AI News
NVIDIA GTC Berlin 2026: DSX POD, OpenShell, and Huang's Plan for Europe's AI Factories
Jensen Huang opened NVIDIA's GTC Berlin keynote at the Tempodrom on October 21 with a claim and a warning. The claim: AI is "the beginnings of a new industrial revolution." The warning was aimed at his European audience: a country that only consumes AI, rather than producing it, will not control its own future in the field. Over the keynote, Huang set out what producing AI requires — AI factories that ship in containers, a safety architecture for autonomous agents, and open models that companies and governments can train on their own data. Here is what was announced, and what it means.
The pitch: Europe should produce AI, not just consume it
Huang described AI as a five-layer cake. Energy powers advanced computers; those computers form an AI factory; the factory runs models that generate tokens and digital intelligence; and that intelligence is integrated into every application. The order matters, he argued: "Long before the AI models emerge, the infrastructure has to exist." He compared it to the networking that had to precede the internet and the grid that had to precede electricity.
The European setting was deliberate. NVIDIA has committed nearly $10 billion to Europe's technology ecosystem and works with more than 12,000 partners. Huang opened by asking what L'Oréal, Revolut, Spotify, Merck, Vodafone, Zara, Porsche, Nestlé and HSBC have in common: all are European companies working with NVIDIA on AI. The conference itself runs three days, with more than 100 sessions on agentic AI, robotics, physical AI, open models and accelerated computing.
DSX POD: an AI factory that arrives in containers
The headline announcement was the DSX POD, a reference design for complete Vera Rubin AI factories. Partners build, cable and test the systems inside modular containers and deliver them integrated and validated, so onsite deployment shrinks from six months to just 14 days once the site's utilities are ready. The engineering bottleneck moves from construction to preparation.
NVIDIA also claims its DSX AI Factory Platform uses up to 100% less water for cooling and up to 30% less power for the same AI output. Those are NVIDIA's figures, and they will stand or fall on the first real deployments. But the strategic point is clear: an AI factory becomes something a company or a country can order and switch on, rather than a multi-year construction project.
OpenShell: a sandbox between agents and the real world
If DSX is about building AI capacity, OpenShell is about trusting what runs on it. NVIDIA introduced the OpenShell Reference Architecture for Safe AI Agents, built on NVIDIA OpenShell — an open-source, policy-controlled runtime that uses kernel-level sandboxing to separate an agent from the systems it can affect.
Huang's definition of an agent was precise: a model, plus a harness for orchestration and memory, plus a sandbox, plus knowledge, plus tools. Because an agent's privileges differ at every launch, static permissions are not enough; policy enforcement and monitoring have to adapt dynamically. Alongside it, NVIDIA announced an Agent Knowledge Foundation — models and libraries that connect agents to enterprise data in documents and databases, including Nemotron tools for document parsing, embedding and reranking, and new Kumo models that make predictions from structured data.
Open models, turned into sovereign intelligence
The third pillar was NVIDIA's open frontier model families: Nemotron for language, Cosmos for world foundation models, GR00T for humanoid robotics, Alpamayo for autonomous vehicles, Ising for quantum-classical computing and Earth-2 for physics AI. Companies can post-train them on their own data — turning proprietary data, in Huang's phrase, into "sovereign intelligence." The platform underneath is large: more than 11 million developers build on NVIDIA's CUDA-X accelerated libraries.
The agent toolkit is already at work in industry. Lila Sciences uses agents to design a potential B-cell lymphoma therapy. At Siemens' Erlangen factory, agents test production plans inside a digital twin before anything changes on the floor. Dassault Systèmes uses engineering agents to refine vehicle design.
Telecom, healthcare, cars and robots
In telecom, NVIDIA made its first commercial phone call on T-Mobile's network with Nokia, using AI-RAN after nearly eight years of work. Huang put the scale at 100,000 sites and about 100 gigawatts of distributed data centers. In healthcare, NVIDIA and Siemens Healthineers are developing AI-enabled adaptive radiation therapy: Nemotron agents coordinate treatment planning, Isaac for Healthcare and Omniverse build a digital twin of the device and the patient, and Holoscan on IGX Thor enables real-time adaptive delivery as the patient moves.
In automotive, the Hyperion platform anchors the announcements: Bolt with Lucid, Mercedes-Benz, Volkswagen Commercial Vehicles and Einride showcasing level 3/4 and robotaxi use cases; Volvo Cars adopting Hyperion and the Alpamayo Stack; and next-generation vehicles from Volkswagen's van division, JLR, Volvo and Mercedes-Benz built to be Hyperion-compatible. In robotics, NVIDIA showed Agile ONE SN, a reference humanoid built on Isaac GR00T with a 31-degree-of-freedom body, expected next year and designed to cut researcher setup from weeks to days. Agility Robotics, Figure, KUKA and NEURA Robotics build on the same stack.
The takeaway
Huang closed with a laptop. RTX Spark, launched the week before, packs 1 petaflop of compute and 128GB of memory into a slim machine. He compared it with the DGX-1 he hand-delivered to OpenAI in 2016: "the only difference is that I can hold this." The onstage demo — Hugging Face's Microduck robots, trained via Isaac Sim on RTX Spark — made the same point as the factories and the sandboxes: NVIDIA's argument is that AI capacity should be buildable, shippable and safe to run, at every scale. Whether Europe acts on it will be measured in factories actually built. DSX POD is designed to make that decision easier.