Summary
- Supermicro says Vera Rubin NVL72 racks are now shipping with its DLC-2 liquid-cooling stack.
- A 5MW scalable unit contains 1,152 GPUs and can be repeated towards gigawatt-scale designs.
- Facility infrastructure includes cold plates, CDUs, heat rejection, networking, storage, and commissioning services.
Supermicro says it has begun shipping Nvidia Vera Rubin NVL72 racks alongside a facility-scale infrastructure package that takes liquid cooling, networking, storage, and commissioning from the rack level into multi-megawatt AI data centre design.
The company’s Data Center Building Block Solutions architecture packages 1,152 Vera Rubin GPUs into what it describes as a 5MW scalable unit. Multiple units can be repeated to create larger clusters, with Supermicro publishing design blueprints extending towards gigawatt-class deployments.
The NVL72 rack uses 72 Rubin GPUs and 36 Vera CPUs connected through Nvidia’s latest NVLink architecture. Supermicro’s associated DLC-2 cooling portfolio includes direct-to-chip cold plates, in-rack or in-row coolant distribution, rear-door heat exchangers, liquid-to-air equipment, and cooling towers.
At this density, the compute hardware and the building services can no longer be specified independently. The power and thermal envelope of the rack determines pipework, CDU capacity, electrical distribution, heat rejection, controls, and commissioning requirements across the hall.
Five megawatts becomes a repeatable compute block
Supermicro’s 5MW blueprint places 1,152 GPUs and 331TB of HBM4 memory into one repeatable infrastructure unit.
The concept resembles modular electrical design: rather than engineer every campus from a blank sheet, the developer starts with a defined block and repeats it as land and power become available.
That approach can shorten design work and simplify procurement, but only where the surrounding facility is capable of accepting the block. Five megawatts of IT equipment still requires upstream transformers, switchgear, UPS or alternative ride-through systems, distribution, network connectivity, and substantial heat rejection.
Supermicro specifies rack thermal designs reaching approximately 227kW for the NVL72 architecture. Other Rubin configurations in its portfolio extend still higher, demonstrating how rapidly individual-rack loads are moving beyond conventional air-cooled layouts.
The cooling stack therefore becomes part of the compute bill of materials. In-row CDUs can serve multiple racks, while facility heat-rejection equipment has to be sized for the aggregate load rather than each server generation in isolation.
Commissioning moves across the IT-facility boundary
Supermicro is also offering site surveys, design, integration, deployment, and L11/L12 testing as part of the package.
That matters because high-density AI systems expose a traditional organisational boundary inside data centres. IT teams historically procured servers while facilities teams operated power and cooling. Rack-scale AI systems require the two groups to commission equipment together.
A fault in a pump, CDU, busway, switchboard, network fabric, or control system can now remove an unusually large block of compute. Conversely, server behaviour can impose rapid electrical or thermal transients on infrastructure outside the rack.
The commissioning process therefore has to test not simply whether a server boots, but whether the complete chain from utility supply and electrical protection through liquid flow, control sequences, network fabrics, and workload behaviour responds correctly under failure conditions.
Supermicro’s announcement also indicates that supply is moving beyond reference designs. Shipping hardware does not mean every announced Vera Rubin data centre has been completed, but it starts the process through which facility designs have to become operational rather than conceptual.
European relevance comes from the same constraint facing projects globally: a Rubin rack cannot be deployed merely because a customer can buy the GPUs. The site needs suitable electrical density, cooling interfaces, commissioning expertise, and heat rejection before the hardware becomes usable compute.
DataCentral previously reported Civo’s plan to deploy Vera Rubin systems in new UK liquid-cooled facilities from early 2027. Supermicro’s shipping announcement puts more definition around the type of facility infrastructure those deployments will require.
The scale is the important change. Liquid cooling has moved from a technology used to solve an unusual rack to a design assumption for complete multi-megawatt compute blocks. At that point, the cooling system is no longer an accessory to the server. It is part of the data centre’s primary production infrastructure.
