During the 2026 CES exhibition, NVIDIA CEO Jensen Huang delivered the keynote speech titled “NVIDIA Live” in Las Vegas, where he rarely mentioned the GeForce graphics card lineup and instead focused entirely on data centers, autonomous vehicles, robotics, and open models.
In this speech, Huang announced that NVIDIA has partnered with Mercedes-Benz, and the first AI autonomous vehicle has officially entered mass production. The safest Mercedes-Benz CLA, certified by NCAP, is equipped with NVIDIA’s automotive model Alpamayo. It is expected to launch in the US market in Q1, enter Europe in Q2, and potentially land in Asia as early as Q3 this year.
The highly anticipated Vera Rubin has also entered full-scale production. This is NVIDIA’s flagship product after Blackwell, not a single chip, but the name of the next-generation AI platform. A single Rubin Pod consists of 16 cabinets with a total of 1,152 GPUs, with each cabinet housing 72 Rubin units. Each Rubin is actually integrated from two GPU chips. Compared to the Grace CPU and Blackwell GPU, the new generation Vera CPU shows significant improvements in single-thread performance, memory capacity, and overall computational metrics.
Also debuting at this event is the new generation data center processor BlueField-4, capable of partitioning large-scale data centers into multiple independent blocks for simultaneous operation by different users. The new AI data center network core Spectrum-X Ethernet Co-Packaged Optics (CPO) switching platform features key technology developed jointly by TSMC and NVIDIA, based on the innovative COOP (Co-Optimized Optics Packaging) process.
At the model level, Huang emphasized that openness will be the key to the AI ecosystem. NVIDIA has now released fully open models across six major domains, including Clara (medical), Earth-2 (climate), Nemotron (reasoning and multimodal), Cosmos (robotics and simulation), GR00T (embodied intelligence), and the autonomous vehicle model Alpamayo.
NVIDIA’s First AI Autonomous Vehicle to Launch in Q1
At CES 2026, Huang first mentioned Alpamayo, positioning it as the world’s first open visual-language-action (VLA) model with reasoning capabilities. Alpamayo is trained end-to-end, directly linking sensor inputs to steering, braking, and acceleration outputs. Its training data combines extensive real-world driving demonstrations, synthetic data generated by Cosmos, and hundreds of thousands of finely annotated samples. Unlike traditional systems, Alpamayo not only makes decisions but can also explain in real-time why it took certain actions.
Huang announced that NVIDIA’s first autonomous vehicle will be on the road in the first quarter of this year. It is expected to launch in the US in Q1, in Europe in Q2, and as early as Q3–Q4 in Asia.
Structurally, the model layer is Alpamayo, while the application layer is the Mercedes-Benz CLA. This Mercedes-Benz CLA has just passed certification and is now in mass production. It has also received NCAP safety ratings, being rated as the safest car globally. This autonomous vehicle features two modes: one with an AI reasoning-enabled autonomous driving system, and another with a fully traceable, rule-based traditional autonomous system. Built-in safety policies and assessment modules will determine in real-time whether AI should take over or revert to the conservative traditional system.
Huang announced that Vera Rubin has fully entered mass production with detailed specifications.
One of the core focuses of the speech was NVIDIA’s official announcement that the Rubin platform has entered full-scale production. Rubin is positioned as the successor to Blackwell and is NVIDIA’s first co-designed six-chip AI platform, starting from data centers, integrating computing, networking, storage, and software stacks.
Within the NVIDIA ecosystem, Vera Rubin is not a single chip but the name of the next-generation AI platform: the core is to connect Vera CPU and Rubin GPU via NVLink‑C2C into the Vera Rubin superchip, which is then stacked into the Vera Rubin NVL72 rack-scale AI supercomputer, used for agentic AI, long-context inference, and AI factory workloads.
A single Rubin Pod consists of 16 cabinets with a total of 1,152 GPUs, with each cabinet housing 72 Rubin units. Each Rubin is actually integrated from two GPU chips. The Vera CPU, designed specifically for supercomputers, achieves twice the watt-efficiency of existing top-tier CPUs under power-constrained scenarios and features extremely high data transfer rates.
Compared to previous Grace CPU and Blackwell GPU, the new Vera CPU shows significant improvements in single-thread performance, memory capacity, and overall computational metrics. When directly connected to Rubin GPU, it forms a massive, battleship-like AI computing core.
The Rubin platform’s core components include Rubin GPU, Vera CPU, NVLink 6, Spectrum-X Ethernet Co-Packaged Optics (CPO), ConnectX-9 SuperNIC, and BlueField-4 DPU. Under NVFP4 precision, each Rubin GPU can deliver up to 50 petaflops of inference performance. Huang stated that through integrated chassis and network design, Rubin aims to eliminate all bottlenecks and reduce the total cost of ownership (TCO) per AI token to about one-tenth of previous levels.
New Data Center Processor BlueField-4
On the hardware front, NVIDIA also revealed deployment details of Vera CPU and BlueField-4 for the first time. Vera CPU is designed for power-limited scenarios, achieving twice the performance per watt of current top-tier CPUs, with significant improvements in single-thread performance and memory capacity, tailored for AI supercomputers and data centers.
Huang further introduced NVIDIA’s new data center processor BlueField-4, capable of partitioning large-scale data centers into multiple independent blocks for simultaneous operation by different users. It offloads virtualization, security, and north-south network traffic management tasks from the CPU, making it a standard component for each computing node. He also announced NVIDIA’s push for industry-standard system architectures, enabling the entire ecosystem and supply chain to share common components.
Since a single MGX system comprises about 80,000 components, changing specifications annually would cause enormous waste. Therefore, major system manufacturers from Hon Hai, Quanta, Wistron, to HP, Dell, and Lenovo can directly adopt new platforms within their existing manufacturing systems.
It is noteworthy that although Vera Rubin’s total power consumption is twice that of the previous Grace Blackwell, its intake airflow is nearly identical, with cooling water temperature maintained at 45°C. Data centers may not even need chilled water systems, effectively cooling supercomputers with hot water.
Data Center CPO Switching Platform, Key Technology from TSMC COOP Process
The new generation AI data center network core Spectrum-X Ethernet Co-Packaged Optics (CPO) switching platform features key technology developed jointly by TSMC and NVIDIA—the COOP (Co-Optimized Optics Packaging) process.
This process integrates silicon photonic optical components directly into the switch chip package core, rather than using traditional external optical modules. A single switch chip can support up to 102.4 Tb/s of lateral bandwidth expansion and provide up to 512 high-speed connections at 200Gb/s per port.
This article, Jensen Huang CES 2026: Vera Rubin Full Production, AI Autonomous Vehicles Q1 Launch, Key Process from TSMC, first appeared on Chain News ABMedia.
