Jensen Huang unveils the 2.5-ton Vera Rubin platform at CES 2026, determined to conquer physical AI

CES 2026 recently marked a historic milestone in the AI industry as Jensen Huang, CEO of NVIDIA, brought a 2.5-ton “miracle machine” onto the stage. Unlike previous years focused on consumer graphics cards, this time Jensen Huang shifted entirely to enterprise computing systems, unveiling the Vera Rubin platform—a leap forward in NVIDIA’s GPU development history.

Within 48 hours, Jensen Huang appeared at three major events: NVIDIA Live, an AI technology seminar with Siemens, and Lenovo’s TechWorld. This continuous presence is no coincidence—it reflects NVIDIA’s overall strategy to build an AI ecosystem for real-world applications.

Vera Rubin: A Quantum Leap in NVIDIA Chip Architecture

Named after the renowned astronomer, the Vera Rubin platform represents a fundamental innovation. For the first time in NVIDIA’s history, an entire generation of products not only seeks to improve but also redesigns six different types of chips simultaneously, now entering mass production.

The reason for this change is clear: Moore’s Law has slowed down. Traditional approaches no longer keep pace with the AI models’ 10x annual growth. Jensen Huang and NVIDIA have chosen the path of “extreme co-design”—innovating at all levels simultaneously.

The six chips include:

• Vera CPU: 88 custom Olympus cores, 1.5 TB system memory (3x Grace), supporting 176 threads with NVLink C2C bandwidth of 1.8 TB/s
• Rubin GPU: NVFP4 inference power of 50 PFLOPS (5x Blackwell), integrated with third-generation Transformer engine
• ConnectX-9: 800 Gb/s programmable Ethernet with RDMA
• BlueField-4 DPU: Dedicated processor with 64 Grace CPU cores, 150TB of contextual memory
• NVLink-6 Switch: Connects 72 GPUs as a unified cluster
• Spectrum-6: Optical technology with 512 channels, 200Gbps each

Performance Surpassing Expectations

The Vera Rubin NVL72 system delivers staggering figures. In inference tasks, compute power reaches 3.6 EFLOPS—5 times Blackwell. For training, performance hits 2.5 EFLOPS—3.5 times higher.

Memory capacity reaches 54TB of LPDDR5X (3x previous generation), while HBM4 bandwidth hits 1.6 PB/s, a 2.8x increase. Most impressively, despite performance tripling, transistor count only increased by 1.7 times to 220 trillion.

Even from a cost perspective, Vera Rubin demonstrates strength. Training a 100 trillion parameter model requires only 1/4 of Blackwell’s systems, and token creation costs just 1/10. Energy efficiency, measured as throughput per watt and dollar, increases tenfold—meaning a data center worth $50 billion could double its revenue.

Spectrum-X Networking Technology: “Free” $5 Billion Advantage

Huang precisely calculated: Spectrum-X, a dedicated end-to-end Ethernet network platform for generative AI, can boost throughput by 25%, saving $5 billion for a gigawatt data center. He confidently states, “This network system is almost free.”

This technology uses TSMC’s COOP process, integrating silicon photonics, allowing GPUs and BlueField-4 to span tens of thousands of devices while functioning as a single memory pool.

Addressing the “Long Tail” of Context Memory

A major challenge in AI has been KV Cache—key-value memory, or “task memory.” When conversations are long and models large, HBM memory becomes overloaded.

Vera Rubin solves this by deploying BlueField-4 in each server node. Each node has 4 BlueField-4 units, each with 150TB of contextual memory, providing 16TB per GPU without reducing data transfer speed (still 200Gbps).

Enhanced Security

Vera Rubin supports Confidential Computing—encrypting all data during transmission, storage, and computation, including PCIe, NVLink, and CPU-GPU communication. Enterprises can deploy models confidently without fear of data leaks.

Open Source and Agent Trends

Huang emphasized the importance of open-source software. He especially praised DeepSeek V1—the first open-source inference system, which surprised the world. His slides list models like Kimi k2 and DeepSeek V3.2 as top in their fields.

While current open-source models may lag by about six months behind top commercial models, new models emerge every six months. This rapid iteration is why startups, giants, and researchers don’t want to miss out—even NVIDIA.

NVIDIA is building an open ecosystem spanning biomedical, physical AI, agent models, robotics, and autonomous driving. They are also developing advanced models like La Proteina (protein synthesis) and OpenFold 3, along with the multi-billion-dollar DGX Cloud supercomputer.

Physical AI: Conquering the Real World

If large language models solve the “digital world” problem, Jensen Huang’s next ambition is to conquer the “physical world.”

He proposes a “three-core computer” architecture for Physical AI: training computers (built from GPUs), inference computers (cortical modules on robots or cars), and simulation computers (Omniverse and Cosmos).

Alpamayo: A Self-Driving System with Reasoning Capabilities

Based on this architecture, Huang officially announced Alpamayo—the world’s first autonomous driving model capable of genuine reasoning and inference.

Unlike traditional autonomous systems, Alpamayo doesn’t just execute rigid commands. When faced with complex, unseen traffic situations, it can reason like a human driver: “It will tell you what it’s about to do and why.”

Mercedes CLA equipped with Alpamayo technology will debut in the US in Q1 2026, followed by Europe and Asia. The vehicle is rated as the safest car globally by NCAP, thanks to NVIDIA’s unique “dual safety stack” design—when the AI model lacks confidence, the system switches to traditional safety mode.

Robot Strategy: From Boston Dynamics to Factories

All robots will be equipped with Jetson mini computers, trained in Isaac Simulator on the Omniverse platform. NVIDIA is integrating this tech into industrial ecosystems like Synopsys, Cadence, Siemens.

Huang invited humanoid robots, quadrupeds from Boston Dynamics and Agility, and cute Disney robots onto stage. But he emphasized the key point: “The biggest robot is actually the factory.”

In the future, chip design, system design, factory simulation—all will be accelerated by physical AI. Huang even told the robot team: “You will be designed in the computer, manufactured in the computer, and even tested and verified in the computer before facing gravity.”

Jensen Huang’s Comprehensive Strategy

In the midst of AI bubble debates, Huang seems eager to prove what AI can truly achieve. Beyond unveiling Vera Rubin’s power, he invests heavily in applications and software.

From initially making chips for the virtual world, NVIDIA now directly demonstrates and focuses on Physical AI—self-driving, humanoid robots—to step into the physical realm. As Huang states: “Only when the battle begins can the weapons keep selling.”

CES 2026 proved that with Huang at the helm, NVIDIA isn’t just selling chips—they’re building the future of physical AI.