Supermicro has introduced Data Centre Building Block Solutions blueprints for NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8, aimed at AI data centre deployments from 5MW to 1GW.
The blueprints package compute, storage, networking, liquid cooling, power distribution and site infrastructure into a single deployment model. They are built around a scalable 1,152-GPU unit that can be repeated for larger installations.
The launch reflects growing demand for larger AI data centres and the practical constraints operators face as they move from pilot clusters to sites with far greater power and cooling requirements. Supermicro is targeting customers planning new AI facilities or retrofits, where available power is often the starting point for system design.
Under the blueprint model, Supermicro provides site surveys, project design, integration, deployment and support through a dedicated team. The approach is intended to reduce the number of separate supplier relationships typically involved in AI infrastructure projects, which can span servers, storage, networking, racks, cooling systems, power equipment and services.
One of the central features is direct liquid cooling. Supermicro said its DLC-2 stack is designed for high heat capture and includes cold plates, cooling distribution units, manifolds, rear-door heat exchangers, cooling towers and its SMC PG25-A coolant.
For customers without facility water infrastructure, Supermicro is also offering liquid-to-air options, including a 200kW configuration for one rack and a 500kW configuration for two racks.
The blueprints cover several layers of deployment, from rack-level integration to in-row cooling distribution units and site-level infrastructure. The designs include medium-voltage transformers, low-voltage distribution, rack-level power shelves and battery backup options.
Each Vera Rubin NVL72 rack includes four 110kW power shelves with redundant 18.3kW power supply units, according to Supermicro. The portfolio also includes battery energy storage options for sites that need backup power.
Reference architecture
The new blueprints align with NVIDIA's latest reference architecture for Vera Rubin systems. That matters because buyers of large AI clusters are increasingly looking for standardised deployment patterns that can shorten design cycles and reduce integration risk.
At the centre of the design is a scalable unit containing 1,152 NVIDIA Rubin GPUs and 331TB of HBM4 GPU memory. The blueprint also includes 16 compute racks, six networking racks, four high-performance storage racks and two context memory storage platform racks.
Networking support includes NVIDIA Spectrum-X Ethernet and NVIDIA Quantum-X800 InfiniBand, with throughput of up to 1.6TB/s for the compute fabric. Supermicro added that options will be available for silicon photonics networking with co-packaged optics.
Storage is based on Supermicro's Petascale server platform for NVMe-tier application storage and model checkpointing. The context memory storage platform racks are intended for long-context inference, retrieval workloads and related AI processing tasks.
Software layer
Supermicro is also bundling management software with the blueprint. Its SuperCloud Composer and SuperCloud Director tools provide infrastructure control from bare-metal management to multi-tenant workload orchestration, alongside NVIDIA AI software including NVIDIA AI Enterprise and NVIDIA Run:ai.
The software is aimed at operators running either single-tenant or multi-tenant environments. It also includes asset tracking, with data on physical equipment and sensors across components such as cooling distribution units.
Much of the integration work is handled before delivery through Supermicro's manufacturing operations in the US. That includes racking, cabling and system testing at both system and cluster level, followed by on-site placement, power and cooling connections, software installation and validation.
Supermicro said it has already deployed liquid-cooled AI facilities with more than 100,000 GPUs. It is now using that experience to present a more standardised method for customers looking to expand beyond individual clusters to very large AI sites.
Charles Liang, President and Chief Executive Officer of Supermicro, said the offering is designed to give customers a single route from design through deployment. "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," Liang said.
He added: "We have delivered some of the earliest and largest liquid-cooled AI factories, and that experience is built into every Blueprint - so our customers can move from design to fully operational faster than ever before."