英伟达把 AI 工厂做成产品，把智能体做成工作负载
过去五年围绕 GPU 的叙事正在升级为整套 AI 工厂方案。
UPDATED 16:52 EDT / OCTOBER 21 2026
For most of the past half-decade, the narrative around Nvidia Corp. has largely centered on the graphics processing unit. The questions have been who has allocation, how much faster the next generation is and whether anyone can close the gap.
At GTC Berlin early today, Chief Executive Jensen Huang spent surprisingly little of his keynote on silicon.Instead, he argued that artificial intelligence constraints have shifted to deployment speed, power, security, and the ability to put enterprise data to work, and that Nvidia intends to have an answer for each.
“We were early to observe that, in fact, AI is a manufacturing industry,” Huang told the audience.“Rather than a tool that’s used by us, it is now part of machinery, an infrastructure, factories that are producing intelligence.”
Those comments from Huang tie together most of the massive news payload from Berlin.A summary of the news includes Nvidia introducing DSX POD, a reference design for complete Vera Rubin AI factories built by partners in containers; launching the OpenShell Reference Architecture to run agents safely; expanding its Agent Toolkit with an Agent Knowledge Foundation and a new family of structured-data models called Kumo; and deepening its push into telecom with Nokia and European operators.
Individually, these are product launches. Together, they show Nvidia moving from selling the most important component in AI to selling the operating model for AI.
One of the more interesting moments in the keynote came when Huang did back-of-the-envelope math on stage. A 1-gigawatt AI factory costs about $50 billion and should generate roughly $40 billion in annual revenue. Delay it by a quarter, and about $10 billion in revenue is lost. “$10 billion worth of revenue will make anybody anxious,” he said.
That’s an interesting and correct way to think about AI infrastructure economics. The cost that keeps AI builders up at night is no longer the price of hardware; it’s the cost of hardware sitting idle while crews pull cable. A modern AI build is a construction project, with kilometers of copper and fiber assembled on site by scarce, specialized labor, especially in Europe.
DSX POD embeds that work into the factory and hides the complexity from the customer. Each pod holds seven Vera Rubin NVL72 racks, with roughly 20 kilometers of NVLink copper and 17 kilometers of scale-out networking, all cabled, plumbed, and tested before shipment. Nvidia says that deployment at a prepared site drops from six months to as little as 14 days.
The more interesting question is where this fits.DSX POD won’t replace gigawatt campuses, which still benefit from economies of scale.The sweet spot is the hundreds-of-megawatts tier, where an operator has a site and power but lacks the time or permitting runway for a full land, power, and shell build.
That’s exactly where Europe’s neoclouds, sovereign AI projects and telcos operate.The 14-day figure assumes the site is ready, with delivery set for early 2027.Containers compress integration, but they don’t fix grid interconnect queues.
DSX POD shifts the long pole in the tent; it doesn’t remove it.
The DSX platform beneath the pod is where Nvidia does some of its most important engineering work, which tends to fly under the radar.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 2 gigawatts.
DSX MaxLPS directs power to where workloads need it, and Nvidia says this delivers the same output with up to 30% less power.
The piece I’d watch most closely is DSX Flex.During the Euro 2020 match between England and Germany, millions of Britons put the kettle on at halftime, and the grid saw roughly a one-gigawatt spike.Today, when a grid operator needs a data center to back off, the process is mostly manual, often a phone call.
DSX Flex automates that interaction, allowing AI factories to defer or throttle work and return capacity.
That matters because data centers that behave like flexible loads are easier to connect and more readily accepted by communities. The industry must stop treating data centers purely as consumers on the grid and start treating them as partners to grid operators. Hyperscalers are already moving in this direction; the difference is that Nvidia is trying to make it a standard.
The second half of the keynote shifted from the factory to what runs inside it, increasingly agents rather than models.Huang described an agent as consisting of five parts: the model, the harness, the sandbox, knowledge and tools.
He was clear about the priority.“The agent is an autonomous system that has access to files, resources and networks, as well as tools, and we need to make sure that the privileges we provide this autonomous system are minimal and that we constantly monitor it,” he said.
Much of the AI safety debate has centered on model alignment and guardrails, assuming a well