NVIDIA's Rubin Architecture Runs Coolant at 45°C, Eliminates Fans and Water

A 50-megawatt hyperscale facility burning $4 million a year on cooling alone can save that entire amount by switching to NVIDIA's Rubin architecture. That is the bottom-line consequence of running coolant at 45°C instead of the traditional 38-40°C that hot tubs use.

Rubin is the first NVIDIA generation to achieve 100% liquid cooling. Every chip, every networking component runs on a closed loop of 75% water and 25% propylene glycol. No fans, no cold aisles, no perforated front panels. Walk into a Rubin AI factory and the noise drops from 85+ decibels to near silence.

The 45-Degree Difference

Raising chiller plant temperature by just one degree cuts cooling energy costs by about 4%, per industry estimates. NVIDIA's DSX reference design pushes coolant inlet temperature to 45°C (113°F). The coolant exits the chips at roughly 55°C, having absorbed the heat load directly from the silicon surface.

That higher inlet temperature is the unlock. In many climates, outdoor dry coolers - essentially big radiator coils - can reject the heat without mechanical chillers. Ali Heydari, NVIDIA's director of data center cooling and infrastructure, says the system has zero water consumption. Conventional cooling towers consume roughly 2.6 million gallons of water per megawatt per year. Rubin's closed loop is filled once and runs for the life of the facility, recirculating the same coolant.

No Fans, No Water, No Chill

Richard Whitmore, president and CEO of Motivair (Schneider Electric's advanced cooling arm), says the transition to liquid cooling became mandatory as watts per chip crossed the air-cooling threshold. NVIDIA's thermal team redesigned the entire server: every heat-generating component now gets a cold plate, and a single inlet-outlet loop routes liquid across multiple high-power chips. The result is a cleaner tray-level architecture and a sealed front panel where air-cooled servers have perforated bezels.

In the right geography - think Scottish Highlands, not Phoenix - the system runs chiller-less year-round. Even in warmer climates, chillers turn on only a few days per year. The data center ambient temperature becomes irrelevant; the liquid does all the work.

The Engineering That Made It Work

Previous liquid-cooled servers were hybrids: GPUs and CPUs got cold plates, but memory and networking relied on finned heat sinks and moving air. NVIDIA's team reworked every component's thermal path to use liquid, a problem nobody had solved at rack scale. Rubin servers now pack six rack units of compute into two, because the cooling takes less space and enables higher density.

That efficiency matters as AI workloads keep growing faster than infrastructure investment. Without this shift, the energy cost of running AI at scale would track hardware growth. By running coolant hotter than a hot tub, NVIDIA closes the gap between compute demand and energy budget.

Expect every hyperscaler building for Rubin to adopt this design. The ecosystem is already keeping pace - Motivair has worked alongside NVIDIA's roadmap for nearly a decade. The next step is waste heat recovery: repurposing that 55°C output to heat nearby buildings, turning a cost center into a utility.

Source: Hotter Than a Hot Tub: The 45°C Breakthrough to Cool AI's Biggest Machines
Domain: blogs.nvidia.com