Nvidia's DSX POD cuts AI factory build to 14 days

For most of the past half-decade, Nvidia's story has been about the GPU itself — who gets allocation and how much faster the next generation runs. At GTC Berlin, Jensen Huang spent surprisingly little of his keynote on silicon, arguing instead that AI constraints have moved to deployment speed, power, security and putting enterprise data to work. That framing ties together the Berlin announcements from containerized factories to agent security tools, showing Nvidia moving from selling the most important component in AI to selling the operating model for AI.

The DSX POD packaging reflects a specific economic concern Huang laid out on stage: a gigawatt AI factory costs about $50 billion and should generate roughly $40 billion in annual revenue, so a quarter's delay costs about $10 billion. The cost that keeps builders up at night is no longer hardware price but hardware sitting idle while crews pull cable. DSX POD shifts that long pole by pre-cabling and testing racks before shipment, though it does not remove grid interconnect queues. The sweet spot is the hundreds-of-megawatts tier, where operators have a site and power but lack the permitting runway for a full build — exactly where Europe's neoclouds, sovereign AI projects and telcos operate.

Beneath the pod, Nvidia's DSX platform adds warm-water cooling, power direction and DSX Flex, which automates the interaction between grid operators and data centers. On the agent side, Huang described agents as five-part systems needing minimal privileges and constant monitoring, and argued no model can be trained to understand every business's policy requirements — hence deterministic controls enforced outside the model. Nvidia also said physical AI revenue reached nearly $11 billion in the first half of 2026, about 30% higher than all of last year, a figure that supports its case that the technology is now good enough for companies without in-house robotics teams.