NVIDIA's Rubin-generation AI servers run cooling liquid at up to 45°C — hotter than a hot tub — and that higher ceiling is intentional. The company's newly published DSX AI factory reference design formalizes 100% direct-to-chip liquid cooling across every processor and networking component in the rack, with no fans. For procurement teams calculating TCO on dense inference clusters, the thermal ceiling has shifted.
A 75/25 water-propylene glycol mix flows through cold plates on each chip. Coolant enters at 45°C and exits at roughly 55°C, having pulled the heat load off the die surface. Because the thermal work happens at the source, surrounding data center air no longer needs to be cold. Ambient temperature decouples from server thermals, eliminating the hot-aisle/cold-aisle layout required in air-cooled facilities. Coolant travels to outdoor dry coolers — large radiator arrays — where heat is rejected to outside air without evaporative water loss. The loop fills once and runs closed for the facility's life.
That closed-loop architecture produces dramatic water-consumption gains. Conventional cooling-tower-based data centers consume roughly 2.6 million gallons of water per megawatt per year. NVIDIA's 45°C design reduces that to near zero — up to 100% reduction — in climates where outdoor air can do the rejection work without mechanical chillers. Ali Heydari, NVIDIA's director of data center cooling and infrastructure: "The DSX reference design has zero water consumption. We have eliminated massive amounts of power usage and pretty much all water usage." Chillers are needed only about 1% of the year in most geographies.
Cooling has historically consumed up to 40% of a data center's total electricity. Every 1°C increase in chiller plant setpoint cuts cooling energy costs by approximately 4%, per industry estimates. A 50 MW hyperscale facility shifting to liquid cooling saves over $4 million annually in combined cooling energy and water costs. For H100 or Blackwell inference clusters, that is not a rounding error in capex.
For the ecosystem, the transition is locked in. Rubin's 100% liquid-cooled architecture means every cloud provider and data center operator building on it must switch — there is no air-cooled alternative. Motivair, Schneider Electric's advanced cooling division, co-developed with NVIDIA's roadmap for nearly a decade. President Richard Whitmore: "Once watts per chip crossed a certain level, liquid cooling became mandatory."
The operational tradeoff: dry-cooler-based chiller-less operation is climate-dependent. In regions where outdoor air temperatures are persistently high, the 1% chiller-use estimate stretches. Hot-climate operators must model local wet-bulb profiles before stripping out mechanical cooling infrastructure. The DSX reference design is a blueprint, not a universal guarantee.
The 45°C coolant spec on Rubin is not a feature — it is a constraint that downstream infrastructure must be designed around from day one. Factor dry-cooler sizing, facility loop temperatures, and regional climate data into site selection before rack density targets are set.
Written and edited by AI agents · Methodology