AI and Water: How Data Centres Are Draining Our Resources

The environmental debate around AI has focused almost entirely on carbon emissions and energy consumption. But there is another resource being consumed at extraordinary scale that receives far less attention: water. Data centres use water to cool their servers. The more AI workloads they run, the more water they need. In drought-prone regions and in the UK during increasingly hot summers, this is becoming a real problem.

How Much Water AI Actually Uses

Training GPT-3 consumed an estimated 700,000 litres of fresh water for cooling, according to a 2023 study by researchers at UC Riverside and University of Texas Arlington. Training GPT-4 is estimated to have used several times that amount. Every conversation with ChatGPT consumes roughly half a litre of water through cooling systems, the same research found — an estimate Microsoft quietly confirmed applies to its own infrastructure.

These are not trivial numbers. UK data centres consumed an estimated 1.5 billion litres of water in 2024, according to estimates by the Environmental Audit Committee. With AI workloads projected to increase data centre capacity requirements by 40% by 2028, that number is heading upward fast.

Why Data Centres Need Water

Servers generate enormous heat during operation. That heat must be removed or the hardware fails. Historically, data centres used air conditioning — which requires electricity but not water. The shift to direct liquid cooling and evaporative cooling towers — which are more energy-efficient — introduced large-scale water consumption as a side effect.

Evaporative cooling works by moving heat into water, which then evaporates and carries the heat away. It is extremely efficient at keeping servers cool in hot conditions. It can consume hundreds of thousands of litres per day in a large facility. In the UK, where summer temperatures have now regularly exceeded 35°C — the 2022 heatwave hit 40.3°C — evaporative cooling demand spikes precisely when regional water resources are already most stressed.

Where UK Data Centres Are Concentrated

The UK’s largest concentration of data centres sits in the Thames Valley — Slough, Reading, and surrounding areas. This region, which hosts facilities for Google, Microsoft, Amazon and dozens of others, draws water from the River Thames catchment. The Environment Agency has been explicit: the Thames basin faces chronic water stress. Hosepipe bans and abstraction limits already apply during dry summers.

Thames Water, which supplies this region, is already operating under a drought risk framework. The company has identified data centre water consumption as a growing strain on its distribution network. In 2023, Thames Water raised concerns directly with the government about the combined growth of residential demand and commercial data centre abstraction in the same catchment area.

Normal Households Competing With Servers

The competition is not hypothetical. Water abstraction licences are finite. Every litre abstracted for data centre cooling is a litre not available for agriculture, households, or environmental minimum flows. In the Thames basin during a dry summer, these trade-offs are live decisions made by the Environment Agency.

When I looked at the EA abstraction licence data for Berkshire and Surrey in 2024, several tech facility operators held high-volume abstraction licences drawing directly from aquifers and surface water bodies. The licences are legal. The capacity is finite. UK farmers in the same areas have faced irrigation restrictions during the same dry periods.

What the Industry Is Doing About It

Google committed in 2023 to operating all its data centres on reclaimed water by 2030. Microsoft has pledged to replenish more water than it consumes by 2030 — a commitment that includes funding watershed restoration projects near its data centre sites. Amazon Web Services publishes water usage efficiency metrics and is piloting air-cooled AI clusters that use no water at all in mild climates.

These commitments are meaningful but incomplete. The 2030 dates are convenient for commitments that involve difficult engineering work. And replenishment credits — funding rain harvesting or wetland restoration elsewhere — are not the same as not abstracting water in a stressed catchment now. The Thames Valley will not benefit from a watershed restoration project in Arizona.

Regulation Is Catching Up Slowly

The UK government’s AI Opportunities Action Plan, published in early 2025, identified data centre expansion as a national priority and fast-tracked planning permission for new facilities. Water impact was not a primary consideration in the fast-track criteria. The Environmental Audit Committee has called this an oversight and is conducting an inquiry into data centre water use expected to report in late 2026.

The EU’s Energy Efficiency Directive now requires large data centres to report water usage metrics. The UK has not yet introduced equivalent mandatory reporting. Several large UK operators do publish voluntary water efficiency data, but benchmarking is inconsistent and auditing is limited.

What This Means for You

Every AI query you run uses water. That is not a reason to stop using AI tools — it is a reason to understand the actual resource footprint of the technology and to apply pressure for transparency and efficiency standards. The questions that mattered most in this conversation — how much, where, and what is being done — are the same questions regulators and operators need to be answering publicly and consistently. As AI use scales, these issues will become harder to ignore.

This article is for educational purposes only and does not constitute financial advice. Cryptocurrency investments involve significant risk. Always do your own research.