NVIDIA says its next-generation Rubin AI servers mark a major shift in data center design by using fully liquid cooling across the system, a setup the company says can reduce cooling energy use and dramatically cut water demand.

The company described Rubin as its first AI infrastructure platform built around 100% liquid cooling, meaning chips and networking components are cooled through closed-loop liquid systems rather than fans and air moving across hardware. NVIDIA says the design allows the coolant to run as warm as 45 degrees Celsius, or 113 degrees Fahrenheit, while still keeping processors within safe operating ranges.

That higher operating temperature is central to the efficiency gains NVIDIA is promoting. By moving heat away from chips directly through liquid, the company says data centers can rely less on traditional mechanical cooling systems, which often consume large amounts of electricity and water. In some climates, NVIDIA says the setup can operate with dry coolers alone and avoid chillers for most of the year.

Cooling has long been a major cost in data centers. NVIDIA pointed to industry estimates showing that cooling can account for as much as 40% of a facility’s electricity use. The company also cited estimates suggesting that every 1 degree Celsius increase in chiller plant temperature can cut cooling costs by about 4%. At hyperscale, those savings can be substantial, with NVIDIA saying a 50-megawatt facility could save more than $4 million annually in cooling-related energy and water costs by shifting to liquid-cooled infrastructure.

NVIDIA said the approach can also sharply reduce water use. In favorable locations, the company says its 45-degree liquid cooling architecture could bring facility water consumption close to zero, compared with conventional cooling-tower systems that use roughly 2.6 million gallons per megawatt per year. The company described the system as closed-loop, with coolant filled once and recirculated for the life of the facility.

The Rubin platform is also designed differently from older liquid-cooled systems, which often used a hybrid approach. NVIDIA said previous designs still depended partly on air cooling for some components. In the Rubin architecture, the company says it reworked the cooling system so that all major parts of the server are cooled by liquid, including those beyond the main processors.

The physical design changes as well. NVIDIA said Rubin servers use sealed front panels instead of the perforated bezels seen on air-cooled machines, and that fully liquid-cooled racks can support higher density. In one example, the company said a system that previously took up six rack units can now fit into two.

The company’s cooling partner ecosystem is adapting to the new approach. Motivair, Schneider Electric’s advanced cooling division, said it has worked with NVIDIA for years and that liquid cooling became necessary once chip power levels rose beyond what air cooling could handle.

NVIDIA framed the shift as part of a broader change in how AI facilities are built. The company said its DSX AI factory reference design supports the liquid-cooled architecture and can enable systems with no fans, no cold aisles, and, in the best-case scenario, no evaporative water use.

As AI workloads continue to grow, NVIDIA argues that improvements in cooling will be essential to keeping the energy footprint of data centers under control. The company says liquid cooling at warmer temperatures is one of the most important ways to make that possible.