AI data centre water use has become a public concern as computing campuses grow larger. Servers do not drink water, but facilities may use it to remove heat, while the power stations supplying electricity can have their own water footprint.
The correct question is not whether every data centre uses a fixed amount. Cooling choices, climate, workload, electricity source and operating hours can produce very different results.
Where data centres withdraw and consume water
Many large facilities circulate water through closed loops inside the building. The final heat rejection may use cooling towers, which evaporate water, or dry coolers, which transfer heat to outside air. Evaporative systems can use less electricity in hot conditions but consume more water.
Air-cooled designs limit direct water withdrawal, although their electricity use may rise during extreme heat. Some campuses use reclaimed wastewater instead of drinking-water supplies.
Why AI changes the calculation
AI accelerators concentrate more heat per rack than traditional cloud servers. That encourages direct-to-chip liquid cooling. Liquid inside the facility does not necessarily mean high water consumption: a sealed loop can recirculate fluid, while the external heat-rejection system determines evaporative losses.
Workload timing also matters. Operators may shift some computing to cooler hours or locations when latency and grid conditions allow.
Microsoft’s reported water-use effectiveness and its limits
Microsoft says its average water-use effectiveness fell from 2.3 litres per kilowatt-hour in early data-centre generations to 0.27 litres per kilowatt-hour in 2025. The company reports a 25% reduction in water intensity toward a goal of 40% by 2030.
Those figures cover Microsoft’s owned fleet and use the company’s methodology. They do not describe every location, local peak demand or the water associated with electricity generation.
Air cooling and heat reuse
Google says new facilities in Texas and Sweden will use air cooling to limit water consumption. Its Swedish project is also designed for potential heat recovery, where warm output could support nearby buildings.
Heat reuse works best when a steady customer is close to the data centre. Long pipes, low-temperature heat and seasonal demand can weaken the economics.
How communities can judge a proposal
Useful disclosures include expected annual and peak water withdrawal, the source of that water, consumption rather than withdrawal, drought plans, cooling technology and projected water-use effectiveness. Developers should also explain how estimates change as computing capacity expands.
Local context is essential. A low-water facility in a stressed basin may matter more than a larger user in a water-abundant region. Transparent, site-specific reporting is more informative than a global average.
The real goal: more computing per constrained resource
Water, power and land are becoming design limits for AI infrastructure. Better chips and software can reduce the energy required for each useful output, while improved cooling can lower both water and electricity demand.
The industry’s challenge is to show absolute local impacts as well as efficiency gains. A facility can become more efficient while still using more total water if it grows quickly.
For a closer look at the cooling hardware, read how liquid cooling removes heat from AI servers.
Sources and further reading
- Microsoft: data-centre water-intensity progress and methodology
- Google: Texas data centre and air-cooling design
- Google: Sweden data centre, air cooling and heat-recovery readiness
Company sustainability figures are self-reported. This article is informational and does not provide investment advice.
Leave a Reply