Nvidia CEO Jensen Huang talks about his company's latest innovations at CES 2026

Key Concepts

• AI as a New Computing Platform: The shift from traditional programming to AI training represents a fundamental platform shift comparable to the advent of cloud and mobile computing.
• Open vs. Frontier AI: The importance of open-source AI models is emphasized as a catalyst for broader innovation, rapidly approaching the capabilities of proprietary “frontier” models.
• Vera Rubin Architecture: A revolutionary, fully redesigned system comprising six custom chips designed to overcome the limitations of Moore’s Law and meet the exponentially increasing demands of AI.
• Physical AI & Agentic Systems: The expansion of AI beyond language models into understanding and interacting with the physical world, exemplified by robotics and autonomous vehicles.
• Infrastructure Redesign: The necessity of innovating across the entire chip stack – CPU, GPU, networking, memory – simultaneously to achieve significant performance gains.

---

The AI Revolution & Platform Shift (Part 1)

Jensen Huang’s CES 2024 keynote framed the current moment as a simultaneous platform shift – moving beyond cloud and mobile to AI – and a fundamental change in how software is developed, shifting from programming to training. This shift is estimated to drive a $10 trillion modernization of computing and hundreds of billions in annual VC funding. Huang argued AI isn’t just an application, but a new platform for applications, necessitating a complete reinvention of the computing stack, with GPUs replacing CPUs as the primary processing unit.

He traced the evolution of AI, highlighting key milestones: 2015 (BERT), 2017 (Transformers), 2022 (ChatGPT), 2023 (Reasoning Models – GPT-01) introducing “test-time scaling,” and 2024-2025 (Agentic Systems) demonstrating reasoning, planning, and tool use. Two major branches of AI were identified: Large Language Models (LLMs) and Physical AI – AI that understands and interacts with the laws of nature, termed “AI Physics.”

Huang emphasized the importance of open models, citing DeepSeek R1 as a catalyst for innovation, currently six months behind frontier models but rapidly closing the gap. Nvidia is investing in building its own AI supercomputers (DGX clouds) to develop open models, showcasing projects like Protein Folding (La Protina, OpenFold 3, EVO 2), Earth-2 AI (Forecast Net, Cardiff), Neimotron (Neimotron 3 in development), Cosmos, and Groot.

Nvidia unveiled Alpamo, an end-to-end trained autonomous vehicle AI, open-sourcing both models and training data. The Mercedes-Benz CLA, powered by Alpamo, has achieved the highest safety rating from NCAAP. To address increasing computational demands, Nvidia introduced the Vera Rubin architecture, comprising six custom-designed chips: Vera CPU, Reuben GPU, Connect X9, Bluefield 4 DPU, MVLink switch, and Spectrum-X Ethernet Photonics, delivering 100 petaflops of AI performance.

Vera Rubin: A Deep Dive into the New Architecture (Part 2)

The Vera Rubin platform represents a significant leap in AI infrastructure. The core is the Reuben pod, comprising 1152 GPUs across 16 racks, each rack containing 72 Vera Rubin units, each with two GPU dies. This generation necessitated a complete redesign of six chips due to the slowing of Moore’s Law and the exponential growth in model size (10x larger) and token generation (5x per year).

The Vera CPU boasts 88 cores utilizing “spatial multi-threading” (effectively 176 threads) and exceptional I/O performance, delivering twice the performance per watt of leading CPUs. The Vera Rubin GPU increases single-threaded performance, memory capacity, and chip size, achieving a 5x increase in floating-point performance with only a 1.6x increase in transistor count compared to Blackwell, attributed to “extreme code design.”

A novel innovation is the MVF FP4 tensor core, dynamically adjusting precision based on transformer layer requirements, offering superior throughput and precision retention. The system is housed in a revolutionized MGX chassis, reducing assembly time to 5 minutes with 80% liquid cooling. Inter-rack communication relies on Spectrox Nick, leveraging Nvidia’s Melanox acquisition and featuring programmable RDMA. Spectrum-X, an AI-optimized Ethernet, has made Nvidia the largest networking company globally, delivering 25% higher throughput and potentially saving $5 billion in a $50 billion data center.

Bluefield 4 manages KV cache (key-value cache) – AI’s working memory – offloading virtualization, security, and networking tasks, providing each GPU with an additional 16 terabytes of memory at 200 gigabits per second. The MVLink 6 switch enables every GPU to communicate with every other GPU simultaneously, handling 240 terabytes per second of bandwidth. The system utilizes two miles of shielded copper cables for the MVLink spine and is designed for confidential computing with encryption across all buses.

Performance projections indicate a 10x increase in training throughput for a 10 trillion parameter model, reducing training time to one month, and a 10x increase in factory throughput, with the cost of token generation reduced to one-tenth that of Blackwell. The platform leverages TSMC’s new “coupe” process, integrating silicon photonics directly onto the chip with 512 ports at 200 Gbits per second.

---

Conclusion

Nvidia’s CES 2024 keynote unveiled a vision of AI as a transformative computing platform, demanding a complete reimagining of hardware and software. The Vera Rubin architecture represents a bold response to the challenges of scaling AI, prioritizing a holistic redesign across the entire system stack. The emphasis on open models, coupled with the advancements in physical AI and agentic systems, positions Nvidia as a key driver of the next wave of AI innovation, promising significant performance gains, cost reductions, and broader accessibility.