NVIDIA has detailed a 45°C liquid-cooling design for its upcoming Rubin AI infrastructure, the world’s first fully liquid-cooled AI compute platform that eliminates fans and near-eliminates water consumption for high-density workloads.
The system runs a closed-loop coolant mixture of 75% water and 25% propylene glycol at temperatures up to 45°C (113°F) directly across chip-mounted cold plates, capturing heat at the source rather than relying on energy-intensive air cooling that traditionally accounts for up to 40% of total data center electricity use 1.
NVIDIA Rubin 45°C Liquid Cooling Eliminates Almost All Data Center Water Use
Per NVIDIA’s official DSX AI factory reference design, the Rubin platform is the first to achieve 100% liquid cooling for every chip and networking component, with no fans anywhere in the system 1. The reference design outlines end-to-end best practices for designing, building, and operating the full AI factory infrastructure stack, including closed-loop cooling loop specifications.
Ali Heydari, NVIDIA’s director of data center cooling and infrastructure, notes the dry-cooler-based design eliminates evaporative water use for all but roughly 1% of the year in most climates 1. For a 1-megawatt AI compute deployment, this cuts annual water consumption from 2.6 million gallons — the standard for conventional cooling-tower-based systems — to near zero 1.
The loop is filled once and recirculates for the facility’s operational life, with no new water required to cool chips even as heat loads climb with denser AI compute.
Do NVIDIA Rubin systems require fans or cold aisle layouts for full performance?
Traditional air-cooled data centers require carefully managed hot and cold aisle layouts, plus fan systems that generate noise levels of 85 decibels or higher — loud enough to require ear protection for on-site staff 1. NVIDIA’s design removes all of that overhead: coolant flows directly through cold plates mounted on processors, exiting the rack at roughly 55°C after absorbing chip heat 1.
There is no dependency on cool ambient air for the servers to operate at full validated performance 1. Unlike traditional systems that ramp up energy-intensive mechanical cooling during hot weather, the 45°C coolant loop allows outdoor dry coolers (large external radiator coils) to reject heat efficiently for most of the year in favorable climates 1.
This eliminates the need for chillers entirely in many regions 1. The data center ambient temperature can run as high as outdoor summer air, as the liquid cooling loop handles all heat removal independently of facility air temperatures.
What coolant mixture does the NVIDIA Rubin 45°C cooling system use?
The closed-loop cooling system uses a proprietary mixture of 75% water and 25% propylene glycol, formulated to operate reliably at temperatures up to 45°C without corrosion or degradation of cooling components 1. The mixture is sealed within the loop for the entire operational lifespan of the data center facility, requiring no top-ups or replacement under normal operating conditions 1. This eliminates ongoing water consumption associated with cooling tower top-offs or bleed-and-fill maintenance cycles common in traditional air-cooled and evaporative cooling systems.
Who is adopting NVIDIA’s liquid-cooled AI infrastructure for high-density workloads?
Rising power densities for modern AI chips have crossed the threshold where air cooling is no longer viable, per Richard Whitmore, president and CEO of Motivair (Schneider Electric’s advanced cooling division) 1. Whitmore has collaborated with NVIDIA on cooling roadmaps for nearly a decade 1. “Once the watts per chip crossed a certain level, liquid cooling became mandatory,” he noted in the NVIDIA blog post 1. Motivair is listed as a key ecosystem partner for NVIDIA’s DSX AI factory reference design, providing cold plate and cooling distribution hardware for Rubin-based deployments.
What are the quantified cost savings of NVIDIA Rubin 45°C liquid cooling?
For hyperscale operators, the financial upside of the 45°C liquid-cooled design is directly quantifiable. Raising chiller plant temperatures by just 1°C cuts cooling energy costs by roughly 4%, per NVIDIA’s reference design data 1. For a 50-megawatt hyperscale facility — such as a mid-sized AI training cluster deployment — that translates to over $4 million in annual savings from combined cooling energy and water costs when switching to the 45°C liquid-cooled architecture 1.
Cooling alone has historically accounted for up to 40% of a data center’s total electricity consumption, making it the single largest area for efficiency gains at hyperscale operators 1. The design also frees up significant data center floor space, as it requires far less physical infrastructure than traditional air-cooling systems for the same compute density 1.
Bottom line: For AI data center operators and hyperscalers building high-density compute infrastructure, NVIDIA’s 45°C liquid-cooled Rubin design delivers a path to save over $4 million annually for a 50-megawatt facility via reduced cooling energy and water costs, with no performance tradeoffs for AI workloads, while eliminating nearly all water consumption for cooling in most climates.
