The cloud has a physical address, and increasingly, a water bill. The data centres that train and run artificial intelligence models are among the fastest-growing industrial consumers of electricity and water in the world: a single large facility can draw as much power as a mid-sized city and, where cooling uses evaporation, drink millions of litres of water a day. As AI construction accelerates across the United States, Europe and Asia, planners and communities are asking questions the industry answered casually for years — starting with how much water, from which river, in which drought.
The scale of the build-out explains the scrutiny. The International Energy Agency tracks data-centre electricity demand roughly doubling across this decade under AI-driven growth, with the United States, China and Europe hosting the great majority of capacity. Water use has been far less transparent: historically few operators published it, definitions varied, and the largest share — the water consumed by the power stations supplying the electricity — often sat outside company reporting altogether. Our front page has covered the planning fights this opacity produces; the numbers underneath are now finally improving.
Where a data centre’s water actually goes
Direct water use is mostly about heat. Servers convert nearly all their electricity into heat that must leave the building. Evaporative cooling towers reject that heat cheaply by evaporating water — efficient in energy terms, thirsty in water terms. Alternatives exist on a spectrum: air cooling uses no water but more electricity in hot climates; closed-loop liquid cooling, increasingly standard for AI hardware, recirculates coolant and slashes consumption; and the newest designs pipe heat into district heating networks, turning a waste product into a civic asset. The right choice depends on local climate and scarcity, which is precisely why campaigners argue siting decisions should weigh water stress as heavily as land price and fibre routes.
Indirect use — water evaporated by power generation — can exceed direct use, particularly where grids rely on thermal plants. This is why two identical facilities can have radically different total footprints, and why renewable electricity procurement reduces the water ledger as well as the carbon one. Industry metrics such as water usage effectiveness exist, but inconsistent boundaries make comparisons treacherous; regulators in Europe now require standardised reporting of energy and water indicators for larger facilities, a model other regions are studying for their own permitting rules — developments our technology coverage follows as they spread.
Communities learn to negotiate
The politics have matured quickly. Early data centres were welcomed anywhere for their construction jobs and tax base, and permitted almost without conditions. Drought-stricken regions have changed the terms: moratoria have been debated or imposed in parts of the Netherlands, Ireland and the American Southwest; new approvals increasingly carry conditions on water sourcing, reuse of treated wastewater, heat recycling and genuine additionality in renewable supply. Operators, needing social licence for projects that run for decades, have begun publishing site-level water data and funding watershed restoration — in the best cases, returning more clean water to stressed basins than their facilities consume.
Technology is helping at the margin that matters. AI itself now optimises cooling in real time, cutting energy overheads that once ran at forty per cent of a facility’s draw towards ten. Immersion cooling and direct-to-chip liquid systems designed for AI racks reduce both water and energy intensity per computation. The uncomfortable arithmetic, acknowledged by the IEA and operators alike, is that efficiency per task improves while total tasks explode — so absolute footprints depend on planning choices, not just engineering.
The planning question for the AI decade
Data centres are becoming critical infrastructure, as essential to a modern economy as ports or power stations, and they are starting to be planned like it: clustered where clean power and water are genuinely abundant, integrated with heat networks, and assessed on full resource footprints rather than headline investment alone. The regions that get this right will host the industry durably. Those that waved through thirsty facilities in dry places are already discovering that the water bill, like the truth about any resource, arrives eventually — usually during a drought, and always in public.
