Nvidia has introduced a closed-loop liquid cooling system for its Rubin chips. The company says the coolant enters the loop at about 113°F and exits at about 131°F.
Practical effect
Because the coolant’s baseline temperature is higher than most ambient air, data centers can reject heat with outdoor dry coolers instead of relying on traditional chillers. Chillers can consume a large share of a facility's electricity—nearly 40%—and Nvidia claims the closed system can reduce water use by up to 100% because the loop is filled once and sealed for the building's life.
Why this matters for data centers and the environment
The approach targets two major operational burdens: water consumption and cooling-related electricity use. If the cooling loop is hermetically sealed and needs no replenishment, water use for cooling can fall toward zero. Replacing chillers with outdoor dry coolers can significantly cut the energy dedicated to cooling, easing pressure on local water supplies and electrical grids.
Context and timing
Earlier this year, local governments delayed or blocked more than 75 data center projects citing concerns about water availability and grid strain. Nvidia’s cooling design directly addresses those concerns by lowering both water demand and chiller-related power draw.
Additional figures and suppliers
Data-center infrastructure supplier Vertiv estimates that each additional degree of warmer inlet temperature yields about a 4% energy saving, meaning higher-temperature cooling baselines can compound efficiency gains.
Remaining challenges
The cooling system reduces peripheral resource demands, but it does not eliminate the fundamental power consumption of the compute hardware itself. GPUs and other compute resources continue to be significant electricity consumers.
Conclusion
Nvidia’s closed-loop liquid cooling for Rubin chips could materially reduce water use and the share of power devoted to cooling in data centers, addressing several regulatory and community concerns. However, overall data-center energy demand will still be driven largely by the compute workloads they host.
