The water it takes to answer you

When you type a question into an AI assistant and hit send, you’re not just consuming electricity. Somewhere in a data centre, most likely in Virginia, Iowa or Arizona, a cooling tower is evaporating fresh water to carry heat away from the GPUs processing your request. That water doesn’t come back. It rises as steam and joins the atmosphere, permanently removed from the local water supply.

The per-query numbers sound small at first. Running a GPT-class inference for 10 to 50 queries consumes around 500 millilitres of water^[1]OECD.AI / Li et al., UC RiversideView source →, accounting for both direct on-site cooling and the indirect water used to generate the electricity powering those servers. That works out to roughly 10–50 ml per response, depending on answer length and facility efficiency.

What counts as water used?

Not all of an AI system’s water shows up in the same place. Researchers split it into three scopes, and the indirect, off-site water usually dwarfs what a data centre draws at the tap.

The trouble is the scale

The per-query numbers sound manageable in isolation. The problem is volume. ChatGPT alone processes an estimated 2.5 billion prompts per day^[2]UN University INWEHView source →. At even a conservative 10 ml per query, that is 25 million litres evaporated daily, from a single product. And AI usage is still growing rapidly.

Once a model is deployed, inference, not training, drives the vast majority of ongoing energy and water use. Estimates suggest 80 to 90 per cent of a model’s total energy draw comes from serving user queries^[2]UN University INWEHView source → post-deployment, not from the original training run. Training GPT-3 consumed an estimated 700,000 litres of freshwater on-site. But that number is dwarfed by the cumulative drip of inference at scale.

Location is everything

The same query can have a wildly different water footprint depending on where it is processed. Research suggests a 10,000-fold difference between the most and least efficient solutions^[6]ImpakterView source →. You have no way of knowing which facility answered your query.

The same query, answered in two places, can differ ten-thousand-fold in the water it costs.

The aggregate becomes geopolitical

By 2027, global AI infrastructure is projected to consume between 4.2 and 6.6 billion cubic metres of water per year^[7]Great Andhra / Li et al.View source →, nearly half the United Kingdom’s total annual water withdrawal. In 2025 alone, AI data centres consumed nearly 1 trillion litres^[3]Mordor Intelligence / BarchartView source →, roughly the annual water use of 1.8 million Americans.

The communities living closest to these facilities, whose aquifers and reservoirs are being drawn down, are often not the ones benefiting from the AI running inside them. In Uruguay, plans for a water-intensive data centre coincided with a 2023 drought that depleted Montevideo’s freshwater reserves and made tap water unsafe to drink^[2]UN University INWEHView source →.

What the industry is doing

The clearest lever is cooling technology. Moving from evaporative towers to closed-loop, direct-to-chip cooling can dramatically cut on-site water use^[8]Tom’s HardwareView source →. Microsoft’s newest AI data centres are designed to operate with near-zero water consumption. Google reports replenishing 64 per cent of its freshwater withdrawals via watershed restoration, targeting 120 per cent by 2030^[9]Google / HowMuchWaterDidIWaste.comView source →.

But offsets are not the same as not using the water in the first place. The uncomfortable truth is that the environmental cost of an AI query is not abstract. It is a real volume of freshwater, taken from a real place. The question worth asking, and the one the industry has mostly avoided, is whether the answer was worth it.