Nvidia has unveiled a radical approach to AI data centers, with its Rubin generation reference design promising near-zero water consumption through 100 % liquid cooling . The company claims the new design eliminates “ huge amounts of power consumption and virtually all water consumption ,” addressing one of the most contentious issues facing modern AI data centers . The move comes at a time when public opinion is putting increasing pressure on tech companies to address their environmental impact.
See also: A Chinese startup dethroned Nvidia in the benchmark it helped create
The cooling challenge in modern data centers is not new, but it has become dramatically more acute in recent years. Modern GPU racks consume 120–140 kW or more per unit, a level that traditional air-cooled systems cannot effectively manage. Conventional water tower cooling systems consume approximately 2.6 million gallons of water per megawatt per year, mainly due to evaporative losses. Furthermore, cooling can account for up to 40% of a data center’s total electricity consumption , making efficiency a critical factor.
Nvidia ’s solution with the Rubin platform is a bold one: running servers at temperatures up to 45 degrees Celsius (113°F) , with full closed-loop liquid cooling. According to Ali Heydari , Nvidia ’s director of data center cooling , “ the NVIDIA DSX reference design for AI factories has zero water consumption .” Heat is captured directly at the chip level and transferred through liquid circuits that operate at much higher temperatures, allowing external dry coolers to efficiently dissipate heat most of the time, without the need for chillers .
Nvidia Rubin: Technical details and financial benefits
Josh Parker , Nvidia ’s chief sustainability officer , says the reference design reduces water consumption “ from about 2.6 million gallons per megawatt per year for conventional cooling tower systems to almost zero — a reduction of up to 100 percent .” This is achieved by operating without a chiller in favorable climates, using only dry coolers . For a 50-megawatt installation , this translates to energy savings of more than $4 million per year . It’s worth noting that increasing the temperature of the cooling system by 1°C reduces cooling costs by about 4 percent , making Nvidia ’s 45°C threshold a significant technological leap.
See also: Redesigned Siri will use Nvidia chips for fast and private cloud computing
The Rubin platform, specifically the Vera Rubin NVL72, supports up to 144 GPUs per rack, delivering 7+ exaflops of AI performance and 5 petaflops of FP64. The architecture is designed to serve both hyperscalers and research institutions that need compute power for climate modeling, quantum chemistry, and energy exploration. Nvidia says that “every cloud provider and data center operator that builds for Rubin is making this transition.”
The ecosystem around the Rubin platform includes major manufacturers and infrastructure providers. Companies such as Schneider Electric (Motivair) , Vertiv , nVent and Supermicro are developing specialized architectures for the Vera Rubin NVL72 platform . Supermicro says that the Vera Rubin NVL4 racks provide 3.2 MW of scalable units with native FP64 performance , enabling converged HPC and AI infrastructure. Vera Rubin NVL4 systems are expected to be available in Q4 2026 from manufacturers such as Dell , HPE and Supermicro . Schneider Electric emphasizes that “ liquid cooling does not equal high water consumption ” and that water consumption is mainly associated with evaporative cooling, which closed-loop systems avoid entirely.
Despite the impressive results, Nvidia ’s approach doesn’t address all the concerns surrounding AI data centers . The company’s blog post doesn’t address the cost of building this type of data center compared to one that uses less efficient air cooling — a point that Gizmodo also points out . It also doesn’t address concerns related to the construction phase of the facilities or the huge power requirements. Amazon , in a recent report, also praised the higher heat tolerance as part of the efficiency improvements of its mostly air-cooled data centers, indicating that this trend is broader across the industry.
See also: Nvidia RTX Spark: The next generations of the chip for AI computers are being designed
Overall, Nvidia ’s Rubin reference design represents a significant step towards more sustainable AI data centers , setting new standards for cooling efficiency and resource savings. The AI data center market is expected to exceed $150 billion by 2027 , and the adoption of standardized reference designs like Nvidia ’s will accelerate the transition to fully liquid-cooled facilities. However, the broader challenge of the overall energy footprint of AI remains open, requiring continued industry collaboration.
