Supermicro Commences Volume Shipments of NVIDIA Vera Rubin NVL72 Liquid-Cooled Supercomputers
Supermicro officially begins volume shipments of complete, turnkey liquid-cooled NVIDIA Vera Rubin NVL72 racks featuring 72 Rubin GPUs, 36 Vera CPUs, and 20.7 TB HBM4.
On September 23, 2026, Supermicro officially announced commercial volume shipments of its complete Data Center Building Block Solutions (DCBBS) for the NVIDIA Vera Rubin NVL72 platform. Unifying 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs into a single liquid-cooled rack footprint with 20.7 TB of HBM4 memory, Supermicro provides turnkey datacenter integration spanning factory-integrated 1.8MW cooling distribution units, modular manifolds, and L11/L12 validation testing to accelerate gigawatt-scale AI infrastructure deployments.
Turnkey Liquid-Cooling Infrastructure and 1.8MW CDU Deployment
Operating at unprecedented thermal and electrical densities, each Vera Rubin NVL72 rack requires an end-to-end engineered liquid cooling solution. Supermicro comprehensive DCBBS stack provides customized cold plates, quick-disconnect stainless steel manifolds, and dedicated in-row Cooling Distribution Units (CDUs) rated at 1.8 megawatts each with N+1 redundancy. To ensure zero facility downtime, Supermicro executes rigorous Level 11 and Level 12 pre-delivery burn-in testing, delivering plug-and-play liquid loops directly to hyperscale data center floors.
System Architecture & Cooling Highlights
- Rack Compute Topology: Houses 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs distributed across eighteen 1U liquid-cooled compute trays.
- NVLink 6 Interconnect: Nine NVLink switch trays provide 216 TB/s of aggregate bidirectional scale-up bandwidth, unifying the rack into a single accelerator.
- Memory Pool: Integrates 20.7 TB of ultra-high-speed HBM4 memory alongside up to 54 TB of coherent LPDDR5X system memory.
- 1.8MW In-Row CDUs: Custom cooling distribution units equipped with redundant pumps and smart telemetry for facility water loops.
- Scalable Unit Architecture: Standardized modular blueprints scale seamlessly from 5 MW clusters to multi-rack gigawatt clusters with 1,152 GPUs per block.
72 Rubin GPUs, 216 TB/s NVLink Bandwidth, and Gigawatt Scaling
By transitioning from component-level manufacturing to full rack-scale and cluster-scale integration, Supermicro substantially compresses the deployment lead time for frontier AI foundation model builders. Through pre-configured network fabrics spanning compute InfiniBand, RoCE v2, and converged management cabling, enterprise hyperscalers can bring thousand-GPU clusters online in days rather than months, accelerating the commercial execution of multi-trillion parameter reasoning and agentic workloads.
Product Profile Sheet
| Feature | Details |
|---|---|
| Announcement Date | September 23, 2026 |
| Architecture | NVIDIA Vera Rubin NVL72 (72 GPUs / 36 CPUs) |
| Interconnect Bandwidth | 216 TB/s Aggregate (NVLink 6 Switch Trays) |
| Manufacturer | Supermicro (Super Micro Computer, Inc.) |
| Onboard Memory | 20.7 TB HBM4 + 54 TB LPDDR5X |
| Primary Source | Supermicro Investor Relations |
| Shipment Status | Active Commercial Volume Deliveries |
| Thermal Management | Supermicro DCBBS Liquid Cooling (1.8MW CDUs) |