Ask most people what makes ChatGPT work and they will say software. The real answer is buildings — enormous windowless boxes stacked with graphics processors that draw so much electricity, and so much water for cooling, that they are now reshaping the grids and reservoirs of the places they land in.

This is not a niche subject. The datacentre footprint of AI is one of the fastest-growing categories of industrial demand for both electricity and water anywhere on the planet.

Where the energy actually goes

A datacentre’s electricity load splits roughly two ways. About half runs the servers — the actual computation. The rest cools them. GPUs designed for AI training generate substantial heat under load, and if the ambient temperature rises they slow down or fail, so cooling is not optional.

Two cooling systems dominate. Evaporative cooling uses water directly — enormous quantities of it — evaporating warm water so the air above it becomes cooler. It is efficient in terms of electricity but consumptive in terms of water. Closed-loop systems recirculate the same coolant and reject heat to the air; they use more electricity but very little water.

Which system a facility uses is largely determined by climate and by the price of water and electricity in that location. Which is why the operator’s choice of site matters as much as the technology inside it.

The scale

The International Energy Agency estimated global datacentre electricity consumption at around 415 terawatt-hours in 2024, roughly 1.5 per cent of world electricity demand, and projected it to more than double by 2030. AI is the fastest-growing segment within that.

In the United States, datacentres are projected to consume around 8 per cent of national electricity by 2030, up from about 4 per cent in 2024. Ireland already runs above 20 per cent — its grid effectively serves as a European hub for the industry, and the government has paused new connections in the Dublin area under pressure from the transmission operator.

The water figures are harder to pin down because reporting is patchy and companies dislike disclosing them. Published estimates suggest a single hyperscale facility using evaporative cooling can withdraw several hundred million litres of water per year — comparable to the consumption of a small town.

The training-versus-inference distinction

Two categories of AI compute exist and they behave very differently.

Training a large model happens once, over weeks or months, and consumes enormous power in a concentrated burst. Inference — actually answering user queries — is smaller per event but constant, and now accounts for the majority of ongoing AI energy demand as models are deployed at scale.

The industry’s own projection is that inference will dominate over the next decade. That makes efficiency gains in individual queries — which providers have made steadily — genuinely helpful, and also insufficient on their own, because query volume is growing faster than efficiency is improving.

Where the water comes from matters

A datacentre built in a wet, cool climate — Ireland, Norway, the Pacific Northwest — has different environmental effects than one built in Arizona, Chile or Spain. Water drawn from an aquifer in a water-stressed region is not equivalent to water drawn from a Nordic river, even at identical litres per year.

Yet several of the fastest-growing datacentre markets are precisely in water-stressed regions — because land is cheap, tax incentives are strong, and electricity is available. This is a genuine externalisation, and it produces exactly the local conflicts one would expect: municipalities negotiating datacentre developments while their reservoirs run low.

Why this connects to the wider story

The AI industry has argued, credibly, that it will be a driver of clean-energy investment because it needs so much power. Several of the largest firms have signed contracts for new nuclear reactors, renewable capacity and long-term power purchase agreements at scales that would not otherwise have happened.

That claim is not wrong, and it is not sufficient. New capacity takes years to build. Datacentres are being commissioned now. In the gap, the grid delivers whatever it has — which in most jurisdictions still includes substantial fossil generation. The near-term emissions impact of the AI buildout is real, and it has already begun.

Sources

International Energy Agency Electricity 2024 report and subsequent updates; Uptime Institute datacentre industry surveys; EirGrid statements on Dublin grid connections; published academic research on datacentre water consumption.