The Hidden Bill Behind Every AI Answer

One conversation with an AI system can use as little as 10 to 25 millilitres of water for cooling the servers that run it. Generating a single image can multiply that figure up to thirty times over. Multiply that by billions of daily interactions worldwide, and the number stops being a rounding error.

Global data centre electricity consumption reached about 415 terawatt-hours in 2024. The International Energy Agency projects it will climb to roughly 945 terawatt-hours by 2030 — nearly triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria, three countries home to more than 650 million people combined. If data centres were a country today, they would already rank as the world's fifth-largest electricity consumer, between Japan and Russia.

Where the strain is already visible

This is not a distant, theoretical problem. In Ireland, data centres already consume more than 20 percent of the country's metered electricity — enough that the national grid operator has effectively paused approvals for new facilities. A standard, medium-sized data centre can use roughly 110 million gallons of water annually, comparable to the yearly needs of a town of 50,000 people. Google has reported its global data centre water use reaching 8.1 billion gallons.

Engineers face a genuine trade-off, not a simple fix: evaporative cooling is energy-efficient but consumes enormous volumes of water; air-cooled systems save water but demand far more electricity to run. There is no cooling method that avoids both costs at once — only a choice of which resource to spend more of, and where.

Two ways to read the same numbers

One view treats this as a serious, underappreciated externality: the benefits of AI are distributed globally and abstractly, while the resource cost — strained local grids, competition for freshwater, community pushback — lands concretely on specific towns and regions that often see little direct benefit from the technology built on their doorstep. Under this view, the current pace of expansion is outrunning the infrastructure and oversight needed to manage it responsibly.

A different view, argued by researchers at institutions like ITIF, cautions against treating raw consumption figures as the problem in themselves. Data centres, they point out, are not the primary driver of rising global electricity demand overall — and electricity use only becomes a genuine policy problem when it produces a concrete harm: higher household bills, a less reliable grid, or measurable environmental damage. Under this view, the right response is better measurement of impact per unit of useful work, not blanket alarm over scale.

Both views can be true at once in different places. Ireland's grid strain is a documented, local reality. Whether that generalizes into a global crisis, or remains a set of specific bottlenecks to be managed case by case, is exactly the kind of question that gets decided by people who rarely have to answer for it in public.

What would you ask?

• Who should be accountable for a data centre's local water and electricity impact — the AI company, the cloud provider, the local utility, or the regulator who approved it?
• Should companies be required to disclose the resource cost of an AI service the way they disclose ingredients or emissions, so users can actually see what a query costs?
• Is it fair for communities to bear the local strain of a technology whose benefits are captured mostly elsewhere?
• Are efficiency gains in cooling technology keeping pace with the growth in demand, or merely slowing the rate at which the problem gets worse?
• Other — tell us in the comments 👇

Is this a question that matters to you? You can put it directly to the people making these decisions at Cores.Media.