Nvidia GTC Berlin 2026: Jensen Huang Full Keynote

In this talk

1. 0:20

   Europe’s AI economy

   Huang says NVIDIA invested nearly $10 billion in Europe over the past year and has more than 12,000 partnerships across the continent. He frames AI as a new industrial sector, with activity spanning energy, computing, data centers, models, and applications. He points to companies including Schneider Electric, Siemens, and Nebius as part of Europe’s growing ecosystem.

2. 7:03

   Building and operating AI factories

   Huang describes NVIDIA as a full-stack platform combining chips, networking, storage, system software, libraries, and models. He presents Vera Rubin as a liquid-cooled system designed for faster assembly and says its performance is 30 times Blackwell’s for the workloads he discusses. NVIDIA DSX combines digital-twin planning, power and cooling controls, and factory-built pods; Huang says DSX Pod could reduce on-site deployment from months to days.

3. 34:50

   Agents need security, knowledge, and tools

   Huang says agent use is driving higher token consumption because agents reason, iterate, and call tools. He outlines an agent setup with a model, harness, secure sandbox, company knowledge, and tools, and introduces OpenShell for enforcing access policies. He also describes using open models and company data to build specialized agents, with examples in drug research, factory operations, and vehicle engineering.

4. 59:56

   Physical AI and robotics

   Huang describes physical AI as systems that understand and act in the physical world, including telecommunications, autonomous vehicles, and humanoid robots. He says NVIDIA’s Hyperion architecture is being adopted by additional vehicle brands, naming Volkswagen, JLR, Volvo, and Mercedes. NVIDIA’s Agile One reference robot, developed with Agile Robots, is planned for release in the first quarter of 2027.

5. 1:10:21

   AI agents on personal computers

   Huang introduces RTX Spark computers as systems for running AI agents and their tools locally, with workloads able to extend to the cloud. He demonstrates a workflow that uses an agent, models, FreeCAD, Blender, Isaac Sim, and cloud training to teach a small robot to navigate. He presents the example as a three-computer process: model training, simulation in a digital twin, and deployment to the robot.