At CES 2026, Nvidia CEO Jensen Huang delivered a keynote that experts may remember as the starting gun for the next phase of the artificial intelligence revolution. Furthermore, moving beyond the generative text and image models that defined the last few years, Huang laid out a comprehensive roadmap for “Agentic AI” – systems capable of reasoning, using tools, and, crucially, interacting with the physical world.
Specifically, he declared that the “ChatGPT moment for robotics” is approaching. Consequently, Huang unveiled a suite of new architectures and platforms designed to bridge the gap between digital intelligence and physical reality.
The Shift to Physical AI
Notably, the central theme of Huang’s presentation highlighted the transition from AI that simply recognizes patterns to Physical AI – intelligence that can navigate, manipulate, and understand the laws of the real world.
“AI is moving into physical applications like robots and autonomous vehicles,” Huang stated. Additionally, he emphasized that the next frontier involves teaching AI “common sense” about physics. To facilitate this, Nvidia deploys digital twins and massive simulation environments where robots learn to walk, grasp, and drive before they ever enter the real world.
The Rubin Platform: Infrastructure for Agents
To power these computationally intensive “agents,” Huang introduced the Rubin Platform, a successor to the Blackwell generation. Significantly, Nvidia describes this new infrastructure as a highly efficient, six-chip architecture designed specifically for the data center.
- The Vera CPU: For instance, at the heart of Rubin lies the new Vera CPU. In contrast to general-purpose processors, engineers designed Vera to handle the massive data throughput and complex reasoning demands of agentic AI.
- Efficiency: Moreover, Nvidia claims the Rubin platform offers significant cost and performance gains. The company aims to reduce the cost of generating AI tokens by up to 10x compared to previous generations. Ultimately, this efficiency proves critical for “inference”—the actual running of AI models—which becomes expensive as agents perform multi-step reasoning tasks.
Cosmos & Alpamayo: Teaching AI to Reason
Regarding software, Nvidia introduced two critical pillars for Physical AI:
- Cosmos Platform: First, the Cosmos Platform targets robotics. Specifically, Cosmos provides foundational world models that help robots understand and predict physical interactions. It serves as a training ground where AI learns the consequences of its actions in a simulated environment that obeys the laws of physics.
- Alpamayo Model: Second, the Alpamayo Model focuses on the autonomous vehicle (AV) sector. Unlike simple lane-keeping systems, developers designed Alpamayo to understand complex traffic scenarios and make “common sense” driving decisions. Additionally, Huang announced that Nvidia’s first AV stack is ready for deployment, with the goal of eventually embedding AI capabilities into every vehicle.
Defining Agentic AI
Crucially, Huang provided a clear definition of what separates this new wave of AI from its predecessors. He explained that Agentic AI systems possess three core capabilities:
- Advanced Reasoning: The ability to “think” through a problem rather than just predicting the next word.
- Tool Use: The capacity to call upon external tools, software, or even other AIs to complete a task.
- Multi-Modal Understanding: The ability to seamlessly process speech, video, and sensory data to build a complete picture of the environment.
An Open Ecosystem and Aggressive Cycle
In addition, the keynote signaled a strategic shift for Nvidia. Specifically, Huang emphasized a commitment to an Open Ecosystem, providing open-source infrastructure and models. This approach allows industries—from healthcare to manufacturing—to build customized AI solutions without locked-in limitations.
Furthermore, the announcement of Rubin confirms Nvidia’s shift to an aggressive annual hardware refresh cycle. Consequently, this pace forces the entire industry to accelerate, pushing the boundaries of hardware efficiency to keep up with the software’s voracious appetite for compute.