Summary
- Trane Technologies and Eaton have developed coordinated power and cooling reference designs aligned with Nvidia's DSX architecture.
- The companies claim up to 15% efficiency gains, 30% lower installation costs, and 80% less copper than conventional low-voltage designs.
- The approach reflects growing pressure to design electrical distribution, liquid cooling, and controls as one system as AI rack densities rise.
Trane Technologies and Eaton have combined their electrical and thermal architectures into a reference design intended to shorten the design and deployment cycle for high-density AI data centres.
The design is aligned with Nvidia’s DSX platform and moves more of the electrical distribution architecture towards medium voltage, while coordinating it with cooling systems and controls rather than treating power and thermal plant as separate design packages.
The companies say the approach could improve combined energy efficiency by up to 15%, reduce installation costs by as much as 30%, and cut copper use by up to 80% compared with conventional low-voltage designs. Those are supplier estimates rather than results from a disclosed live data centre deployment, and the actual savings will depend on site topology, load density, redundancy, cooling architecture, and local electrical requirements.
As AI clusters move towards higher rack power densities, the infrastructure between the utility connection and the chip is becoming harder to design in isolated disciplines. Electrical conversion losses, cable runs, switchgear, cooling distribution, control logic, and white-space layout increasingly influence one another.
Trane and Eaton’s design is intended to create a more repeatable pathway from grid connection through electrical distribution and thermal management to the IT load. Eaton’s systems provide the power-distribution element within Trane’s Continuum Rubin DSX platform, while both companies are also incorporating their work into their own Nvidia-aligned reference architectures.
Medium voltage moves closer to compute
One of the most consequential changes is greater use of medium-voltage distribution deeper into the data centre electrical chain. Traditional low-voltage designs can require very large quantities of copper as power demand rises, particularly where current levels become extreme. Raising distribution voltage allows the same power to be moved at lower current, potentially reducing conductor size, losses, and the physical bulk of parts of the installation.
That does not make the design inherently simpler. Medium-voltage equipment introduces different protection, isolation, safety, maintenance, and commissioning requirements. Operators also need to consider how a more integrated architecture affects resilience boundaries, fault containment, maintainability, and the ability to replace components without disturbing the wider system.
The thermal side is changing at the same time. Direct liquid cooling is moving from a specialist high-performance-computing technology towards a mainstream requirement for the densest AI deployments. That creates new interfaces between cooling distribution units, pumps, heat exchangers, facility-water systems, controls, and the electrical systems feeding the racks.
Coordinating those elements at reference-design stage can remove part of the redesign work that occurs when a compute platform is specified after the building’s mechanical and electrical assumptions are already fixed. It also gives developers a repeatable template that can be adapted across multiple campuses rather than recreating the entire electrical and cooling topology for every project.
The strongest claims in the announcement still need to be tested against live projects. A 30% installation-cost reduction or 80% reduction in copper use would be material at hyperscale, but the baseline design and boundaries used for comparison will determine how transferable those figures are.
The broader direction is less ambiguous. AI facility design is moving away from a sequence in which utility, electrical, mechanical, and IT systems are specified independently. Higher-density compute makes those interfaces part of the primary engineering problem. The value of reference architectures will therefore depend not only on how quickly they can be deployed, but on whether they preserve the resilience, serviceability, and operational flexibility expected from critical infrastructure once the facility is live.
