AI’s thirst is getting bigger, but not in the way most people assume. By 2030, AI data centres could use up to 600 billion gallons, or 2.27 billion m3, of water, largely because the electricity needed to run them will itself require more water to generate. That shifts the debate away from server-room cooling alone and straight into the power grid.
That matters because the industry has spent years talking up better cooling as the fix, while the real problem is now moving upstream. GPU power consumption nearly doubles with each new generation, so even if chip cooling improves, the wider energy appetite keeps pulling water demand along for the ride.
Cooling helps, but electricity is the bigger bill
The easy story is that data centres guzzle water to keep hardware from frying. The less convenient version is that most of the future water use will be indirect, coming from power plants rather than the racks themselves. In other words, you can make the servers cleaner and still end up with a bigger water footprint if the electricity mix stays dirty and power-hungry.
Tom’s Hardware says some sectors still use far more water than AI, and even more than all data centres combined. Fair enough. But AI infrastructure is scaling fast enough that the trend line is the part to watch, especially as larger GPU clusters keep pushing utilities harder.
Microsoft’s zero-water pitch has limits
Microsoft says its newest AI data centres use cooling systems efficient enough to operate with ”zero water” consumption. That sounds terrific for press releases and local planners, but the catch is obvious: closed-loop systems usually draw more energy than evaporative designs. Less water at the facility can still mean more water somewhere else.
That trade-off is why water and power are now fused into the same political problem. Public resistance to new data centres is already rising near residential areas, so developers are being forced to prove they can scale without turning themselves into both an electricity and water headache.
The short list of fixes is not very pretty
- Shift from evaporative cooling to closed-loop, direct-to-chip systems.
- Use mobile gas turbine units as a stopgap for data centre power needs.
- Build around renewables, nuclear power, and water-recovery systems for the long term.
The problem with the quick fixes is that they solve one headache by creating another. Mobile gas turbines may avoid heavy water use, but they bring ugly carbon emissions with them. The longer-term route is more sensible, but also slower, pricier and far less convenient for anyone trying to stand up AI capacity yesterday.
Xylem’s research points to a harder reality: by 2050, most indirect water use from data centres is expected to come from electricity generation. That makes the next phase of AI infrastructure less about clever chip cooling and more about whether the energy system can keep up without draining rivers, aquifers and public patience in the process.