Summary
- Trane has added two 250MW zero-water cooling architectures to its Nvidia DSX reference-design portfolio.
- One uses air-cooled chillers and free cooling, while the second packages water-cooled chillers and dry coolers into factory-built modular plants.
- The designs show how cooling architecture is increasingly being engineered around the amount of scarce site power that can ultimately reach compute.
Trane Technologies has introduced two new 250MW AI data centre cooling architectures using zero-water heat rejection, extending its reference-design work around Nvidia’s DSX platform.
The first design combines direct-to-chip liquid cooling with air-cooled chillers and integrated free cooling. Trane says the architecture can improve cooling-system efficiency by up to 25% and release as much as 22MW of electrical capacity for compute compared with the baseline against which it was modelled.
The second architecture packages water-cooled chillers and closed-loop dry coolers into factory-built modular cooling plants supplied with Stellar Energy. Trane says the configuration can improve chiller power efficiency by up to 16% and reallocate around 8MW towards the AI load while maintaining zero water usage effectiveness at the heat-rejection stage.
Both designs integrate LiquidStack’s GigaModular coolant distribution technology. The CDU platform is rated at up to 14MW under the conditions specified by the supplier, illustrating the scale of liquid-cooling infrastructure now being designed around campus-level AI loads.
The key change is not simply the size of the cooling equipment. Reference architectures are increasingly being engineered around the amount of scarce electrical capacity that remains available to the IT load after pumps, fans, chillers, controls, and heat-rejection systems have taken their share.
That puts cooling design directly into the capacity calculation. On a site where grid access is constrained, a more efficient mechanical system can make additional megawatts available to servers without changing the incoming electrical connection. The value of those gains depends on climate, redundancy, operating conditions, and the final IT load, but the principle is becoming more important as campuses approach utility limits.
Zero-water heat rejection also addresses another constraint. Evaporative systems can reduce electrical consumption in some climates, but large projects increasingly have to account for water availability, local restrictions, and the operational consequences of using potable or treated water for cooling.
The modular-plant approach shifts part of the mechanical programme into a factory setting, where chillers, pumps, controls, pipework, and heat-rejection equipment can be assembled and tested before delivery. That can reduce site labour and commissioning risk, although final integration with the data hall and utility systems still has to be proven on location.
Trane has been developing reference designs with both Nvidia and electrical-equipment suppliers as the boundary between power and cooling becomes less distinct. High-density facilities need those systems to be coordinated early because changing pipe sizes, plant layouts, electrical distribution, or redundancy late in construction can create major cost and programme penalties.
The two 250MW designs reinforce a broader shift towards treating thermal management as a capacity-allocation problem rather than a supporting service. The useful output of a data centre is compute, and the infrastructure surrounding it is increasingly judged by how little power and water it consumes while keeping that compute within operating limits.
