Nvidia is bringing factory-style prefabrication to artificial intelligence data centers, a shift expected to sharply shorten the time it takes for massive capital spending to turn into revenue. Data center operators can start running their facilities sooner and reduce the burden of leaving expensive computing equipment idle, while Nvidia is seen extending its reach beyond semiconductors into data center infrastructure as a whole.
The DSX POD, which Nvidia unveiled at GTC Berlin 2026 in Germany, is designed to house seven server racks of Vera Rubin, its next-generation AI computing system, inside a single container, according to CNBC and other outlets on the 21st. The container is similar in size to a standard freight container and weighs about 28 metric tons. Cables linking the server racks and network equipment, along with liquid cooling piping, are installed at the factory and put through operational testing before the unit is shipped by truck.
Seventeen global companies, including Dell Technologies, Hewlett Packard Enterprise and Foxconn, are working together to build and supply the DSX POD. Nvidia provides the standard design, and server makers such as Dell and HPE and manufacturers such as Foxconn build the systems based on it. Nvidia plans to start supplying the systems through its partners early next year. Pricing was not disclosed.
The design also aims to improve power delivery and cooling efficiency. Nvidia has made it possible to use dry cooling, which releases heat into outside air, with the Vera Rubin system. The server racks are supplied with 800-volt direct current power to reduce losses in power transmission.
Conventional data centers require servers and cooling equipment to be brought on site and then installed and connected one by one. The DSX POD, by contrast, completes much of that work at the factory, minimizing on-site labor. Nvidia expects the approach to cut on-site installation time to as little as 14 days from six months. Shortening the period in which costly AI computing equipment sits unused brings forward the point at which it starts generating revenue. The figure assumes a site where supporting infrastructure is already in place, however, and does not include the time needed for site preparation or grid connection.
Nvidia Chief Executive Jensen Huang also stressed in his keynote speech that day how the timing of a data center's start-up affects profitability. "A 1-gigawatt AI data center costs about $50 billion," he said. "If you assume it generates $40 billion in annual revenue, a delay of just three months in starting operations means losing $10 billion in revenue opportunity." Huang added that with the new approach, the time between the arrival of computing equipment and the start of actual token production could be cut to about one month from four to five months, saying it amounts to a complete transformation of the global supply chain and will also change the way data centers are installed and built.
The technology industry read the announcement as a move by Nvidia to expand beyond AI chips and rack-level sales into data center infrastructure as a whole. CNBC noted that Nvidia, once a semiconductor company, now supplies a range of products and solutions needed to run data centers, including networking and storage.