# NVIDIA CEO Jensen Huang Leaves Everyone SPEECHLESS (CES Supercut) Source: https://www.youtube.com/watch?v=tXGlS460b_w Recap page: https://rapidrecap.app/video/tXGlS460b_w Generated: 2026-01-06T17:10:14.583+00:00 --- ## Quick Overview NVIDIA CEO Jensen Huang revealed the Vera Rubin platform, which is in full production, featuring the Vera CPU and Reuben GPU, delivering significantly higher performance than Blackwell through extreme co-design strategies, including new MVF FP4 tensor cores and revolutionary liquid-cooled MGX chassis architecture. **Key Points:** - NVIDIA is now in full production of the Vera Rubin system, driven by the necessity to advance computation annually to meet the 10x annual increase in AI model size and computation demand. - The Vera CPU delivers two times the performance per watt of the world's most advanced CPUs in a power-constrained world, boasting 88 physical cores with 176 threads using spatial multi-threading. - The Reuben GPU offers 5x Blackwell's floating-point performance while only having 1.6 times the transistor count of Blackwell, achieved through extreme co-design across the entire stack. - A key innovation is the MVF FP4 tensor core, an entire processing unit that adaptively adjusts precision to maximize throughput without sacrificing necessary accuracy in transformer models. - The MGX chassis assembly time reduced from two hours to five minutes, and the system is 80-100% liquid-cooled using water at 45°C, enabling significant data center power savings. - NVIDIA introduced Spectrum X, AI Ethernet, and Bluefield 4 Data Processing Units (DPUs) to handle intense east-west and north-south traffic, with Bluefield 4 providing 150 terabytes of context memory backing store per rack. - The system achieves a 10x throughput improvement over Hopper in factory throughput (revenue potential) and is projected to require only 1/4th the number of systems compared to Blackwell to train a 10 trillion parameter model in one month. **Context:** NVIDIA CEO Jensen Huang presented a supercut of announcements, likely at CES, focusing on the escalating race for AI advancement, which demands exponential increases in computational power annually. The context is driven by the fact that AI models are increasing by a factor of 10 yearly, necessitating aggressive innovation beyond the slowing pace of Moore's Law, leading NVIDIA to implement 'extreme co-design' across their entire hardware and software stack. ## Detailed Analysis The core message delivered by Jensen Huang is that NVIDIA is combating the intense AI competition by radically advancing computation every single year through extreme co-design, exemplified by the newly launched Vera Rubin system, which is now in full production. This system incorporates the Vera CPU, which doubles performance per watt over existing CPUs, and the Reuben GPU, achieving 5x the floating performance of Blackwell with only 1.6x the transistors, demonstrating that performance gains now rely on architectural innovation rather than just transistor density alone. Crucially, they developed the MVF FP4 tensor core, which dynamically manages precision for massive throughput gains. Furthermore, the infrastructure supporting these chips has been completely overhauled: the MGX chassis assembly time dropped to five minutes, and the system utilizes 45°C liquid cooling, which is so efficient it allows data centers to cool supercomputers with hot water, saving significant power. Networking is revolutionized by Spectrum X AI Ethernet and the new Spectrum X silicon photonic switch, which handles data equivalent to twice the world's internet traffic bandwidth. Context memory limitations are addressed by integrating Bluefield 4 DPUs, backing each GPU with an additional 16 terabytes of context memory on the rack fabric. Overall, this complete stack redesign—from chip to chassis to networking—is projected to deliver 10 times the factory throughput revenue potential of Hopper and drastically cut the cost of generating AI tokens. ### Vera Rubin System Unveiling - Vera Rubin is in full production - It features the Vera CPU and Reuben GPU - Performance gains necessitate 'extreme code design' annually. ### Chip Architecture Innovations - Vera CPU offers 2x performance per watt in power-constrained environments - Reuben GPU achieves 5x Blackwell floating performance with only 1.6x transistor count - Introduction of MVF FP4 tensor core for adaptive precision management. ### Infrastructure and Cooling Breakthroughs - MGX chassis assembly time reduced from 2 hours to 5 minutes - System is 100% liquid cooled using 45°C water, eliminating chillers and saving data center power - Power smoothing implemented to handle instantaneous 25% current spikes during 'all reduce' operations. ### Networking and Data Management - Spectrum X is AI Ethernet, enabling larger, low-latency communication - Bluefield 4 DPUs provide 150TB context memory backing store per rack, augmenting HBM - New Spectrum X silicon photonic switch handles 240 TB/s cross-sectional bandwidth. ### Performance Projections - Training a 10 trillion parameter model in one month requires 1/4th the number of Rubin systems compared to Blackwell - Factory throughput (revenue per watt) is projected to be 10 times higher than Blackwell.