NVIDIA’s DSX POD moves more AI data-centre assembly into the factory

NVIDIA is extending the idea of a preassembled computing system to a larger piece of the data centre. Its DSX POD reference design, announced on 21 October, brings Vera Rubin equipment and supporting infrastructure together in modular containers, with integration and functional checks completed before shipment. Partners are expected to start delivering units in early 2027.[1]

The company says deployment at a prepared site could fall from six months to as little as 14 days. The starting condition matters to the schedule: power, cooling, networking and supporting site infrastructure must already be ready. Factory work and site preparation can proceed separately, bringing the two together when the equipment arrives.[1]

The unit being tested gets bigger

NVIDIA’s technical account explains the change in engineering terms. Partners can check an installed deployment unit before dispatch, including its configuration, network connections, sensors and cooling integrity. Transport is part of the design problem: equipment restraints, weight distribution and handling arrangements have to protect the assembled system. Defined checks after arrival establish whether the delivered unit matches the one tested in the factory.[2]

The design also specifies where standardisation ends and local engineering begins. Core containers use an 800-volt direct-current input design. Partners provide the upstream electrical arrangements appropriate to the site. Internal liquid-cooling plumbing can be assembled and tested in the factory; the external heat-removal infrastructure and final commissioning still belong to the receiving location.[2]

For an operator, this changes how a delay is investigated. A fault found during factory testing can be addressed while work continues at the site. A late utility connection still controls when the equipment can run. The useful planning distinction is between shortening equipment integration and shortening the whole project; the two schedules meet at commissioning.[2][1]

A manufacturing plan behind the reference design

Crusoe provides one example of how a partner intends to turn the design into capacity. Its newly announced Spark Liquid module is designed to support up to 3.5 megawatts of computing equipment, including NVIDIA GB300 and Vera Rubin systems. The company says existing Spark modules already host workloads for Ollama. That establishes an operating history for the modular approach, while the new liquid-cooled version remains a production plan.[3]

Crusoe separately confirms that Spark Liquid is joining the DSX POD partner ecosystem and that production is expected to begin in early 2027 at its Brighton, Colorado, factory. The company both manufactures infrastructure and operates a cloud service, making the handover from a built module to customer workloads part of the same business. Those are Crusoe’s plans and operating claims, rather than evidence of delivered DSX POD installations.[4]

Cooling gives the physical design another consequence. In its broader DSX announcement, NVIDIA describes a closed liquid loop that carries heat to outdoor dry coolers, which release it to the air without evaporating water. The company says this can bring facility cooling water consumption close to zero. Local climate determines whether additional mechanical cooling is needed during the hottest periods.[5]

The scope of that water claim is the facility’s cooling system. It does not establish the water footprint of electricity generation or equipment manufacturing. For a site planner, the relevant question is how the proposed heat-removal system performs in that location’s conditions. A standard module still has to operate in a particular climate and alongside a particular energy supply.[5]

For operators, the commercial prize is a shorter interval between paying for computing equipment and putting it to work. Realising that benefit requires the factory schedule and local construction to meet successfully.[1]

Report an error or suggest a correction