Watch the full keynote on YouTube: NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang

At NVIDIA GTC Berlin 2026, CEO Jensen Huang took the stage to outline the next phase of the global AI transformation. Highlighting Europe’s pivotal role in the new AI economy, Huang presented a comprehensive vision spanning next-generation datacenter architectures, hardware-enforced agent security, open foundation models, physical AI robotics, and personal AI supercomputers.

1. Europe and the Rise of the 5-Layer AI Economy

Huang opened the keynote by reflecting on NVIDIA’s 33-year history in Europe, noting that early graphics processors like the NV2 and Riva 128 were manufactured and distributed through European partners like SGS Thomson and Elsa. Today, NVIDIA collaborates with over 12,000 European companies and startups, having invested over $10 billion across the region in the past year alone.

Huang framed artificial intelligence not simply as software, but as the engine of a New Industrial Revolution. He introduced a 5-layer framework representing the end-to-end AI economy:

1. Energy & Power Infrastructure: Driven by industrial partners like Schneider Electric and Siemens Energy.
2. Accelerated Computing Hardware: Datacenters, supercomputers, and telecom base stations transformed into AI factories.
3. AI Models & Intelligence: Spanning language, biology, physics, climate science, and robotics.
4. AI-Native Applications & Platforms: Emerging software startups and platforms (e.g., Mistral, Nebius, Mebius).
5. End-Industry Integration: Automotive, manufacturing, drug discovery, healthcare, and financial services.

2. Next-Gen Hardware Architecture: Vera Rubin & Spectrum-X CPO

Pushing performance beyond traditional semiconductor limits, Huang detailed NVIDIA’s Vera Rubin architecture—the successor to Blackwell. Over the past decade, NVIDIA’s “Extreme Co-Design” methodology across chips, systems, and algorithms has delivered a 1-million-fold increase in AI compute performance.

Key hardware announcements included:

• Vera CPU: Designed for heavy agentic AI sessions with an 8-instruction-per-clock core and ultra-high-bandwidth LP-DDR memory.
• Vera Rubin NVL72: A 100% liquid-cooled system connecting 72 GPUs into a unified domain via NVLink switches, delivering 20 PetaFLOPS per GPU.
• Spectrum-X Co-Packaged Optics (CPO): A 512-port, 200 Gbps Ethernet switch integrating silicon photonics directly onto the chip package in partnership with TSMC (CuPy technology).
• Performance Benchmark: Vera Rubin achieves up to 30x higher performance than Blackwell for Large Language Model (LLM) and agentic processing.

3. NVIDIA DSX Architecture & DSX Pods

Building AI factories at scale requires solving thermal, power, and construction constraints. NVIDIA introduced NVIDIA DSX, an open reference architecture for gigawatt-scale AI factories developed with partners including PTC, Dassault Systèmes, Schneider Electric, Jacobs, Siemens, and Procore.

Features include:

• DSX MaxLPS: Closed-loop power management reducing power consumption by up to 30% under peak load.
• Waterless Dry-Cooling: Liquid cooling handling coolant temperatures up to 45°C, eliminating evaporation and saving up to 54 million liters of water daily per site.
• NVIDIA DSX Pod: Factory-assembled, containerized 800V DC power pods that reduce datacenter deployment timelines from 4–5 months down to approximately 1 month.

4. Agentic AI Security: OpenShell & NVIDIA Sentry

As AI evolves from simple chat interfaces to autonomous agents with system execution capabilities, security becomes paramount. Huang introduced OpenShell and NVIDIA Sentry to establish hardware-enforced boundaries for AI agent execution:

• NVIDIA OpenShell: An open-source (Apache 2.0) lightweight runtime environment that isolates AI agents inside secure sandboxes, enforcing network and file system access policies at the Linux kernel level.
• NVIDIA Sentry (on Vera BlueField-4): A hardware-isolated trust domain operating outside the host OS. If an agent attempts an unauthorized action, Sentry intercepts the connection and quarantines the agent within milliseconds.
• Partnership Announcement: Anthropic and NVIDIA collaborated to integrate OpenShell into Claude agentic workflows.

5. Enterprise Knowledge Foundations & Open Frontier Models

Huang emphasized NVIDIA’s commitment to open science by providing open weights, training scripts, and datasets for its specialized foundation models:

• Nemotron: Language and multimodal models.
• Cosmos: World foundation models for physical AI.
• GR00T & Alpamayo: Models for humanoid robotics and autonomous navigation.
• Agent Knowledge Foundation: A suite including Nemotron Parse, Nemotron Embed, Nemotron 3 Omni, and Kumo for real-time predictive analytics across enterprise SQL databases and unstructured documents.

6. Physical AI, Robotics & Autonomous Mobility

Physical AI represents the next major wave of automation, where models reason about gravity, friction, and spatial mechanics.

• NVIDIA DRIVE Hyperion: An open industry standard sensor and compute architecture (14 cameras, 9 radars, 3 LiDARs) for autonomous vehicles.
• Automotive Fleet Expansion: The Volkswagen Bus, Volvo, next-gen Mercedes-Benz, and Jaguar Land Rover (JLR) joined the Hyperion network for autonomous ride-hailing fleets.
• Humanoid Robotics: NVIDIA unveiled Agile One, a turnkey reference humanoid robot developed in partnership with Agile Robots, equipped with onboard spatial compute and pretrained physics skills.

7. Personal Supercomputing: RTX Spark & DGX Station

To close the keynote, Huang introduced a new category of personal AI devices designed to run local agentic workflows:

• RTX Spark: Powered by the custom NVIDIA N1X chip, developed in a 4-year secret project with Microsoft. RTX Spark delivers 1 PetaFLOPS of AI performance and 128 GB of unified memory in a compact desktop/handheld form factor.
• DGX Station (RTX Spark Edition): A desktop workstation running Windows that packs 20 PetaFLOPS of compute and 754 GB of memory.
• Live Demonstration: On stage, Huang used RTX Spark to generate, train, and test motor policies for dual-legged “Micro Duck” robots (by Pollen Robotics) in Isaac Sim before deploying the policy to physical robots live.

Summary Overview

Segment	Announcement / Technology	Key Specs / Highlights
Compute Architecture	Vera Rubin GPU & NVL72	100% liquid-cooled, 30x faster than Blackwell, 72-GPU NVLink domain
Networking	Spectrum-X CPO	Co-packaged optics with TSMC CuPy, 512 ports at 200 Gbps
Datacenter Infrastructure	DSX Pods	Factory-built container pods; cuts deployment time from months to 30 days
Agent Security	OpenShell & Sentry	Kernel & BlueField-4 hardware sandboxing; sub-millisecond agent quarantine
Autonomous Mobility	DRIVE Hyperion Partners	VW Bus, Volvo, Mercedes-Benz, and JLR join autonomous robotaxi network
Robotics Reference	NVIDIA Agile One	Reference humanoid robot co-developed with Agile Robots
Personal AI Hardware	RTX Spark & DGX Station	Desktop AI supercomputer with up to 20 PetaFLOPS compute