Supermicro’s Vera Rubin NVL72 Racks: Specs, Cooling and the 5 MW Blueprint

Supermicro announced on September 23, 2026, that it had begun shipping complete NVIDIA Vera Rubin NVL72 racks integrated with its Data Center Building Block Solutions (DCBBS) and DLC-2 direct-liquid-cooling stack. The offering is about more than delivering GPUs: Supermicro is packaging the rack with the infrastructure design needed to deploy it in a larger AI data center. 4

What is in a Vera Rubin NVL72 rack?

One rack combines 72 NVIDIA Rubin GPUs and 36 Vera CPUs across eighteen liquid-cooled compute trays. Each GPU has up to 288 GB of HBM4, giving the rack up to 20.7 TB of GPU memory. Supermicro’s datasheet specifies 1.6 PB/s of aggregate HBM4 bandwidth and 260 TB/s of scale-up bandwidth through sixth-generation NVLink. 1

NVLink connects the GPUs within the rack. ConnectX-9 SuperNICs, BlueField-4 DPUs and Spectrum-X Ethernet are part of the broader networking design for connecting and operating systems beyond one rack. 218

How does the cooling reach beyond the rack?

The NVL72 is a liquid-cooled system, so deployment also requires a way to distribute coolant and handle heat at facility scale. Supermicro’s DLC-2 offering spans cold plates and manifolds, coolant distribution units (CDUs), rear-door heat exchangers and cooling towers. Its listed options include 500 kW liquid-to-air sidecars and 2.2 MW in-row CDUs. These are infrastructure options, not a claim that either cooling unit alone serves every deployment. 85

How does a 5 MW design scale toward 1 GW?

DCBBS blueprints turn a power envelope into a proposed bill of materials for compute, networking, power and cooling. Supermicro describes a Scalable Unit for a 5 MW power envelope with 1,152 Rubin GPUs and 331 TB of HBM4—the GPU count of 16 NVL72 racks. The plan is to repeat these units for larger clusters and sites, with blueprints extending toward 1 GW. That is a deployment architecture, not a statement that a gigawatt-scale Vera Rubin site is already operating. 428

Why does this launch matter to Supermicro?

The shipment advances Supermicro’s shift from supplying individual AI servers toward supplying racks alongside the networking, cooling and facility planning around them. Its reported fiscal 2026 sales were $39.1 billion, versus $22.0 billion in fiscal 2025, and management’s fiscal 2027 revenue target is $65 billion–$72 billion. That target is a forecast, not revenue secured by the NVL72 launch.

NVIDIA positions Vera Rubin for large-scale AI reasoning, including agentic workloads. Supermicro’s commercial premise is that customers building such systems will need deployable clusters and supporting infrastructure—not just accelerators. Whether the new racks contribute materially to its outlook will depend on customer deployments. 174