NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang

Summary

NVIDIA CEO Jensen Huang’s Berlin GTC keynote framed AI as a new industrial revolution: not simply software running on computers, but an industry that manufactures “tokens”—digital intelligence—through large-scale AI infrastructure. The keynote focused on Europe’s opportunity to build that infrastructure and use AI across business, science, manufacturing, and robotics.

Europe’s AI economy

• Huang said NVIDIA has worked in Europe for 33 years and cited partnerships with companies including L’Oréal, Spotify, Merck, Vodafone, Porsche, and HSBC.
• He said NVIDIA invested nearly $10 billion in Europe over the past year, announced investments in 14 startups during the week of the keynote, and has more than 12,000 partnerships across Europe.
• He described AI as a five-layer system: energy, advanced computing, AI factories, models, and applications. His argument was that regions should aim to produce AI—not just consume it—and that infrastructure must precede the models and applications it supports.
• He predicted that telecom networks could evolve into distributed AI infrastructure by making use of their many sites and data centers.

NVIDIA’s AI infrastructure and products

Huang presented NVIDIA as a full-stack AI factory platform spanning chips, systems, networking, storage, software, libraries, and models.

• Vera Rubin: NVIDIA’s next-generation system, described as in full production. Huang highlighted a cable-free, fully liquid-cooled computer that he said takes less than five minutes to assemble. NVLink switch trays connect 72 GPUs into a large-scale system.
• Spectrum-X: NVIDIA’s Ethernet networking platform. Huang showcased a co-packaged-optics switch, saying it was in volume production and offered 512 ports at 200 gigabits per second.
• DSX and DSX Pod: DSX is a reference architecture for designing and operating AI factories, with digital-twin planning for power, cooling, performance, and construction. The announced DSX Pod packages Vera Rubin systems, networking, and cooling into a factory-integrated, tested system. NVIDIA said this could cut on-site deployment from months to days. The presentation also claimed DSX’s power controls could deliver the same output with 30% less power and described water-saving cooling approaches.
• CUDA and CUDA-X: Huang emphasized NVIDIA’s software libraries and algorithms as core assets. CUDA-X libraries support workloads such as scientific computing, simulation, data processing, and graphics, and can give AI agents practical tools to perform tasks.
• Performance claims: Huang said NVIDIA’s AI performance had improved by one million times over ten years through “extreme co-design” across chips, systems, software, and algorithms. He also compared Vera Rubin’s performance with Blackwell, and Blackwell’s with Hopper; these were keynote claims.

AI agents, enterprise use, and security

Huang described an AI agent as having five components: a model, harness, secure sandbox, knowledge foundation, and tools. The model reasons and plans; the harness coordinates work; company data provides context; and tools let the agent act.

• Open Shell was introduced as an open-source secure runtime for agents. The presentation said it isolates agents in sandboxes, enforces policies on system calls, files, and network access, and routes network requests through a supervisor.
• NVIDIA Sentry, shown running on BlueField 4, was presented as an optional hardware-isolated monitoring layer that can detect suspicious activity, cut an agent’s connection, and preserve tamper-evident records.
• Huang stressed that agents need only the permissions required for a particular task, and that access controls should be enforced and monitored dynamically.
• NVIDIA’s open-model portfolio was described as covering language, biology, robotics, autonomous navigation, quantum computing, and physical-world simulation. Huang said NVIDIA provides model weights, data, and scripts so users can inspect and reproduce models, then customize them with proprietary data.
• For company knowledge, he highlighted tools for parsing documents, embedding and retrieving information, ranking search results, and analyzing structured data.
• The keynote’s examples showed agents being applied to drug discovery, factory planning, and engineering design. In the engineering example, agents used design and simulation software to adapt a car-door design, identify a vibration issue, and refine the result for human review.

Physical AI, robotics, and autonomous vehicles

Huang presented physical AI—systems that understand and act in the real world—as a major next stage of AI.

• Applications discussed included telecom networks that use AI to interpret radio conditions, adaptive robotic radiotherapy, autonomous vehicles, and general-purpose humanoid robotics.
• NVIDIA’s Hyperion autonomous-vehicle architecture combines sensors, computing, and safety systems. Huang said Volkswagen, JLR, Volvo, and Mercedes models were joining the Hyperion-compatible ecosystem for potential robotaxi use.
• He introduced NVIDIA Agile One, a reference robot developed with Agile Robots, intended to give developers a pre-integrated robotics platform.
• Huang described a three-computer workflow for physical AI: train the model, develop and test skills in a digital twin such as Omniverse, then deploy the validated system to a robot or vehicle.

Personal AI computers and demonstration

• RTX Spark was presented as a new compact AI computer line, with a dedicated N1X chip developed in partnership with Microsoft. Huang said the system offers one petaflop of performance and 128 GB of memory.
• He also showcased a Windows-capable DGX Station, claiming 20 petaflops and 754 GB of memory.
• The keynote included a practical demonstration in which an AI agent used FreeCAD, Blender, Omniverse/Isaac Sim, Isaac Lab, and related tools to create a course, simulate and train a small robot, and then deploy its learned behavior to the physical robot.

Guides, tutorials, and reviews

There were no product reviews or formal step-by-step tutorials. The keynote did provide:

1. A conceptual framework for building agents from five components.
2. An explanation of how Open Shell and Sentry are intended to secure autonomous agents.
3. A demonstration of a physical-AI development workflow, from agent-directed scene creation and simulation through cloud training and deployment to a real robot.

Main speaker and sources

• Jensen Huang, NVIDIA founder and CEO, delivered the keynote and product announcements.
• Supporting material came from NVIDIA-produced videos and demonstrations featuring company systems and partners, including Siemens, Lyra Sciences, Dassault Systèmes, and robotics developers.