Nvidia unveils DSX POD AI-factory design and agent infrastructure at GTC Berlin

CEO Jensen Huang told the audience that AI has become a manufacturing industry. “We were early to observe that, in fact, AI is a manufacturing industry,” he said. “Rather than a tool that’s used by us, it is now part of machinery, an infrastructure, factories that are producing intelligence.”

DSX POD holds seven Vera Rubin NVL72 racks, with about 20 kilometers of NVLink copper and 17 kilometers of scale-out networking, all cabled, plumbed and tested before shipment. Nvidia says deployment at a prepared site falls from six months to as little as 14 days. The design is aimed at operators in the hundreds-of-megawatts tier, such as Europe’s neoclouds, sovereign AI projects and telecom companies, rather than replacing gigawatt campuses. Delivery is set for early 2027, and the 14-day figure assumes a site is ready; containers compress integration but do not remove grid interconnection queues.

At the keynote, Huang calculated that a one-gigawatt AI factory costs about $50 billion and should generate roughly $40 billion in annual revenue, so a one-quarter delay loses about $10 billion. “$10 billion worth of revenue will make anybody anxious,” he said. The calculation frames time to first token as a central metric for AI infrastructure economics.

Nvidia also described DSX platform features under the pod. Warm-water, closed-loop cooling lets operators in suitable climates skip evaporative towers, which Nvidia estimates saves 54 million liters of water per day at two gigawatts. DSX MaxLPS directs power to workloads and Nvidia says it delivers the same output with up to 30% less power. DSX Flex automates interactions with grid operators, allowing AI factories to defer or throttle work and return capacity; Nvidia cited a roughly one-gigawatt spike on the British grid when millions of people put kettles on at halftime during a Euro 2020 match between England and Germany.

On agents, Nvidia launched the OpenShell Reference Architecture to run agents safely and expanded its Agent Toolkit with an Agent Knowledge Foundation and a family of structured-data models called Kumo. Huang described an agent as five parts: the model, the harness, the sandbox, knowledge and tools. He said agents are autonomous systems with access to files, resources, networks and tools, and that the privileges granted to them must be controlled.

Nvidia also said it is deepening its push into telecom with Nokia and European operators. The company’s questions have for years centered on GPU allocation and generational speed; at GTC Berlin, Huang spent little of the keynote on silicon and instead focused on deployment, power and security.