NVIDIA's Rubin AI platform ships with 100% liquid cooling across every chip and networking component — a full departure from the hybrid air-and-liquid configurations that have defined data center thermal management for the past decade.

The details, published by NVIDIA on 22 June 2026, confirm that Rubin operates at a 45°C warm-water cooling temperature, meaning the facility infrastructure itself can run on water that would have been considered too warm for conventional cooling loops just a few years ago. That 45°C threshold matters operationally: it means Rubin-based deployments can reject heat into cooling towers or dry coolers at ambient temperatures that previously required chilled-water plants, cutting both capital expenditure and the energy overhead of mechanical refrigeration.

From Hybrid to Homogeneous

The shift from partial to total liquid coverage is the harder engineering story here. Earlier GPU generations used liquid cooling selectively — GPU dies, occasionally memory stacks — while leaving networking ASICs, power delivery, and optics on air. Rubin closes that gap. NVIDIA's blog post states the platform extends liquid loops to every chip and every networking component in the rack. That eliminates the mixed-mode thermal planning that has plagued hyperscale build-outs, where air-cooled periphery constrained overall rack density even when the primary compute was already on liquid.

The progression from Blackwell is measurable. NVIDIA's Blackwell platform, detailed in an April 2025 post, already posted a 300x improvement in water efficiency over earlier generations and delivered a 12% increase in compute throughput while reducing energy draw through its liquid cooling implementation. Rubin takes those gains as a baseline, not a ceiling.

The Water Efficiency Nuance

There is a distinction worth holding carefully. NVIDIA's cooling architecture reduces water consumption inside the data center — the coolant circulating through on-premises liquid loops — but that is not the same as eliminating the facility's total water footprint. As TechCrunch noted on 22 June 2026, the broader water problem for AI infrastructure also involves the upstream water consumption at power generation facilities, which is a function of grid mix and geography rather than chip-level thermal design. Rubin's architecture addresses the on-site side of that equation. The off-site side remains a grid and siting problem.

That is not a criticism of the engineering. It is a scoping note for anyone planning a sustainability brief around these specs.

Production Status and Timing

Rubin entered production as of CES 2026 in January, where Data Center Frontier reported the 45°C warm-water capability alongside the production announcement. The June 2026 blog post from NVIDIA provides the fuller technical picture and confirms the all-liquid scope that the January coverage had only partially captured. For infrastructure teams already planning Rubin deployments, the June documentation is now the authoritative specification reference.

What This Changes for Operators

At 45°C supply temperature, Rubin-compatible facilities can frequently operate without mechanical chillers in moderate climates, running economizer-only modes for a larger fraction of the year. That has direct bearing on PUE calculations, water tower sizing, and — for colocations selling AI-optimised space — the capital cost of retrofitting older halls.

The all-liquid mandate also restructures the deployment conversation. Operators who have deferred liquid cooling infrastructure investments while running hybrid Hopper or Blackwell configs no longer have a middle path with Rubin. The platform requires full CDU and manifold infrastructure from rack one. That is a harder conversation for enterprise buyers with existing air-cooled halls, and a cleaner one for greenfield hyperscale builds designed around liquid from the ground up.

The Blackwell-to-Rubin arc — from selective liquid coverage to complete elimination of in-rack air cooling, from water efficiency gains measured in the hundreds of times over legacy systems to warm-water operation that removes chiller dependencies — suggests NVIDIA is treating thermal architecture as a first-class design constraint rather than an afterthought to compute density. Given where GPU TDPs are heading, that sequencing is logical. The next question for the industry is whether facility standards bodies and colocation providers can keep pace with the infrastructure requirements the silicon roadmap is now dictating.