NVIDIA's newest Rubin generation AI servers represent the world's first fully liquid-cooled infrastructure, where every chip and networking component is cooled entirely by liquid in a closed loop with no fans. The cooling liquid operates at up to 45 degrees Celsius (113 degrees Fahrenheit) — hotter than typical hot tubs at 38 to 40 degrees — which paradoxically makes the systems significantly more energy efficient. This approach is detailed in the NVIDIA DSX AI factory reference design, which outlines best practices for designing, building, and operating AI factory infrastructure.

Historically, cooling alone has accounted for up to 40 percent of data center electricity consumption. The liquid-cooled approach enables dramatic reductions in these costs. Industry estimates show that raising chiller plant temperatures by just one degree cuts cooling energy costs by approximately 4 percent. At hyperscale, a 50-megawatt facility can save over $4 million annually in cooling-related energy and water costs by switching to liquid-cooled infrastructure.

The design uses a closed-loop system with coolant composed of 75 percent water and 25 percent propylene glycol. In favorable climates, this architecture enables chiller-less operation using outdoor dry coolers, reducing facility cooling water consumption from roughly 2.6 million gallons per megawatt per year for conventional systems to near zero — up to a 100 percent reduction in water use. According to Ali Heydari, director of data center cooling and infrastructure at NVIDIA, the reference design eliminates massive amounts of power usage and nearly all water usage, with dry-cooler designs requiring mechanical chillers only about 1 percent of the year in some climates.

Traditional air-cooled data centers require large volumes of cooled air and energy-intensive cooling infrastructure, particularly in hot weather. The Rubin architecture captures heat directly at the chip and transports it through liquid loops operating at much higher temperatures. The data center ambient temperature becomes flexible since the liquid does all the cooling work, and the same liquid recirculates in a closed loop requiring no new water consumption. Processors continue to operate at full performance because liquid-cooled cold plates keep device temperatures within validated operating limits, with coolant entering fully liquid-cooled chips at 45 degrees Celsius and exiting at roughly 55 degrees after absorbing heat across the chip surface.

The shift eliminates traditional data center challenges. Walk-in freezer-like conditions are no longer necessary, and the architecture removes the need for the careful choreography of hot aisles and cold aisles. Previous hybrid liquid-cooled servers retained air-cooling for some components using finned heat sinks. The Rubin servers are completely redesigned for liquid cooling with cleaner thermal engineering, featuring clean sealed front panels instead of perforated bezels. The fully liquid-cooled design enables higher rack density than air-cooled servers, fitting systems that previously occupied six rack units into just two units — delivering more compute in less space with significantly reduced noise.

An additional benefit is waste heat recovery potential, where residual heat from AI factory operations can be repurposed to heat nearby commercial or residential buildings. Cooling partners like Motivair, the advanced cooling division of Schneider Electric, have aligned with this transition. According to Richard Whitmore, Motivair's president and CEO, once power density per chip crossed a certain threshold, liquid cooling became mandatory.

The industry has long held a misconception that a cold data center is an efficient one. In reality, silicon processors can sustain far warmer environments while maintaining full performance. The liquid cooling methodology at up to 45 degrees Celsius represents one of the most important tools available to address the gap between growing compute demand for AI workloads and the energy cost of running AI at scale.