NVIDIA introduces DSX POD reference design for AI factories

NVIDIA introduced DSX POD at GTC Berlin as a reference design for complete Vera Rubin AI factories, to be built by partners in modular containers and validated before delivery.[1] NVIDIA describes the architecture as standardized, with factory-built container types and defined interfaces.[2] The approach shifts integration and installation of computing, networking, cabling, rack-level power and liquid cooling to the manufacturing facility.[1] DSX POD is a reference design that will ship from NVIDIA’s manufacturing partners.[5]

NVIDIA says DSX POD can reduce deployment time at a prepared site from six months to as little as 14 days.[1] Units arrive ready to connect to the site’s power, cooling, networking and other supporting infrastructure.[1] The time-saving claim therefore does not mean customers can skip site preparation: NVIDIA says customers still need prepared infrastructure and must make the connections onsite.[3] Conventional data-center construction and designs integrated onsite remain alternatives alongside DSX POD.[2] Teams assessing the timeline should distinguish the work shifted to the factory from work that remains at the site.

NVIDIA says the design’s initial focus is large-scale, greenfield, liquid-cooled AI infrastructure, including NVIDIA Cloud Partners.[3] The company says DSX POD AI factories are to be delivered worldwide in early 2027.[1] NVIDIA is working with partners including Anduril, Armada, Bull, Crusoe, Dell Technologies, HPE, Lenovo, Schneider Electric and Vertiv.[1] NVIDIA says DSX POD units are validated before delivery.[1]

Crypto Briefing reports that a DSX POD holds seven Vera Rubin NVL72 racks, together containing 504 Rubin GPUs and 252 Vera CPUs, and has an enclosure roughly the size of a standard shipping container.[4] CNBC reports that NVIDIA puts the pod’s weight at about 28,000 kilograms, while Crypto Briefing reports it contains 17 kilometers of Spectrum-X networking cable.[5][4] NVIDIA has not announced a price; CNBC reports analysts estimate a single Vera Rubin rack could cost more than $7 million.[5] Huang said a 1-gigawatt AI factory costs about $50 billion to build and could generate around $40 billion in annual revenue, and warned a three-month delay could mean roughly $10 billion in lost revenue opportunity.[4]