Nvidia turns the AI factory into a product, and agents into the workload
SiliconANGLE Zeus Kerravala ● Covered by 5 sources
Nvidia says AI is now a factory business, not just a chip business. It’s pushing containers, security, and grid-friendly load shifting to sell the whole build, not just the GPU.
Based on reporting by SiliconANGLE, Zeus Kerravala — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
At GTC Berlin, Jensen Huang spent far less time talking about chips than Nvidia usually does. The pitch was broader and, frankly, more ambitious: the bottlenecks in AI are now deployment speed, power, security, and getting enterprise data into use. Nvidia wants a product for each one.
The headline move is DSX POD, a reference design for complete Vera Rubin AI factories built by partners in containers. Nvidia says a prepared site can go from deployment in six months to as little as 14 days. Each pod carries seven Vera Rubin NVL72 racks, with about 20 kilometers of NVLink copper and 17 kilometers of scale-out networking already cabled, plumbed and tested before it ships. That won’t replace giant campuses, but it does target the middle ground: hundreds of megawatts, where operators have land or power but not much time.
Huang also did the sort of math that makes infrastructure people wince. He put a 1-gigawatt AI factory at about $50 billion in cost and roughly $40 billion in annual revenue, then said a quarter of delay could mean about $10 billion in lost revenue. In that world, the pain is not just hardware cost. It’s hardware sitting idle while crews finish the build. Nvidia’s DSX platform tries to turn that messy construction process into something closer to an appliance.
Power is part of the pitch too. Nvidia says warm-water, closed-loop cooling can save 54 million liters of water per day at 2 gigawatts. DSX MaxLPS is designed to route power where workloads need it, with up to 30% less power for the same output. DSX Flex goes after a different problem: making AI factories behave like flexible grid loads instead of stubborn ones. That matters in places like Europe, where grid queues can be as important as hardware queues.
Then came the security story. Nvidia described an agent as a bundle of five parts: the model, the harness, the sandbox, knowledge and tools. The OpenShell Reference Architecture is built to keep privileges minimal and monitor agents constantly, with Vera CPUs, Vera Rubin models, and BlueField-4 plus Nvidia Sentry acting as an isolated monitor that can cut off an agent in milliseconds. CrowdStrike, Palo Alto Networks, Zscaler, Cisco and others are already lining up support.
The last piece is data, and this is where Nvidia’s argument gets sharper. Huang said enterprises hold an estimated 70% of the world’s data, most of it not ready for agents. Nvidia’s Kumo models are meant to answer many what-if questions against the same structured data, instead of treating every question like a one-off project. That’s a direct shot at the fantasy that AI kills software. Nvidia’s version is more practical: the software and databases that feed agents become more valuable, not less.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI everyone keeps missing: the boring stuff is the moat. Chips are flashy, but the winner is whoever makes power, security and enterprise data look less like a science project. Nvidia has figured out that selling the factory beats arguing over benchmark charts, which is bad news for rivals still acting like this is only about silicon.
Read more about this at: SiliconANGLE