Nvidia Unveils Modular AI Factories, Promises 14-Day Deployment

Nvidia unveiled DSX POD on October 21, 2026 at GTC Berlin, a modular container-sized reference design for Vera Rubin AI data centers. The company said pre-integrated systems plug into prepared power, cooling, and networking and can move from shipment to operation in 14 days instead of about six months, letting AI workloads run while permanent facilities are still built. CEO Jensen Huang told Deirdre Bosa that cutting install time from months to weeks would be game-changing, and said in a statement the pods bring factory-built speed so customers can scale capacity and start revenue-generating production faster. Nvidia said DSX closed-loop liquid cooling with dry coolers can cut cooling-water use by up to 100%—about 54 million liters a day at a 2-gigawatt site, weather permitting—and MaxLPS can cut electricity use by up to 30% for the same output. Nscale reported a 49% inference-throughput gain on GB300 racks in Iceland inside a fixed power budget; a UK trial with National Grid and Nebius cut chip-cluster draw by up to 40% during simulated grid strain while critical work continued. European adopters named include Nscale, Siemens, Schneider Electric, and others; delivery with partners such as Dell, Foxconn, HPE, and Schneider is slated to begin early 2027. The Register put seven NVL72 racks and 504 GPUs in each 28-metric-ton container. Nvidia gave no general-availability dates for new DSX software features still in developer preview.