ASRock's Liquid-Cooled NVIDIA HGX B300 Server Doesn't Use Any Water
Instead of the traditional water-based cold plate or immersion systems, HyperCool uses a two-phase dielectric fluid that vaporizes at the heat source and re-condenses in a closed loop. The fluid is non-conductive and non-corrosive, which means that leaks, while still undesirable, don't destroy hundreds of thousands of dollars of hardware when they happen. The goal is to deliver the efficiency of liquid cooling without the operational risks or maintenance complexity that have limited adoption in conventional data centers.
The 4U16X-GNR2/ZC is a factory-assembled, fully-warrantied platform that can scale to eight HGX B300 servers per 42U rack. It's designed for workloads like reasoning AI, agentic inference, and other compute-heavy applications that push GPU thermals beyond what air cooling can reasonably handle. According to ASRock Rack, the setup shortens deployment times and reduces infrastructure costs by arriving pre-integrated and tested rather than requiring custom plumbing or retrofits on site.
ASRock Rack is also showing air-cooled HGX B300 systems and MGX-based configurations using RTX PRO 6000 Blackwell GPUs and BlueField-3 DPUs at the show, but the ZutaCore-cooled model is the clear centerpiece. It reflects an industry trend toward turnkey, self-contained thermal solutions tailored for "AI factories," as NVIDIA calls them, where energy efficiency, safety, and density are becoming as critical as raw compute power.
While details like price, performance metrics, and long-term serviceability remain to be seen, this collaboration suggests that direct-to-chip liquid cooling—particularly in "waterless" form—is moving from niche curiosity to mainstream infrastructure. With AI hardware outpacing traditional thermal limits, innovations like this are simply a matter of survival in the age of the plus-1-kilowatt GPU.
