What Supermicro Announced for NVIDIA Vera Rubin: DCBBS Blueprints at Computex 2026

At the rack level, the flagship configuration delivers staggering performance. The NVIDIA Vera Rubin NVL72 rack, which unifies 72 Rubin GPUs with 36 Vera CPUs, serves as a single rack-scale accelerator and achieves 3.6 Exaflops of NVFP4 inference and 1.6 PB/s of HBM4 bandwidth . This performance enables the system to target up to 10x throughput per watt and one-tenth the token cost compared to its Blackwell-based predecessors .

Rack Configurations: NVL72 vs. HGX NVL8

Supermicro's blueprints accommodate two primary server configurations, offering flexibility for different infrastructure needs :

• NVIDIA Vera Rubin NVL72 (Rack-Scale): This is the premier system, packing 72 Rubin GPUs and 36 Vera CPUs into a single liquid-cooled rack. It functions as a cohesive unit, interconnected by NVIDIA NVLink 6, ConnectX-9 SuperNICs, BlueField-4 DPUs, and Spectrum-X Ethernet. This architecture treats the entire rack as a single, powerful accelerator .
• NVIDIA HGX Rubin NVL8 (2U System): For data centers requiring a more traditional form factor, this 2U node houses 8 Rubin GPUs. It is described as the most flexible HGX platform yet, uniquely supporting not just NVIDIA Vera CPUs but also next-generation x86 CPUs from other vendors. A single rack can accommodate nine of these systems, reaching a density of 72 Rubin GPUs per rack .

Cooling the Next Wave: The DLC-2 Direct Liquid Cooling Stack

Managing the immense thermal output of a 1.6 PB/s bandwidth rack is a critical design challenge, and the DCBBS Blueprints address it with an integrated DLC-2 Direct Liquid Cooling stack. This is not an optional add-on but a core, pre-engineered component designed for near-total heat capture and power efficiency .

The stack comprises a complete ecosystem of components :

• Cold plates affixed directly to GPUs and CPUs.
• Cooling Distribution Units (CDUs) available in in-rack and in-row configurations.
• Cooling Distribution Manifolds (CDMs).
• Rear door heat exchangers and cooling towers.
• The proprietary SMC PG25-A ultra-high electrical impedance coolant.
• A key flexibility option is the Liquid-to-Air (L2A) Sidecar CDU, specifically designed for facilities that lack a pre-existing liquid-cooling infrastructure, enabling a phased hybrid cooling approach .

Power and Resilience: 33 kW Shelves and Battery Backup

The Blueprints also address power delivery and system resilience, two non-negotiable aspects of hyperscale AI infrastructure. The solutions integrate 33 kW Power Shelves directly into the rack designs . To protect against power interruptions and prevent the loss of in-flight computation, the blueprints incorporate Battery Backup Units (BBUs) that provide 90 seconds of 48V DC power. This window allows the system to perform a state checkpoint, saving work before a graceful shutdown . The entire facility-side infrastructure is designed from the ground up for single-tenant or multi-tenant environments and scales across a massive 5 MW to 1 GW power envelope .

Beyond Hardware: An End-to-End Deployment and Lifecycle Model

Arguably the most significant differentiator is Supermicro's bundled services and software model, which moves the offering from a hardware sale to a turnkey deployment. The DCBBS Blueprints include a dedicated team of Supermicro experts that manages the complete project lifecycle . This spans:

• Site survey and project design.
• On-site integration and deployment by Supermicro professionals.
• Ongoing post-deployment support.

Underpinning this service layer is the SuperCloud software suite, which provides unified control for the entire infrastructure. It handles deployment automation, offers developer tools, and enables multi-tenant GPU cloud management for service providers . Customer engagements for the DCBBS Blueprints have begun, with physical deployments aligned with the NVIDIA Vera Rubin platform launch scheduled for the second half of 2026 .