Rubin platform ships 100% liquid cooling at 45°C inlet for zero-water AI factories
NVIDIA's 45°C liquid cooling spec removes both energy and water bottlenecks at hyperscale. The change forces immediate redesign of power, siting, and cost models for AI factories. Adoption will be gated by dry-cooler availability rather than server shipments.
NVIDIA's DSX design eliminates air cooling and evaporative towers for Rubin-generation clusters. Every GPU, CPU, and switch is cooled in a closed liquid loop. Dry coolers reject heat directly when ambient conditions allow, with chillers required less than 1% of annual hours in suitable climates. This removes the 40% of facility power historically spent on cooling and ends water evaporation losses.
Data from the reference architecture shows a 50 MW installation saves more than $4 million yearly in combined energy and water costs. Raising loop temperatures by one degree reduces chiller work approximately 4%. The 45°C specification enables chiller-less operation across wide geographic ranges, a threshold air cooling cannot reach at Rubin power densities.
Prior coverage focused on efficiency gains while omitting the hard infrastructure constraint: water availability now limits siting for clusters above 100 MW. The closed-loop design also reduces Scope 3 emissions tied to water treatment and transport. Operators must now re-rate PUE calculations and negotiate power-water contracts together.
Next deployments will standardize on dry-cooler retrofits for Blackwell-to-Rubin migrations. Facilities without access to dry-cooler real estate will face 15-20% higher TCO from retained chillers.
NVIDIA: 60% of Rubin-based clusters deploy dry-cooler loops by end of 2026.
Sources (3)
- [1]Primary Source(https://blogs.nvidia.com/blog/liquid-cooling-ai-factories/)
- [2]Supporting Source(https://www.nrel.gov/docs/fy23osti/85067.pdf)
- [3]Supporting Source(https://www.opencompute.org/documents/ocp-liquid-cooling-ocp-acs-2024)