NVIDIA's Rubin Goes 100% Liquid-Cooled at 45°C, Slashing Data Center Water Use

NVIDIA’s upcoming Rubin AI infrastructure is the first to drop fans entirely and cool every chip and networking component through a closed liquid loop. The key trick is running coolant—a 75% water, 25% propylene glycol mix—at temperatures up to 45°C (113°F). Because the silicon stays within validated limits even as coolant enters at 45°C and exits around 55°C, performance holds steady while the loop runs far hotter than conventional designs. That higher operating temperature is what makes the system efficient: heat is pulled directly off the chip and rejected through outdoor dry coolers (essentially large radiator coils), so in favorable climates mechanical chillers barely run.

The payoff is in energy and water. Cooling has historically consumed up to 40% of a data center’s electricity, and NVIDIA’s DSX reference design aims for near-zero water consumption by skipping evaporative cooling towers. The company cites cuts from roughly 2.6 million gallons of water per megawatt per year to nearly nothing, and estimates a 50MW facility could save over $4 million annually. The loop is filled once and recirculated for the life of the building. Server density also improves sharply—hardware that once took six rack units now fits in two—with sealed front panels replacing perforated bezels and noise dropping below the 85-decibel threshold that requires ear protection.

The shift is being driven by necessity rather than preference: once per-chip power crossed a certain wattage, air cooling stopped being viable, according to Schneider Electric’s Motivair division. Because Rubin ships as a fully liquid-cooled platform, every operator building for it must adopt the approach, effectively making it an industry default. The main caveat is geography—chiller-less operation works best in cool climates like Scotland, less so in Phoenix—though even warm regions move closer to the chiller-free ideal. NVIDIA also flags waste-heat recovery, reusing residual heat to warm nearby buildings, as a further upside as AI compute demand continues to outpace other infrastructure growth.