What NVIDIA Revealed at CES 2026: Physical AI, Self Driving Cars, Supercomputer.

Enter the Physical AI era

CEO Jensen Huang positioned this moment as the start of a new phase in AI evolution. Instead of systems that merely generate text, images, or code, physical AI is designed to understand the laws of the real world, reason through complex situations, and take safe, intelligent action. NVIDIA sees autonomous vehicles, robotics, industrial automation, and humanoid machines as the next major frontier for AI-driven growth.

Alpamayo: reasoning-based autonomy

Alpamayo is NVIDIA’s open model family for autonomous driving, built around explicit reasoning rather than pure pattern recognition. It applies chain-of-thought-style decision logic to driving scenarios, allowing the system to explain and adapt its actions instead of reacting blindly to sensor inputs.

The model has been trained on more than 1,700 hours of driving data and uses a 10-billion-parameter architecture. NVIDIA describes Alpamayo as a “thinking” self-driving system designed to handle rare, high-risk edge cases that typically challenge conventional autonomy stacks.

Real-world driving demonstration

To showcase the technology, NVIDIA demonstrated a Mercedes-Benz CLA navigating city streets autonomously. During the demo, the vehicle anticipated hazards such as a ball rolling into the road—recognizing it as a possible signal that a child might follow.

Huang said the upcoming Mercedes-Benz CLA, targeted for mid-decade production, is being engineered to prioritize safety above all else, leveraging Alpamayo’s more human-like judgment and situational awareness.

Rubin and Vera Rubin: rack-scale AI computing

On the data center side, NVIDIA unveiled Rubin, a tightly integrated, six-chip AI platform combining Rubin GPUs, Vera CPUs, networking, and data-processing components into what the company calls a rack-scale AI supercomputer.

According to NVIDIA, Rubin delivers approximately five times the AI training performance of Blackwell while reducing inference token costs to roughly one-tenth. The platform is also optimized for large mixture-of-experts models, requiring far fewer GPUs and significantly lower token costs than previous generations.

A notable absence—and a clear signal

For the first time in nearly half a decade, NVIDIA’s CES keynote did not introduce a new consumer GeForce GPU. The omission reinforced the company’s strategic pivot toward enterprise AI, hyperscale data centers, autonomous vehicles, and robotics.

Taken together, Alpamayo and the Rubin platform outline NVIDIA’s long-term vision: extending AI beyond the cloud and embedding it directly into cars, machines, and physical systems that operate in the real world.