Summary
- Schneider compared modelled 100MW air- and liquid-cooled AI facilities using adiabatic heat rejection.
- In the Paris case, annual on-site cooling water fell from about 108,000m³ to 51,000m³.
- The figures are modelling results, not measured performance from an operating data centre.
Schneider Electric has modelled a 53% reduction in annual on-site cooling water use for a 100MW AI data centre in Paris by changing the cooling architecture and operating conditions around high-density compute.
The company’s White Paper 220 compares air-cooled and optimised liquid-cooled AI designs using adiabatic heat rejection. In the Paris model, annual cooling water falls from approximately 108,000 cubic metres to 51,000 cubic metres.
A corresponding Dallas model shows a 48% reduction, from around 382,000 to 197,000 cubic metres. The geographic difference illustrates how climate affects the same basic cooling strategy.
The numbers are modelled scenarios rather than meter readings from operating 100MW sites. That distinction is important because actual water consumption depends on workload, weather, cooling-water temperatures, control settings, equipment loading, and how frequently different heat-rejection modes operate.
Even so, the analysis demonstrates how facility design can shift the water consequences of dense AI deployment before a site reaches construction.
Liquid cooling moves heat, it does not eliminate it
Direct liquid cooling is often described as a response to rack density, but installing cold plates or liquid loops at the IT equipment is only one part of the thermal chain. The captured heat still has to be moved away from the rack and ultimately rejected or reused.
The facility can do that through dry coolers, cooling towers, adiabatic systems, heat exchangers, or combinations of those technologies. Each has different implications for electricity use, water consumption, ambient-temperature limits, and capital cost.
Operating temperature also matters. Higher allowable coolant temperatures can increase the number of hours in which a facility rejects heat without mechanical refrigeration or heavy evaporative assistance. That can reduce both electricity and water requirements, depending on local conditions.
Schneider’s study therefore treats water consumption as an architectural result rather than an unavoidable fixed quantity attached to a megawatt of AI capacity.
That is increasingly relevant for planning. Water demand has become a contested issue around data centre proposals in several markets, particularly where projects are discussed through maximum design requirements rather than expected annual operating consumption.
Facility metrics need boundaries
On-site cooling water is also only one part of the water footprint associated with computing. Electricity generation can have water impacts upstream, while semiconductor manufacturing and equipment production add embodied resource use outside the facility boundary.
A design can therefore reduce water drawn at the data centre without necessarily reducing every part of the wider water footprint by the same proportion.
For operators and planners, the immediate value of the Schneider comparison is narrower and more practical: two facilities serving a similar IT load can impose materially different local water requirements because their mechanical systems are different.
The same applies to PUE. Higher-density liquid-cooled equipment can improve some aspects of thermal efficiency, but the complete site still includes pumps, heat exchangers, controls, heat-rejection plant, electrical conversion losses, and remaining air-cooled components.
As AI campuses increase in size, small percentage differences translate into large absolute resource flows. A cooling decision repeated across 100MW is no longer a marginal mechanical choice.
The Paris modelling suggests that water demand can be reduced substantially through design, but it also raises the bar for project disclosure. Developers making low-water claims increasingly need to explain the heat-rejection technology, operating assumptions, climatic basis, and measurement boundary behind the number rather than presenting a single headline figure as a universal property of liquid cooling.
