Jensen Huang Unveils Super Computing New Era, Vera Rubin Becomes Game-Changer in AI Field

Jensen Huang once again took the CES stage. The NVIDIA CEO showcased a grand technological release that demonstrated the future direction of AI computing for the entire industry. Unlike previous years, this year’s highlight was not a traditional consumer-grade graphics card but a 2.5-ton enterprise supercomputing platform called Vera Rubin, which integrates six custom chips. Named after astronomer Vera Rubin, it symbolizes NVIDIA’s continued exploration of new frontiers in the AI universe.

Vera Rubin Reconstructs Chip Design Logic, System Innovation Behind Performance Leap

In traditional R&D models, NVIDIA usually follows a conservative strategy of innovating only 1-2 chips per generation. But Vera Rubin broke this pattern by designing and mass-producing six new chips at once—an unprecedented move in the industry. Jensen Huang called this “Extreme Collaborative Design”—synchronized innovation across chip architecture and the entire platform.

These six chips each have their roles: the Vera CPU uses NVIDIA’s custom 88-core Olympus architecture, with 1.5TB of system memory (tripling the previous Grace), enough to support more complex AI applications; the Rubin GPU is the performance powerhouse, reaching 50 PFLOPS of inference capability, five times that of the previous Blackwell, with only a 1.6x increase in transistor count—showing a leap in design efficiency.

Beyond these “brains,” Vera Rubin also integrates ConnectX-9 network cards, BlueField-4 data processing units, NVLink-6 switch chips, and Spectrum-6 optical Ethernet chips. Each component is meticulously crafted to ensure the entire system operates as a cohesive organism.

In practical performance metrics, the Vera Rubin architecture-based NVL72 system achieves 3.6 EFLOPS in inference tasks—five times the Blackwell generation; training performance reaches 2.5 EFLOPS, a 3.5x increase. Notably, such performance gains come with only a 1.7x increase in transistor count, indicating major breakthroughs in semiconductor process and architecture optimization.

Physical design advances are equally impressive. Previously, supercomputing nodes required 43 cables, took 2 hours to assemble, and were prone to errors. Vera Rubin nodes need no cables—only six liquid cooling pipes—and can be assembled in 5 minutes. The main NVLink network behind the nodes spans nearly 3,200 km of copper wires and 5,000 copper cables, transmitting data at 400Gbps. The engineering complexity prompted Huang to joke that “you probably need a very strong CEO to move it.”

Memory Revolution and Network Acceleration Address Key AI Bottlenecks

A major pain point in AI applications is insufficient context memory. During dialogue tasks, AI generates “KV Cache”—its working memory. As conversations deepen and models grow larger, traditional high-bandwidth memory (HBM) quickly becomes a bottleneck.

Vera Rubin’s solution is deploying BlueField-4 processing units to independently manage KV Cache. Each node is equipped with four BlueField-4s, each with 150TB of context memory, allowing each GPU to gain an additional 16TB of memory—an extraordinary leap from the original ~1TB of integrated memory per GPU. Crucially, this expansion does not reduce access speed; the 200Gbps bandwidth remains constant.

But memory alone isn’t enough. When working memory needs to span dozens of servers and tens of thousands of GPUs, network infrastructure becomes the new bottleneck. Spectrum-X, NVIDIA’s first “AI Generative-Specific” end-to-end Ethernet platform, was introduced. Using TSMC’s COOP process and integrated silicon photonics, it supports 512 channels × 200Gbps transmission rates.

Economically, for a data center costing $5 billion, Spectrum-X can boost throughput by 25%, saving about $500 million—Huang described this network system as “almost free.”

Additionally, Vera Rubin also advances secure computing. All data—during transmission, storage, and computation—is encrypted, including PCIe channels, NVLink, CPU-GPU communication, and other buses. Enterprises can deploy their models externally without worrying about data leaks.

Open-Source Wave and AI Democratization: Next-Gen Models Reshape Industry Ecosystem

A highlight of the conference was Huang’s enthusiasm for the open-source AI community. He specifically mentioned last year’s breakthrough release of DeepSeek V1, calling it the “first open-source inference model” that ignited a wave of innovation industry-wide. Slides showed that domestic models Kimi and DeepSeek V3.2 rank first and second globally among open-source models.

Huang admitted that current open-source models may lag behind industry-leading solutions by about six months, but new breakthrough models emerge every half-year. This rapid iteration keeps startups, tech giants, and research institutions on their toes. NVIDIA itself recognizes it cannot afford to fall behind.

Thus, NVIDIA is no longer just a chipmaker. They have built a multi-billion-dollar DGX Cloud supercomputing cluster, developed cutting-edge applications like the protein synthesis model La Proteina and OpenFold 3, and are expanding an open ecosystem centered on healthcare, physical AI, agents, robotics, and autonomous driving.

The Nemotron series models are also notable, covering speech, multimodal, retrieval-augmented generation, and security. These models perform excellently on various authoritative leaderboards, with many enterprises already deploying them in real-world applications.

AI Physical Embodiment Alpamayo: Autonomous Driving Enters Inference Era

If large language models solve the “digital world” problem, Huang’s next ambition is to conquer the “physical world.” He proposed a “Three Core Computers” architecture to advance physical AI: training computers (based on traditional GPU systems), inference computers (edge deployment in robots or autonomous vehicles as “little brains”), and simulation computers (virtual training environments provided by Omniverse and Cosmos).

The implementation of this architecture is Alpamayo—NVIDIA’s first globally deployed autonomous driving system with reasoning capabilities. Unlike traditional rule-based systems, Alpamayo is an end-to-end deep learning system that addresses the “long tail problem” of autonomous driving.

When encountering unprecedented complex traffic scenarios, Alpamayo doesn’t just follow pre-programmed instructions but can reason and make decisions like a human driver. Even better, it “tells you what it’s going to do next and why.” During live demos, the vehicle demonstrated remarkable ability to break down complex traffic scenes into basic common sense to respond effectively.

Mercedes CLA became the first commercial application of Alpamayo technology. Huang announced that this model will be launched in the US in Q1 of this year, followed by Europe and Asia. The vehicle received the highest NCAP safety rating, thanks to NVIDIA’s unique “Dual Safety Stack”—when the end-to-end AI model lacks confidence in road conditions, the system switches immediately to a more conservative traditional safety mode, ensuring safety.

Robot Ecosystem and Industrial Future: From Virtual to Real Manufacturing

Another major focus was the robot strategy. NVIDIA showcased robots from partners, including humanoid robots and Boston Dynamics’ quadrupeds. Huang emphasized that all robots will be equipped with Jetson microcomputers, trained within the Omniverse-based Isaac Simulator environment.

Even more ambitiously, NVIDIA is integrating this ecosystem into the workflows of industrial giants like Synopsys, Cadence, and Siemens. Huang sees the largest robot as the factory itself.

From bottom to top, NVIDIA’s future blueprint involves AI-driven acceleration of chip design, system design, and factory simulation. A surprise appearance was Disney’s robots, which Huang humorously told, “You will be designed in computers, manufactured in computers, and even tested and certified in computers before facing real gravity.”

If you only watch the latter half of the event, you might mistake it for a product launch from a robotics or model company.

From Chip Supplier to AI Enabler: Huang’s Strategic Turn

Against the backdrop of AI bubble debates, Huang’s deeper logic in this CES presentation is intriguing. Besides the slowdown of Moore’s Law dampening traditional performance optimization, he clearly aims to demonstrate AI’s real value through platform-level products like Vera Rubin—shifting focus from raw compute to real-world applications.

This strategic shift from hardware seller to AI enabler extends beyond chips. NVIDIA is deeply investing in applications and ecosystems—supporting open models, building DGX Cloud, developing Alpamayo autonomous driving, physical AI models, and more—using tangible products and use cases to tell the story of how AI truly changes the world.

A final interesting detail: due to time constraints at CES, Huang prepared many slides that couldn’t be shown live. He made a humorous video compiling these unseen slides, again showcasing his personal style—serious about technological innovation, yet relaxed about industry development.