Nvidia unveiled a reference design that lets partners deliver AI computing power in prefabricated shipping containers, slashing installation time from months to as few as 14 days, according to an announcement made at the company's GTC Berlin conference on Wednesday. The DSX POD system arrives at data centers with racks, networking, and liquid cooling already installed and tested at the factory. The approach aims to help operators start generating revenue from AI workloads faster, as long as the facility already has power, cooling, and network infrastructure ready to connect.
Each DSX POD container houses seven Vera Rubin NVL72 racks holding 504 Rubin GPUs, roughly 10 million components, and weighs approximately 28 metric tons, Nvidia vice president of hyperscale and high-performance computing Ian Buck said in a press briefing before the announcement. The units come with about 20 kilometers of copper cabling and 17 kilometers of scale-out networking preinstalled and tested before shipping. Once a container reaches its destination, teams position it, link it to adjacent units, and conduct final burn-in testing. The system operates without water for cooling because Vera Rubin can run on coolant heated to 45 degrees Celsius, allowing dry coolers similar to automotive radiators to replace evaporative cooling towers. For a 2-gigawatt deployment, Nvidia estimates dry cooling can conserve roughly 54 million liters of water daily compared with cooling towers, though operators in some climates may still require chillers on extremely hot days.
"An AI factory creates value when it starts producing tokens — not when construction is complete," Nvidia founder and CEO Jensen Huang said in a statement. Buck characterized the containers as the fundamental building block for new AI factories, designed to slot into data centers as prefabricated components rather than standalone facilities. The company distinguishes DSX PODs from earlier containerized data center projects by other tech firms, noting these units function as modular parts of larger installations rather than self-contained systems.
Seventeen partners including Armada, Bull, Dell, HPE, Schneider Electric, and Vertiv are building the containers, with deliveries scheduled to begin worldwide in early 2027. The design addresses the lengthy timelines that installing and wiring racks in conventional data centers typically require, which can stretch across multiple months. By shifting assembly and testing to the factory, Nvidia positions the DSX POD as a solution for data center operators racing to deploy AI capacity in a market where speed to operational status directly impacts revenue generation. The prefabricated approach also tackles water consumption concerns that have shadowed data center expansion, particularly in regions facing drought or regulatory pressure over resource use.