← Back to feed

Super Micro Computer (NASDAQ: SMCI) has begun shipping production NVIDIA Vera Rubin NVL72 racks, integrating its proprietary DLC-2 direct liquid cooling stack with NVIDIA’s reference architecture. The announcement marks the shift from early deployments to volume supply for the Vera Rubin platform.

Rack Configuration

Each Vera Rubin NVL72 rack is built around:

• 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs operating as a single machine
• Eighteen 1U compute trays, each housing 4 Rubin GPUs and 2 Vera CPUs
• Nine sixth-generation NVIDIA NVLink switch trays delivering 216 TB/s scale-up bandwidth
• 20.7 TB of HBM4 memory per rack
• Up to 54 TB of LPDDR5X auxiliary memory per rack

The NVLink configuration means all 72 GPUs within a rack communicate at full fabric speed, the key architectural difference from prior-generation multi-node setups requiring InfiniBand between GPU nodes.

Liquid Cooling at Scale

Supermicro is shipping complete liquid cooling infrastructure alongside the compute racks. The in-row cooling distribution units (CDUs) are rated at 1.8 MW per CDU, deployed with N+1 redundancy. Optional rear-door heat exchangers capture residual heat load not handled by direct liquid paths.

CEO Charles Liang: “We have spent years building the liquid-cooling stack, the manufacturing capacity, and the deployment teams for exactly this moment. Our customers can now order a Scalable Unit and receive production-ready systems with end-to-end integration, because we design and build every layer between the cold plate and the cooling tower.”

DLC-2 is Supermicro’s second-generation direct liquid cooling architecture, developed specifically for the Vera Rubin platform’s heat density. Rubin’s throughput-per-watt gains over Blackwell come with a proportionally larger thermal load.

DCBBS Blueprint: 1,152 GPUs Per Scalable Unit

Supermicro’s Data Center Building Block Solutions (DCBBS) Blueprint defines a balanced bill of materials for deployment at scale. The standard Vera Rubin NVL72 Scalable Unit is 16 compute racks:

• 1,152 NVIDIA Rubin GPUs
• 331 TB HBM4 memory
• Balanced cooling, power delivery, high-performance storage, and networking

The DCBBS Blueprint scales from 5 MW to gigawatt-level deployments. Supermicro is managing end-to-end project delivery including site survey, design, integration, testing, delivery, and ongoing support.

Context

CoreWeave was among the first to deploy Vera Rubin NVL72 at rack scale, with preview inference results appearing in MLPerf Inference v6.1 in September 2026. NVIDIA published benchmark data in August showing Vera Rubin NVL72 achieving up to 30x agent throughput per megawatt over GB300 NVL72. Supermicro’s shipping announcement signals the platform is now in production volume, accessible to any operator with sufficient power and cooling infrastructure rather than only hyperscalers with direct NVIDIA relationships.

Wistron’s Texas factory for AI server production, including Vera Rubin hardware, opened in July. The SMCI announcement adds a second major system integrator alongside Dell capable of delivering complete NVL72 rack solutions with integrated thermal management.