Hotter Than a Hot Tub: The 45°C Breakthrough to Cool AI’s Biggest Machines
07:00 · June 22, 2026 · NVIDIA
Hot tubs sit at about 38 to 40 degrees Celsius, warm enough that most people can only soak for about 15 minutes. NVIDIA’s newest AI servers can run their cooling liquid even hotter — up to 45 degrees Celsius, or 113 degrees Fahrenheit. That higher temperature limit is precisely what makes them more energy efficient. […]
Summary
NVIDIA’s Rubin-generation AI servers mark the first full transition to 100 percent liquid cooling, with every processor, networking component and supporting element served by a closed-loop system that operates without fans or cold-aisle infrastructure. Coolant enters cold plates at up to 45 °C—roughly five degrees warmer than a typical hot-tub temperature—exits at about 55 °C and is routed through a mixture of 75 percent water and 25 percent propylene glycol. Because the liquid captures heat directly at the chip, the surrounding data-center air can remain at ambient outdoor temperatures for most of the year.
The approach is formalized in the NVIDIA DSX AI factory reference design, which specifies dry-cooler rejection loops that eliminate evaporative cooling towers. In suitable climates the facility can run chiller-free except for perhaps one percent of annual hours, cutting cooling-related electricity use that historically reached 40 percent of total data-center demand. Industry estimates indicate each additional degree of allowable coolant temperature reduces cooling energy by roughly four percent; at hyperscale, a 50-megawatt installation can therefore save more than four million dollars annually in combined energy and water costs while reducing water consumption from approximately 2.6 million gallons per megawatt-year to near zero.
Beyond efficiency, the sealed liquid architecture removes the perforated bezels and high-speed fans of earlier hybrid designs, lowering rack noise below the 85-decibel threshold that once required ear protection and allowing previously six-rack-unit assemblies to fit in two rack units. The same higher-temperature loop also opens the possibility of waste-heat recovery for nearby buildings. Every cloud provider and operator adopting the Rubin platform must therefore re-engineer its cooling stack, a shift already supported by long-standing partners such as Motivair, now part of Schneider Electric, whose cold-plate and coolant-distribution units have tracked NVIDIA’s power-density roadmap for nearly a decade.
Why it matters
This article is highly relevant as it addresses the critical environmental impact of AI data centers, a major concern in the Netherlands given its dense data center footprint. The breakthrough in liquid cooling offers significant energy and water savings, which is vital for Dutch enterprises and policymakers focused on sustainable AI infrastructure.
