Supermicro DCBBS blueprints scale NVIDIA Vera Rubin AI factories from 5MW to 1GW

Supermicro has introduced new Data Center Building Block Solutions (DCBBS) Blueprints built around NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8, aiming to standardize how large AI clusters are designed and delivered across a wide power envelope, from 5 MW to 1 GW. The company describes the offering as an end-to-end package that spans compute, storage, networking, liquid cooling, power distribution, and site infrastructure, with Supermicro also providing a dedicated services team for delivery and ongoing support.

The DCBBS Blueprints are organized around a repeatable “scalable unit” built with 1,152 NVIDIA Rubin GPUs and 331 TB of HBM4 GPU memory, and the concept is to replicate that building block to reach larger deployments. Supermicro also says the Blueprints align to NVIDIA’s latest reference architecture and integrate NVIDIA Context Memory Storage Platform, NVIDIA Spectrum-X Ethernet, and the NVIDIA Quantum-X800 InfiniBand Platform.

On the cooling side, Supermicro’s DLC-2 direct liquid cooling stack is presented as a full facility-to-rack integration, including cold plates, CDUs, manifolds, rear door heat exchangers, and cooling towers, along with Supermicro’s SMC PG25-A coolant, described as an ultra-high electrical impedance coolant. In a 5 MW example configuration, Supermicro lists 5 MW cooling towers, four in-row CDUs rated up to 1.8 MW each, 16 vertically mounted cooling distribution manifolds, and 576 direct-to-chip copper cold plates (one per host processor module). For sites without liquid cooling infrastructure, Supermicro lists liquid-to-air options, including a 200 kW configuration supporting one rack and a 500 kW configuration supporting two racks.

Power distribution in the Blueprints runs from medium-voltage transformers through low-voltage distribution to rack-level power shelves and battery backup units (BBUs). Supermicro specifies that each Vera Rubin NVL72 rack includes four 110 kW power shelves with redundant 18.3 kW power supply units, and the portfolio includes an option for Supermicro’s Battery Energy Storage System (BESS) for “instant-switching” backup power.

The rack and cluster layouts called out in the Blueprint include 48U and 52U rack enclosure options, 16 compute racks, six networking racks (four compute and two converged) supporting either NVIDIA Spectrum-X Ethernet or NVIDIA Quantum-X800 InfiniBand up to 1.6 TB/s for the compute fabric, four high-performance storage racks based on Supermicro’s Petascale server platform, and two context memory storage platform racks. Supermicro also says options will be available for silicon photonics networking with co-packaged optics (CPO) to eliminate pluggable transceivers.

For operators, the practical value of a “Blueprint” approach is reducing integration friction across cooling, power, networking, and commissioning, where schedule risk often shows up at the interfaces between vendors. But the proof will be in execution: large, liquid-cooled deployments live or die on commissioning discipline, controls integration, and field service responsiveness, not just a bill of materials.

“The NVIDIA Vera Rubin NVL72 platform sets a new standard for AI factory performance, and our DCBBS Blueprints give customers a proven, end-to-end path to build at any scale—from 5MW to 1GW,” said Charles Liang, president and CEO of Supermicro.

Supermicro said the DCBBS Blueprints for NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8 are available for customer engagements, with deployments scheduled for the second half of 2026 aligned with NVIDIA Vera Rubin general availability. More information is available at supermicro.com/en/solutions/dcbbs and supermicro.com/en/accelerators/nvidia/vera-rubin.