Nvidia GTC Berlin 2026: Jensen Huang on Vera Rubin, OpenShell and DSX Pods

Nvidia CEO Jensen Huang used his GTC Berlin 2026 keynote to pitch Europe on building its own AI economy, unveiling factory-built DSX Pods, the OpenShell agent security runtime and a 30x jump from Blackwell to Vera Rubin. Over nearly 90 minutes on stage he laid out Nvidia’s five-layer model of the AI industry, showed off Vera Rubin hardware and the first co-packaged optics Ethernet switch, defined the five parts of an AI agent, announced new Hyperion robotaxi partners, a reference humanoid robot and the RTX Spark PC, and kept going through a protest in the hall. Watch the full keynote on Nvidia’s YouTube channel.

TLDW

Huang framed AI as a new industrial revolution built in five layers (energy, chips, AI factories, models and applications) and told Europe it should produce AI, not just consume it. Nvidia says it has raised AI performance a million times in ten years, Vera Rubin is 30 times Blackwell, and its new DSX Pods ship whole factory-tested AI factory blocks in containers to cut the time from delivery to first token from four or five months to about one. The software half of the keynote was about agents: a model, a harness, a secure OpenShell sandbox, company knowledge and CUDA-X tools, with Anthropic signed on as a partner. He closed on physical AI, from AI-run cell towers and Hyperion robotaxis to a reference humanoid robot and an RTX Spark PC that matches the first DGX-1.

Thoughts

The most revealing number in the keynote has nothing to do with chips. Huang says a one-gigawatt AI factory costs about $50 billion and should generate tens of billions of dollars a year, so at $40 billion a year, a three-month delay in turning it on is $10 billion of lost revenue. That reframes what Nvidia sells. DSX Pods move the whole rack, cabling and testing job into Nvidia’s supply chain, so customers receive working, burned-in blocks in containers and go from delivery to tokens in about a month instead of four or five. It is a strong pitch, and it also explains why Nvidia keeps absorbing more of the data center: power, cooling, networking and now construction schedules. Every step that saves the buyer time also makes the buyer’s factory more of an Nvidia product, which is the part competitors building only a chip cannot match.

Huang directly rejected the idea that AI means the end of software. He argued the opposite: agents will use specialized software far more than human engineers ever did, so tool vendors will have more users than ever and will turn into agent companies that ship agents tuned for their own products, which he says Cadence, CrowdStrike, Dassault and IQVIA are already building. The Dassault demo, where agents adapt a Volkswagen Golf door for the Tiguan, run Simulia and catch a vibration problem before the engineer picks the final design, makes the case concrete. His line that “those knowledge and tools belong to you” is the real message for European industrial companies. In his framing, the model is a commodity you can rent or fine-tune, and the defensible part is decades of proprietary data and specialized software, which is exactly what Europe has plenty of.

OpenShell is the most practical announcement for anyone already running agents. Nvidia’s own video puts the status quo bluntly: today the only thing stopping an agent from misusing access is politely asking it not to delete files or touch your SSH keys. OpenShell enforces a written policy at the system-call level, keeps secrets with a supervisor outside the sandbox so a compromised agent never sees them, checks policy changes with formal verification, and is open source under Apache 2.0. Sentry adds a hardware-isolated monitor on BlueField-4 that can quarantine an agent in milliseconds. The browser analogy is apt, since early web viruses ran wild until browsers strictly limited what a page could do, and the Anthropic partnership to build Claude agents on OpenShell suggests the frontier labs agree that guardrails inside the model are not enough.

Standardization is the thread that ties the hardware and robotics halves together. Huang credits NVLink 72’s fast progress to the fact that every computer maker builds the same configuration, and he wants to repeat that in cars and robots. Hyperion is a fixed sensor and computer kit for robotaxis that any brand can adopt and plug into ride-hailing networks, and the VW van, Volvo, the next Mercedes and a new JLR model are joining. The humanoid pitch follows the same logic: a robot shape that is probably not perfect for anything but good enough for many things could create the volume and data flywheel that fragmented industrial robotics never had. Nvidia’s Agile One reference robot, due in the first quarter, is a bet that small and mid-sized European manufacturers will teach robots tasks rather than program them.

Key Takeaways

Nvidia says it invested almost $10 billion in European startups and partnerships last year, announced 14 more startup investments in the past week, and has over 12,000 partnerships across Europe.
Huang’s five-layer cake of the AI economy runs from energy to computing to AI factories to models to applications, and he told Europe to produce AI rather than only consume it.
He predicts every future telecom operator will run about 100,000 sites, roughly 100 gigawatts of distributed AI infrastructure.
Nvidia describes itself as an algorithm company that builds chips, with CUDA-X libraries as its crown jewels and cuDNN as the library that started modern AI.
Across six architectures from Pascal to Vera Rubin, Nvidia claims a one-million-fold AI performance gain in ten years, against roughly 100x for Moore’s law.
Blackwell was about 30 times faster than Hopper for agentic and LLM work, and Vera Rubin is about 30 times faster again, in two years.
Vera Rubin compute trays have no cables, are 100% liquid-cooled and assemble in under five minutes, versus about two hours for Grace Blackwell.
Spectrum-X now ships the world’s first co-packaged optics Ethernet switch, with 512 ports at 200 Gb/s, in volume production.
DSX AI factories use closed-loop liquid cooling with coolant up to 45°C and dry coolers that Nvidia says can save 54 million liters of water a day at two gigawatts.
DSX Pods arrive factory-tested in containers, cutting site bring-up from months to days and time to first token from four or five months to about one.
An agent has five parts: model, harness, secure sandbox, knowledge and tools; OpenShell and the hardware-based Sentry secure the sandbox, and Anthropic is a launch partner.
Nvidia’s open models (Nemotron, Cosmos, GR00T, Alpamayo) ship with weights, data and training scripts so companies can post-train them to frontier level on their own data.
Nvidia and Nokia completed the first commercial phone call on T-Mobile’s network running on CUDA, after almost eight years of work.
Hyperion robotaxi partners now include the VW van, Volvo, next-generation Mercedes and a new JLR model, and the Agile One reference humanoid ships in the first quarter.
RTX Spark, built with Microsoft over four years around the N1X chip, delivers one petaflop and 128 GB, the same as the first DGX-1 that Nvidia delivered to OpenAI in 2016.

Chapters

26:55 Nvidia’s 33 Years in Europe

After a long pre-show, Huang opens by listing European companies working with Nvidia, from L’Oréal and Spotify to Porsche and Nestlé. He recalls that Nvidia’s early chips were made with SGS-Thomson in France and assembled in Malta, that Germany’s Elsa brought Nvidia into workstations in 2000, and that work with Audi and Mercedes followed. He says Nvidia invested almost $10 billion in Europe last year and now has more than 12,000 partnerships there.

32:32 The Five-Layer Cake of the AI Economy

Huang describes AI as manufacturing rather than just software: factories that produce intelligence for every industry. His five layers are energy, computing, AI factories, models and applications, and each one creates jobs and growth. He names European players at each layer, predicts telecom networks will become distributed AI infrastructure, and stresses that models span biology, chemistry, physics and robotics, not only language.

36:21 From GPU to Full-Stack AI Factory

Huang walks through how Nvidia grew from a 1993 graphics chip into a full-stack AI factory company spanning Vera Rubin GPUs, NVLink switches, BlueField-4, Spectrum-X networking and the Vera CPU. A one-gigawatt AI factory costs about $50 billion, and running it takes enormous system software plus the CUDA-X libraries. He traces the DGX-1, first delivered to OpenAI ten years ago, through six architectures and a million-fold performance gain driven by what Nvidia calls extreme co-design.

44:28 Vera Rubin Hardware on Stage

Huang pulls a cable-free, fully liquid-cooled Vera Rubin tray from a rack and shows the NVLink switch that joins 72 GPUs, each rated at 20 petaflops, into one giant GPU. Protesters interrupt, and he asks the hall to let him finish, noting about 100 watch parties around the world. He then shows the Vera CPU, with eight instructions per clock and LPDDR memory built for agentic sessions, and the Spectrum-X co-packaged optics switch made with TSMC. He says Vera Rubin is about 30 times Blackwell, which was itself about 30 times Hopper.

51:22 DSX AI Factories and Factory-Built Pods

A video introduces DSX, a reference architecture co-designed with PTC, Dassault Systèmes, Schneider Electric, Jacobs, Siemens and Procore that starts with an Omniverse digital twin of the whole site. It claims 30% lower power at the same performance through dynamic power control, plus dry cooling that saves water. DSX Pods link Vera Rubin NVL72 racks with 17 km of Spectrum-X cable and ship fully wired in containers with 800-volt DC power. Huang ties it to economics: every month of delay on a $50 billion factory costs billions in revenue.

64:35 Agents and the OpenShell Secure Runtime

Huang credits OpenClaw, the open-source agent that took off early in the year, with kicking off the agent boom, which sent token use and GPU demand soaring. He defines an agent’s five elements and says safety comes first, because agents are autonomous systems with access to files, networks and tools. OpenShell enforces least-privilege policies outside the agent, Sentry monitors it from BlueField-4 hardware, and Huang says almost every major lab, OS maker and security company has agreed to adopt it, with Anthropic partnering on Claude agents.

72:18 Open Models, Company Knowledge and CUDA-X Tools

Every company will use both closed frontier models and open models it customizes, Huang says, and Nvidia releases its open models with data and scripts so companies can trust their origins. Partners such as AI21, Baseten, CodeRabbit, Harvey, Scale and Palantir show post-trained open models reaching frontier performance. The Agent Knowledge Foundation adds parsing, embedding and prediction tools for company data, and CUDA-X tools give agents real work to do, shown through Lila Sciences, Siemens’ Erlangen plant and a Dassault car-door design. Huang argues AI will give tool vendors more users than ever.

89:42 Physical AI: Cell Towers, Radiation Therapy and Robotaxis

Huang calls physical AI the next wave and names radio networks as a favorite application, since base stations sense and act through radio waves and beamforming. Nvidia and Nokia just completed the first commercial call on T-Mobile’s network running on CUDA. He highlights adaptive radiation therapy that tracks a moving tumor, then robotaxis: Hyperion standardizes 14 cameras, nine radars, three lidars and ultrasonics on two Nvidia computers so any automaker can join ride-hailing fleets, with Volkswagen, Volvo, Mercedes and JLR adding models.

95:21 Humanoid Robots and the Agile One

Huang argues robotics has been held back by fragmented, task-specific designs, while humanoids could become the high-volume standard that drives a technology flywheel. He says no region builds precision mechatronics like Europe, describes the three computers of physical AI (training, Omniverse simulation and the onboard robot computer), and unveils Agile One, a reference humanoid built with Agile Robots that ships preassembled with software in the first quarter of next year.

100:00 RTX Spark, Micro Duck and the Finale

Huang reintroduces the PC as a personal agent and shows RTX Spark, built with Microsoft around the N1X chip, matching the original DGX-1’s one petaflop and 128 GB, plus a Windows DGX Station with 20 petaflops due at year end. A demo uses an agent on RTX Spark to design, simulate and train a small walking robot in the cloud, and Pollen Robotics’ Micro Duck robots then join him on stage. He recaps the day and closes with a music video and thanks to 33 years of European partners.

Notable Quotes

“You don’t want to just be consuming the AI. You want to be producing the AI.”
Jensen Huang, on why Europe should build every layer of the AI economy

“cuDNN is to AI what SQL is to the world’s database industry.”
Jensen Huang, on the library he says started modern AI

“$10 billion worth of revenues will make anybody anxious.”
Jensen Huang, on the cost of a three-month delay turning on a gigawatt AI factory

“But hope isn’t security. Safety isn’t an aspiration. It’s technology and engineering.”
Nvidia’s OpenShell launch video, on agents that are only asked politely not to misbehave

“Some people said that AI is here, it’s gonna be the end of software. It’s gonna be end of tools company. It’s exactly the opposite.”
Jensen Huang, arguing agents will give tool vendors more users than ever

“The way you build an agent is you ask an agent to build it for you.”
Jensen Huang, introducing the RTX Spark Micro Duck demo

“Every company in Europe will become an AI company.”
Jensen Huang, closing the keynote

Watch the full keynote here.