500 ml per 10 to 50 replies works out to roughly 10 to 50 ml each. The gap between the two studies is not dishonesty, it is scope, model age and location. Cooling a data centre in a hot, dry region on a summer afternoon can cost many times more water than the same job at night in a cool climate.

Three ways a data centre drinks

Water enters the picture at three points. Only the first is fully in the operator’s hands, which is exactly why the other two get quietly left out of most “per-prompt” claims.

Small drop, staggering scale

The per-prompt number stays small. The number of prompts does not. Multiply five drops by a planet’s worth of daily AI use and the totals stop being abstract.

Why this is a UK problem, right now

It would be easy to file this under “American problem”. It is not. England is already forecast to be short of water, and the AI build-out is landing squarely on the thirstiest ground.

Environment Agency, reaffirmed December 2024, via the UK Government report.

Into that gap the country is pouring a £14 billion commitment to large data centres and a set of new “AI Growth Zones”. The catch, flagged by the Environment Agency itself, is that the growth-zone policy makes no mention of water at all, and data centres sit outside the national water resources framework. The plumbing is being scaled up while the water plan looks the other way.

The pressure is concentrated. Under a high-growth scenario, one regional supplier estimated that data centres alone could drive almost 30% of all new water demand, around 270 million litres a day. That is a single new industry, quietly becoming one of the largest draws on a stressed network.

The part almost nobody connects

Here is the bridge back to the ground floor, where estates and facilities teams live. Those evaporative cooling systems are not exotic. They are wet cooling towers, the same category of equipment that sits on the roofs of hospitals, plants and large commercial buildings across the country. And in the UK, a wet cooling tower is a notifiable Legionella risk, governed by the HSE’s Approved Code of Practice L8 and the technical guidance in HSG274 Part 1.

So a data centre’s cooling loop is three problems wearing one coat: a consumption problem, a leakage problem, and a water-quality and compliance problem. The organisations that will come out of the AI boom looking competent are the ones already measuring all three, instead of discovering them during a drought or an audit.

What “measuring it” actually looks like

You cannot manage what you cannot see, and most water systems are close to invisible. This is the whole point of smart water monitoring, and it maps neatly onto the three costs above.

Frequently asked