NVIDIA details direct-to-chip liquid cooling for Rubin AI servers, cutting facility water use to near zero

Brilliant 🫡

NVIDIA’s Rubin AI servers can now cool every chip and networking part with 45°C water-glycol coolant instead of cold air.

The big deal with this is that cooling water use can drop from about 2.6M gallons per MW per year to near zero in suitable climates

Traditional data centers cool air, then force that air across servers, so the building needs fans, chillers, cold aisles, and often cooling towers that dump heat by evaporating water.

Direct-to-chip cooling skips most of that air problem by bolting cold plates onto GPUs, CPUs, and networking parts, then pumping water-glycol coolant through them so heat leaves the chip through liquid instead of room air.

The strange part is that warmer coolant can be more efficient, because 45°C inlet coolant and roughly 55°C outlet coolant are hot enough for outdoor dry coolers to reject heat like a car radiator in many climates.

A cooling tower spends water to remove heat through evaporation, while a dry cooler spends fan power to move heat into outside air, so water use can fall from about 2.6M gallons per MW per year to near zero in the right location.

NVIDIA’s design is not water-free, because the loop still uses mostly water mixed with glycol, but it can be closed-loop, meaning the same liquid circulates rather than being continually evaporated.

Single-phase immersion goes further by putting electronics in a non-conductive dielectric fluid, where the fluid stays liquid, pulls heat from many surfaces at once, and can run in a sealed loop without water evaporation.

Another strong claim from the immersion side is not only “less water,” but zero process-water cooling plus easier heat capture, because the whole server bath becomes a heat collector.

Heat reuse works only when someone nearby needs low-grade heat, such as buildings, greenhouses, or industrial preheating, while BESS does not create energy but can shift power demand and provide grid services.

They say that a 50MW AI facility can save over $4M per year in cooling-related energy and water costs by shifting to liquid-cooled infrastructure.

And fully liquid-cooled Rubin servers can shrink a system from 6 rack units to 2, which means more AI compute fits into the same building footprint.