LIVE: CEO Jensen Huang Nvidia GTC Taipei 2026 Keynote

Key Concepts

• Agentic AI: AI systems that can observe, reason, plan, and use tools to perform productive work autonomously.
• AI Factories: Large-scale, highly complex infrastructure designed to generate "tokens" (the unit of AI output) efficiently and profitably.
• Vera Rubin: NVIDIA’s next-generation, multi-rack, pod-scale supercomputing system designed specifically for agentic AI.
• Vera CPU: A new CPU architecture built from the ground up for agents, prioritizing low latency, high bandwidth, and single-threaded performance.
• Extreme Co-design: The methodology of designing chips, racks, networking, and software together to maximize throughput and efficiency.
• NVIDIA OpenShell: A secure, open-source runtime environment for managing AI agents within enterprises.
• Cosmos 3: A frontier foundation model for physical AI, enabling robots to understand and reason about the physical world.
• RTX Spark: A new line of PCs (laptops, desktops, workstations) re-engineered to run agentic AI natively.

---

1. The Arrival of Agentic AI

Jensen Huang declared that "useful AI has arrived" in the form of Agentic AI. Unlike previous AI waves focused on simple generation, agents act as autonomous workers.

• Economic Impact: Software development serves as a primary case study. With AI agents, 30 million software developers are producing nearly three times the output, effectively turning $3 trillion in salaries into $9 trillion in productivity.
• The Computing Pattern: An agent consists of a Large Language Model (LLM) acting as the "brain," a "harness" (orchestrator) acting as the "body," and tools/skills (like CUDA X libraries) that the agent uses to perform tasks.

2. Vera Rubin: The Infrastructure for Agents

Vera Rubin is NVIDIA’s most ambitious system, designed to handle the complexity of agentic workflows.

• Technical Specifications: It features 6 trillion transistors per board, 18,000 components, and utilizes liquid cooling at 45°C.
• Manufacturing Efficiency: Through extreme co-design, assembly time for a rack has been reduced from two hours to five minutes.
• DSX Blueprint: NVIDIA provides the "DSX" reference design to help partners build AI factories that optimize power, cooling, and network integration, ensuring maximum "tokens per watt."

3. Vera CPU: Built for Agents, Not Humans

Traditional CPUs were designed for humans (virtualization, multi-tenancy). Vera is designed for the "impatient" agent.

• Key Features:
  • High IPC (Instructions Per Clock): 10-wide decode engine for superior single-threaded performance.
  • Bandwidth: 3.6 TB/s fabric bandwidth and LPDDR5X memory support.
  • Performance: Delivers 1.8x the agentic sandbox performance compared to traditional x86 CPUs.
• Application: It serves as the "conductor" for the GPU "orchestra," managing memory (KV caching) and tool orchestration.

4. Physical AI and Robotics

NVIDIA is extending agentic capabilities into the physical world.

• Cosmos 3: A foundation model for physical AI that processes pixels, sound, and language to allow robots to reason about their environment.
• Isaac Groot: A reference humanoid robot platform (6 feet, 150 lbs, 31 degrees of freedom) designed to help researchers and universities bypass the difficulty of building robotic hardware from scratch.
• Alpamo2: An open model for autonomous vehicles, enabling cars to "think" and reason through complex traffic scenarios in real-time.

5. Reinventing the Personal Computer (RTX Spark)

In partnership with Microsoft, NVIDIA is reinventing the PC for the first time in 40 years.

• RTX Spark: A new chip architecture (N1X) combining a Blackwell RTX GPU and a custom 20-core Grace CPU.
• Agentic PC: These machines are designed to run agents locally, 24/7, without "meter anxiety" (cloud costs). They are fully compatible with the entire NVIDIA software stack, allowing users to run everything from digital biology simulations to high-end gaming.

6. Synthesis and Conclusion

The core takeaway is that the computing industry has shifted from "application-based" computing to "agentic" computing. NVIDIA has transitioned from being a GPU company to a full-stack AI infrastructure company. By providing the models (Neotron, Cosmos), the hardware (Vera Rubin, Vera CPU), the runtime (OpenShell), and the reference platforms (Isaac Groot, RTX Spark), NVIDIA is positioning itself as the backbone of the next decade of industrial and personal computing. The overarching goal is to maximize "tokens per watt," as compute is now directly synonymous with revenue and profit.