NVIDIA has introduced Rubin, a new generation of AI infrastructure built around a fully liquid-cooled architecture, marking a significant shift in how hyperscale data centers manage heat and energy consumption. The company said Rubin is the first NVIDIA AI platform designed to operate entirely on liquid cooling, eliminating fans from servers and networking equipment through a closed-loop cooling system.
The design is outlined in NVIDIA’s DSX AI factory reference architecture, which provides guidance for building and operating large-scale AI infrastructure. According to the company, the move toward high-temperature liquid cooling is aimed at improving efficiency as AI workloads continue to increase power and cooling requirements across data centers.
NVIDIA noted that cooling has historically accounted for up to 40% of a data center’s electricity consumption. The company said raising cooling system temperatures can reduce energy demand and lower operating costs, while liquid cooling enables heat to be removed more directly from computing components.
The Rubin platform is designed to operate with coolant entering chips at approximately 45 degrees Celsius and leaving at around 55 degrees Celsius. NVIDIA said processors remain within validated operating limits while maintaining full performance. The cooling system uses a mixture of 75% water and 25% propylene glycol circulated through cold plates attached directly to processors.
The company also highlighted potential reductions in water consumption. Rubin’s cooling architecture is designed around closed-loop systems that rely on dry coolers rather than traditional evaporative cooling methods. In suitable climates, NVIDIA said the system can support chiller-less operation, allowing facilities to reject heat through outdoor dry coolers while reducing cooling-related water use to near zero.
The new architecture also changes the physical design of AI infrastructure. Traditional air-cooled facilities depend on large volumes of airflow, fans, and carefully managed hot and cold aisles. Rubin’s servers use sealed front panels and fully liquid-cooled components, enabling higher rack density and reducing dependence on airflow-based cooling systems.
NVIDIA said its engineering teams redesigned cooling for components that were previously air-cooled, allowing multiple high-power chips to be served through simplified liquid-routing systems. The company added that the absence of fans can also reduce noise levels typically associated with large-scale data center operations.
The broader data center cooling industry is also moving toward liquid-based approaches as chip power requirements increase. NVIDIA cited collaboration with Motivair, the advanced cooling division of Schneider Electric, which has worked with the company on cooling technologies for nearly a decade.
Beyond efficiency gains, NVIDIA said the higher-temperature liquid cooling system could support waste-heat recovery projects, allowing residual heat generated by AI operations to be reused in nearby commercial or residential buildings.
The announcement comes as data center operators face growing pressure to improve efficiency while supporting expanding AI workloads. NVIDIA positioned Rubin and its warm-water liquid cooling approach as part of its broader strategy to reduce the energy and water demands associated with next-generation AI infrastructure.
Read Article: Getty Images and OpenAI Strike Multi-Year Licensing Deal for ChatGPT Visuals
