• Cooling can account for 40% of a data centre’s electricity use
• A 50-megawatt facility could save over $4 million annually on cooling
• Coolant enters processors at 45°C and exits near 55°C during operation

NVIDIA revealed that its latest Rubin AI servers can run on cooling liquid as hot as 45 degrees Celsius, warmer than a hot tub, in what it describes as a major efficiency shift for data centres.

The company said Rubin is the first NVIDIA AI infrastructure generation to use 100% liquid cooling, with every chip and networking component cooled in a closed loop and no fans anywhere in the system.

The company said the design, set out in its DSX AI factory reference architecture, can sharply reduce power and water use at hyperscale.

“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,” said Ali Heydari, director of data centre cooling and infrastructure at NVIDIA. He said some climates may still need chillers for about 1% of the year.

NVIDIA said cooling has historically accounted for up to 40% of a data centre’s electricity use, while industry estimates show that each one-degree rise in chiller plant temperature can lower cooling energy costs by about 4%. It said a 50-megawatt hyperscale site could save more than $4 million a year in cooling-related energy and water costs.

In favourable climates, the 45-degree system can reduce cooling water use from about 2.6 million gallons per megawatt a year in conventional cooling-tower systems to near zero.

The company said traditional data centres rely on cooled air, hot aisles and cold aisles, with fan noise at or above 85 decibels. In Rubin systems, a coolant mix of 75% water and 25% propylene glycol runs through cold plates on processors. Coolant enters at 45 degrees Celsius and leaves at about 55 degrees after absorbing heat, while processors continue to run at full performance.

Richard Whitmore, president and CEO of Motivair, Schneider Electric’s advanced cooling division, said “Once the watts per chip crossed a certain level, liquid cooling became mandatory.”

NVIDIA said full liquid cooling also required a redesign of components that had earlier remained air-cooled. It said the new tray-level architecture uses a single inlet and outlet, sealed front panels and higher rack density, allowing systems that earlier used six rack units to fit into two.

Whitmore said that, in the right location, operators could use outdoor dry coolers without refrigeration equipment. NVIDIA added that cloud providers and data centre operators building for Rubin are making that transition.

The company said the approach could also support waste heat recovery for nearby buildings as AI workloads and data centre construction continue to grow.