The "nuclear bomb" of 2.5 tons at CES 2026: how NVIDIA is reshaping the future of AI with Vera Rubin

At CES 2026, Jensen Huang took something unusual to the stage: not a consumer graphics card, but an entire 2.5-ton AI rack server. This symbolic gesture introduces what can be called the true “nuclear bomb” of the event: the Vera Rubin computing platform, a completely redesigned hardware ecosystem to accelerate the training of next-generation AI models.

Six chips, a single vision: the architecture that challenges Moore’s slowdown

Vera Rubin represents a paradigm shift compared to NVIDIA’s past. While traditionally each generation of processors saw the evolution of only 1-2 chips, this time the company has simultaneously redesigned 6 different components, all already in production.

The reason is simple but profound: Moore’s Law is no longer sufficient. AI models grow tenfold each year, but traditional performance improvements can’t keep up. NVIDIA has therefore chosen “synchronized innovation at every level” of the platform.

The six pillars of the “nuclear bomb” of technology:

The Vera CPU integrates 88 custom Olympus cores, supports 176 threads via NVIDIA’s spatial multithreading technology, and offers 1.5 TB of system memory—three times that of the previous Grace generation. NVLink bandwidth reaches 1.8 TB/s.

The Rubin GPU is the real star: with an NVFP4 inference power of 50 PFLOPS (5 times higher than Blackwell), it contains 336 billion transistors and includes the third Transformer engine that dynamically adjusts precision as needed.

The ConnectX-9 network card supports 800 Gb/s Ethernet with programmable RDMA, while the BlueField-4 DPU has been specifically designed to handle new AI storage architectures, combining a 64-core Grace CPU with 126 billion transistors.

The NVLink-6 switch connects up to 18 compute nodes, allowing 72 Rubin GPUs to operate as a single coherent machine, with an all-to-all bandwidth of 3.6 TB/s per GPU. Finally, the Spectrum-6 optical switch uses 512 channels at 200Gbps each, integrating silicon photonics.

Game-changing performance: from 3.5x to 10x improvement

In the Vera Rubin NVL72 system, the performance leap compared to Blackwell is dramatic. NVFP4 inference reaches 3.6 EFLOPS (+5x), while training hits 2.5 EFLOPS (+3.5x). Available memory nearly triples: 54TB of LPDDR5X and 20.7TB of HBM.

But the most impressive data point concerns efficiency. Despite transistors increasing only 1.7 times (to reach 220 trillion), measured AI token productivity per watt-per-dollar grows 10 times. For a $50 billion data center and one gigawatt of power, this directly doubles revenue capacity.

In concrete terms: training a 100-trillion-parameter model requires only 1/4 of the Blackwell systems, and the cost to generate a token drops to about 1/10 of the previous.

Revolutionary engineering: from assembly to cooling

The “nuclear bomb” hardware isn’t just brute power. Vera Rubin’s engineering solves practical problems that plagued previous systems. Old supercomputing nodes required 43 cables and 2 hours of manual assembly; Vera nodes use zero cables, only 6 liquid cooling tubes, and can be assembled in 5 minutes.

The back of the rack contains nearly 3.2 km of dedicated copper cables: 5,000 cables form the NVLink backbone at 400Gbps. As Jensen Huang ironically observed: “They could weigh several hundred pounds—you need to be a fit CEO for this job.”

The memory bottleneck problem: the BlueField-4 solution

Contemporary AI faces a critical bottleneck: the “KV Cache” (AI working memory) grows with longer dialogues and larger models. Vera Rubin addresses this with BlueField-4 processors integrated into the rack, each equipped with 150TB of context memory.

Each node has 4 BlueField-4s, which distribute memory to the GPUs: each GPU receives an additional 16TB of memory beyond its native 1TB, maintaining a 200Gbps bandwidth without speed compromises.

Spectrum-X: the network designed for generative AI

To enable dozens of racks and thousands of GPUs to work as a single distributed memory, the network must be “big, fast, and stable.” Spectrum-X is the world’s first end-to-end Ethernet platform specifically designed for generative AI.

The latest generation uses TSMC’s COOP process and silicon photonics, with 512 channels at 200Gbps each. Jensen Huang’s calculation is persuasive: improving productivity by 25% is equivalent to saving $5 billion on a $50 billion data center. “You could say this network system is almost free,” he emphasized.

From theory to action: physical AI, robotics, and autonomous driving

The focus of the “nuclear bomb” isn’t just hardware. Jensen Huang emphasized how about $10 trillion of compute resources accumulated over the decade are undergoing a complete modernization, but not just hardware—also a paradigm shift in software.

The “three computers” architecture for physical AI:

Training computers based on GPU-class training architectures like GB300. Inference computers, the “brain” that makes real-time decisions on robots and edge devices. Simulation computers (Omniverse and Cosmos) that generate virtual environments where AI learns physical feedback.

Alpamayo: autonomous driving with reasoning capabilities

On this architecture, NVIDIA introduced Alpamayo, the first autonomous driving system with true reasoning ability. Unlike traditional systems, Alpamayo is fully end-to-end and solves the “long tail problem”—when facing never-before-seen road situations, it doesn’t just run mechanical code but reasons like a human driver.

The Mercedes CLA equipped with Alpamayo will launch in the United States in the first quarter of 2026, followed by Europe and Asia. The system received NCAP ratings as the safest vehicle in the world, thanks to NVIDIA’s “double security stack”: when the AI model has low confidence, the system immediately switches to a traditional safety mode.

The robotics strategy: from Boston Dynamics to Disney

NVIDIA showcased how nine major AI and hardware companies are all expanding into robotics. Each robot will use Jetson computers, be trained in Omniverse’s Isaac simulator, and the technology will be integrated into industrial systems from Synopsys, Cadence, Siemens, and others.

Jensen Huang invited humanoid and quadruped robots from companies like Boston Dynamics and Agility, highlighting a fascinating perspective: the biggest robot is the factory itself. The vision is that chip design, systems, and factory simulations will all be accelerated by physical AI.

Why this “nuclear bomb” now?

In a context where skepticism about a supposed “AI bubble” is growing, Jensen Huang prioritized not just raw computing power but concrete applications. The 2.5-ton “nuclear bomb” is both a symbol and a promise: to demonstrate that AI can truly transform both the digital and physical worlds.

NVIDIA, which previously sold “picks for gold prospectors,” now directly enters the market where competition is fiercest—physical AI, robotics, autonomous driving. As Huang himself suggested: “When war is ongoing, you can keep selling weapons.”

But Vera Rubin’s true innovation isn’t building a more powerful “nuclear bomb” hardware—it’s demonstrating that by synchronizing innovation at every level of the platform, traditional limits can be surpassed.