📺 Today’s recommended deep-dive video: https://www.youtube.com/watch?v=tUE2RV9hqWI

Beyond the Token: Jensen Huang’s Vision for the Age of Agentic AI

Jensen Huang returns to Taipei to declare the end of “useful” AI as a mere concept and the beginning of its reality as an economic powerhouse. With the unveiling of the Vera Rubin platform and a new class of “Agent-first” CPUs, NVIDIA is moving beyond simple generation toward autonomous reasoning systems that can use tools and manage memory.

Core Question: How is NVIDIA restructuring the entire computing stack—from chips to PCs to robots—to support the shift from generative AI to autonomous agentic systems?

Highlights

• The Rise of the Agent: AI has transitioned from Large Language Models (LLMs) to Agentic systems that observe, reason, and act using digital tools.
• Vera Rubin Revealed: The successor to Blackwell is a multi-rack, pod-scale supercomputer designed specifically for disaggregated agentic workloads.
• Vera CPU: NVIDIA’s first “Agent-first” CPU prioritizes single-threaded performance and massive bandwidth over traditional core-count scaling.
• RTX Spark: A total reinvention of the PC in partnership with MediaTek and Microsoft, featuring the N1X chip for local agentic processing.

⏱️ Reading time: approx. 12 minutes · Saves you about 109 minutes vs. watching.

Want to take notes while watching? Click the image below and let AI Notebook capture the key points for you 👇

The Arrival of Agentic AI

From Generation to Action

Agentic AI has officially arrived, marking a shift from models that simply generate text to “agents” that perform actual work.

In this new paradigm, the Large Language Model acts as the brain, but it is housed within a software “harness” that manages working memory and tool orchestration.

This shift is creating an explosion in productivity, particularly in software engineering, where the number of code commits is tripling. Jensen Huang highlights that this doesn’t reduce jobs; instead, it makes every developer three times more productive, turning $3 trillion in salaries into $9 trillion of economic value. Because these agents are “impatient” and require immediate tool feedback, they demand a completely different hardware architecture than the human-centric computers of the past.

💡 Digging Deeper

Q: What defines an “Agent” vs. a standard LLM?
A: An agent consists of the model (brain), a harness (body), and access to tools/skills (workshop) that allow it to execute tasks autonomously.

Q: Why does NVIDIA say “more compute equals more revenue”?
A: Because tokens are now profitable units of production; every token generated by an agent in a factory setting directly translates to work performed or services sold.

Q: How does this impact software developers?
A: Rather than replacing them, it amplifies their output, leading companies to hire more engineers to capture the massive productivity gains offered by agentic workflows.

Infrastructure for the Agentic Age

Vera Rubin and the Vera CPU

NVIDIA’s new hardware roadmap is built for “extreme co-design,” moving from individual chips to entire data center racks as the basic unit of compute.

Vera Rubin is not just a GPU; it is a multi-rack, pod-scale supercomputer that integrates GPUs, CPUs, and networking to process the complex loops of agentic thinking.

The standout announcement is the Vera CPU, designed from the ground up for agents rather than humans. Traditional CPUs are built for virtualization and multi-tenant renting, but agents need the lowest possible latency for single-threaded tasks like tool calls and sandbox execution. Vera features the “Olympus” core, which delivers world-class single-threaded performance and three times the bandwidth of current high-performance CPUs. This ensures that the CPU never becomes a bottleneck for the high-value GPUs sitting next to it in the AI factory.

💡 Digging Deeper

Q: Why is disaggregated computing necessary for agents?
A: Agents trigger different parts of a system—thinking on GPUs, using tools on CPUs, and accessing memory via DPUs—requiring a high-speed fabric to tie them together.

Q: What makes Vera Rubin more reliable than previous systems?
A: It features a new PCB midplane design that eliminates thousands of cables and hoses, reducing assembly time from hours to minutes and significantly lowering points of failure.

Reinventing the Personal Computer

The PC as a Personal Robot

The PC is undergoing its most significant reinvention in 40 years, transitioning from a tool for applications to a platform for autonomous agents.

In collaboration with MediaTek and Microsoft, NVIDIA announced the RTX Spark, a new chip category that brings Blackwell-level AI performance to laptops.

This isn’t just a faster processor; it’s a 70-billion transistor powerhouse that runs the entire NVIDIA CUDA software stack locally. This allows users to run “Digital Twins” of themselves or specialized agents for design and gaming without relying entirely on the cloud. Jensen envisions a future where every home has an AI supercomputer—effectively an “R2-D2” or “C-3PO” in the form of a desktop or workstation—running 24/7 to manage a user’s life, security, and professional tasks.

💡 Digging Deeper

Q: What is the N1X chip?
A: It is a custom 20-core Grace CPU built with MediaTek, fused via NVLink to a Blackwell RTX GPU, designed to run 100% of NVIDIA’s AI and graphics software.

Q: How does the agent change the user experience on a PC?
A: Instead of clicking and typing in an application, the user expresses intent to an agent, which then manipulates the software (like Adobe Photoshop or Rhino) to achieve the goal.

Physical AI and Robotics

Teaching Robots to Perceive

The final frontier of the agentic age is Physical AI, where agents inhabit the bodies of cars, factory machines, and humanoid robots.

To solve the massive data problem in robotics, NVIDIA introduced Cosmos 3, a foundation model that can understand the physical world from both first-person and third-person perspectives.

Data is the primary constraint for robots because, unlike text on the internet, physical interactions are hard to scale. Cosmos 3 uses “Compute as Data,” generating physics-accurate synthetic videos and simulations in Omniverse to train robot brains before they ever touch the real world. This stack culminates in the Isaac Groot reference platform—a fully integrated humanoid robot designed for researchers to jumpstart the age of general-purpose robotics.

Key Takeaways

NVIDIA has effectively shifted its identity from a chip designer to an “AI Infrastructure Company.” The announcement of Vera Rubin and Vera CPU signals a departure from general-purpose computing toward a specialized architecture where the software (the Agent) dictates the hardware’s form. By integrating the CPU, GPU, and networking into a single, cohesive “pod,” NVIDIA is optimizing for the “revenue-per-watt” era of the AI factory.

The reinvention of the PC through RTX Spark and the introduction of humanoid reference platforms like Isaac Groot show that NVIDIA aims to own the agentic loop everywhere it exists. Whether in a 100-megawatt data center, a thin-and-light laptop, or a 150-pound humanoid robot, the computing pattern remains the same: a reasoning brain, a secure harness, and a suite of accelerated tools.

Q&A

Q1: What is the significance of the “Vera” name?
A1: Named after astronomer Vera Rubin, it continues NVIDIA’s tradition of naming architectures after legendary scientists, specifically highlighting the scale and complexity of the modern universe of data.

Q2: How does the Vera CPU compare to traditional x86 processors?
A2: It offers roughly 1.8x to 3x performance gains in real-world agentic workloads like SQL processing and stream processing by focusing on single-threaded IPC and internal bandwidth.

Q3: Is the RTX Spark laptop compatible with existing software?
A3: Yes, it is 100% Windows compatible and 100% CUDA compatible, ensuring it runs everything from Microsoft Office to complex scientific simulations.

Q4: What is “Open Shell”?
A4: It is an open-source security harness and runtime for agents that keeps AI models grounded in enterprise security policies and protects user privacy.

Q5: What is the “Nemo-Tron 3 Ultra” model?
A5: It is NVIDIA’s latest open model that is five times faster and 30% cheaper than previous versions, specifically tuned for long-range reasoning and tool use.

Q6: Why is NVIDIA building its own humanoid robot reference design?
A6: To lower the barrier to entry for researchers and universities who would otherwise spend months or years just setting up the hardware and simulation stack.

Q7: How does “Compute as Data” work in Cosmos 3?
A7: Cosmos 3 acts as a world model that predicts future frames of video based on robot actions, allowing an AI to “dream” and learn from millions of simulated physical interactions.