Data centres, AI and water
Data centres use water for cooling. The widely circulated per-query figures are scenario outputs, not measurements — and the largest water use is usually the one nobody counts.
Data centre water use has become a prominent public issue, and the reporting on it is unusually poor — not because the concern is invalid, but because the figures circulated almost never state their assumptions. This page sets out what is actually known and what the honest uncertainties are.
Why data centres use water at all
Servers convert essentially all the electricity they draw into heat, which has to be removed. There are three broad approaches, and they trade water against energy.
| Row | Evaporative cooling | Air cooling | Liquid / direct-to-chip |
|---|---|---|---|
| How it works | Evaporates water to remove heat | Uses fans and chillers, no evaporation | Circulates coolant directly to the chip |
| On-site water use | High — this is the water people are counting | Near zero | Low; often closed loop |
| Energy use | Lower | Higher — more chiller load | Lower, and enables higher density |
| Indirect water via electricity | Lower | Higher | Lower |
| Best suited to | Hot dry climates, where it is most effective — and where water is most scarce | Cool climates and water-stressed sites | High-density compute including AI training |
What WUE measures
Water usage effectiveness is litres of water consumed per kilowatt-hour of IT energy delivered. Evaporatively cooled facilities typically report somewhere in the region of 1 to 2 l/kWh; air-cooled facilities report far lower. It is a useful metric with one large limitation: it covers on-site water only, so it cannot capture the trade-off above.
On per-query water figures
Widely circulated per-query figures derive from academic estimates that combine assumed facility WUE, assumed grid water intensity, assumed model size and assumed hardware utilisation. Change any one of those and the answer moves by an order of magnitude. The original papers are careful about this; the figures that circulate from them usually are not. There is no single per-query water figure, because the answer depends on which facility, which climate, which grid, which model and whether indirect water is counted.
- Facility WUE varies by more than a factor of ten between climates and cooling designs
- Grid water intensity varies by more than an order of magnitude between electricity systems
- Estimates differ on whether training is amortised across queries or excluded entirely
- Hardware efficiency changes materially between generations, so figures date quickly
None of this means the concern is unfounded. Data centres are large, individually visible, growing quickly, and frequently sited in water-stressed regions for reasons — cheap land, cheap power, network proximity — that have nothing to do with water availability. That is a legitimate and specific planning issue. It is better argued from facility-level withdrawal in a named basin than from a per-query number that cannot be defended.
How to assess a specific facility
- Ask for withdrawal and consumption separately — they can differ by a factor of several
- Ask what the water source is: potable municipal supply, reclaimed water, or non-potable industrial supply. Using reclaimed water in a stressed basin is materially different from using drinking water
- Ask about the local basin, not the country. National water stress figures say nothing about whether a specific aquifer can support a facility
- Ask whether indirect water from electricity is included. It usually is not
- Check the disclosure basis: company reported, third-party estimated, from a planning document, or inferred
Hydrionis records the disclosure basis on every data centre record for exactly this reason, and does not mix reported and estimated values in the same figure.
Sources
- Individual corporations — Published corporate sustainability and water reports. Public but restricted · Publisher terms — public access, reuse not clearly granted
- Peer-reviewed scientific literature — Open-access hydrology, water chemistry and water use research. Public but restricted · Publisher terms — public access, reuse not clearly granted
- International Energy Agency — Water–energy nexus analysis. Public but restricted · Publisher terms — public access, reuse not clearly granted