NVIDIA's Rubin AI factory reference design achieves 100% liquid cooling at 45°C with zero water consumption

NVIDIA has published the cooling architecture for its next-generation AI factory infrastructure — and the headline spec is a coolant temperature of 45°C, warmer than most hot tubs, running through a closed loop that uses no water at all. The announcement, published on the NVIDIA blog on June 21, 2026, describes the DSX reference design for Rubin-class AI server deployments and the implications for how large-scale GPU infrastructure handles thermal load.

What's new

The NVIDIA DSX (Data Center System) reference design for AI factories achieves 100% liquid cooling at temperatures up to 45°C. Across the cooling system, the design eliminates all water consumption:

"The NVIDIA DSX reference design for AI factories has zero water consumption — we have eliminated massive amounts of power usage and pretty much all water usage."

Key specifications:

• 45°C coolant temperature: Most liquid cooling implementations run coolant at 20–30°C, requiring chiller plants that consume substantial power. Running at 45°C enables direct heat rejection to ambient air in most climates — eliminating the mechanical refrigeration stage.
• 100% liquid-cooled: All heat-generating components — GPU, CPU, memory, networking — transfer their thermal load to the liquid loop. There is no hot-air exhaust to manage at the rack level.
• Zero water consumption: The closed-loop design avoids cooling tower evaporation entirely. No water draw, no water discharge, no water-related permitting requirements.

Context

AI server power density has escalated sharply with each GPU generation. NVIDIA's Hopper-based servers ran at roughly 700W per GPU; Blackwell-class systems exceeded 1,000W per GPU; Rubin-class systems are expected to push higher still. At these densities, air cooling is physically impractical at rack scale — the volume of airflow required exceeds what data center mechanical systems can deliver in a dense deployment.

First-generation liquid cooling responded to this with chilled water loops, but those still required large chiller plants: high capital cost, significant power draw, and a dependency on water supply. NVIDIA's 45°C specification moves the design toward "free cooling" — ambient-temperature heat rejection without mechanical refrigeration — which has been a target for data center designers for years.

Why it matters

For hyperscalers, cloud providers, and large enterprises evaluating Rubin-class deployments, this cooling specification changes three things:

Energy: Chiller plants typically account for a meaningful fraction of a data center's total power draw, worsening Power Usage Effectiveness (PUE) significantly. Eliminating them improves overall facility efficiency and reduces operating cost at scale.

Site selection: Data centers in water-stressed regions — Arizona, Texas, parts of the EU — have faced scrutiny and permitting friction because of their cooling water requirements. A zero-water design removes that constraint. AI factory deployments can go where power and land are available, not where water rights permit.

Waste heat reuse: Liquid coolant at 45°C is warm enough to be useful. District heating networks, building HVAC in colder climates, and some industrial process heat applications can absorb waste heat at that temperature. Data centers have long explored waste heat partnerships but struggled because typical chilled-water loops run too cold to be useful downstream. At 45°C, the economics change meaningfully.

The DSX reference design is the specification NVIDIA hands to colocation providers and hyperscalers building out Rubin infrastructure. Customers adopting it get a standardized cooling architecture rather than designing bespoke thermal systems for each deployment.