ASUS debuts liquid-cooled Vera Rubin AI POD at GTC, claims up to 10x performance

AI infrastructure is turning into a power-and-cooling problem as much as a chip problem. That is the real story behind ASUS showing off a liquid-cooled AI POD for NVIDIA’s Vera Rubin generation at GTC 2026. The pitch is simple: if you want to run giant training jobs or dense inference clusters, normal racks and normal airflow are starting to break. (press.asus.com) ASUS is trying to package the next step as something enterprises can actually buy and install. ### What did ASUS actually launch? ASUS unveiled a broader liquid-cooled AI infrastructure lineup at GTC, but the centerpiece is its ASUS AI POD built on NVIDIA’s Vera Rubin NVL72 platform. ASUS framed it as an end-to-end system rather than just a server — meaning racks, storage, power, cooling, deployment, and support all bundled into one enterprise pitch. (press.asus.com) ### What is Vera Rubin in this context? Vera Rubin is NVIDIA’s next rack-scale AI platform, and it is not just “a faster GPU.” NVIDIA is designing entire pods around compute, networking, storage, and power as one coordinated system. The company says the Vera Rubin POD can span 40 racks with 1,152 Rubin GPUs and is tuned for agentic AI workloads that need huge throughput, low latency, and lots of memory movement. (press.asus.com) ### Why does liquid cooling matter so much? Because these systems are getting absurdly dense. ASUS says its flagship XA VR721-E3, built on Vera Rubin NVL72, is 100% liquid-cooled and can run at up to 227kW TDP in MaxP mode, or 187kW in MaxQ. (press.asus.com) At that heat level, air cooling stops being the easy answer. Liquid cooling is how vendors keep performance from throttling while also trying to lower power usage effectiveness and total cost of ownership. ### Where does the “10x” claim come from? This is the part worth reading carefully. ASUS says the XA VR721-E3 delivers up to 10x higher performance per watt. NVIDIA separately describes the Vera Rubin NVL72 as offering 10x inference performance per watt over Blackwell in the POD context. (developer.nvidia.com) So ASUS is leaning on a platform-level efficiency story tied to Rubin, not claiming some magical ASUS-only breakthrough in isolation. ### Who is this actually for? Not normal enterprises buying a few GPUs. This is for hyperscalers, research labs, sovereign AI projects, and companies building what NVIDIA now likes to call AI factories. ASUS explicitly says the system is built for massive AI workloads and trillion-parameter models. (press.asus.com) Basically, if your bottleneck is rack density, cooling loops, and utility power planning, this is your lane. ### Why is ASUS involved instead of just NVIDIA? Because NVIDIA designs the platform, but somebody still has to turn that blueprint into deployable infrastructure. ASUS is pitching itself as that integrator. It says it is working with Vertiv, Schneider Electric, and other partners on the power and cooling stack, with redundancy and different cooling options depending on the site. (press.asus.com) That matters because buying AI hardware is easy compared with operating it without downtime. ### What is the bigger shift here? The industry is moving from selling chips to selling whole racks and then whole pods. NVIDIA’s own Rubin roadmap is built around five specialized rack-scale systems working together — compute, CPUs, networking, and storage all treated as one machine. (press.asus.com) ASUS is basically saying: fine, we will bring that architecture into the real data center and handle the plumbing. ### Bottom line? This launch is less about a flashy box and more about a new purchasing model for AI. The scarce thing is no longer just GPU supply — it is deployable high-density compute that can stay cool, fed, and fully utilized. (press.asus.com) ASUS wants to be the company that sells that whole answer. (developer.nvidia.com)