NVIDIA CES 2026 Keynote Recap: Vera Rubin Architecture, Physical AI and the Full-Scale Arrival of Agentic AI
At the just-concluded CES 2026, NVIDIA CEO Jensen Huang took the stage once again in his signature leather jacket and delivered announcements big enough to "reset" the computing industry. The keynote packed in so much it felt like 15 keynotes rolled into one. From the software-defined shift to the all-new Vera Rubin hardware architecture, NVIDIA is telling the world: we are not just a chip company; we are rebuilding the entire AI full stack.
Below is our in-depth recap and technical analysis of the keynote.
1. A Dual Platform Shift: AI Applications and Computing Change at the Same Time
Huang opened by noting that the computing industry goes through a platform shift every 10 to 15 years (PC to Internet, Internet to mobile). What is different this time is that we are facing two platform shifts happening at once:
Applications: apps are no longer "written"; they are built on top of AI.
Computing: software is no longer programmed but trained, and it no longer runs on CPUs but on GPUs.
Over the past year (2025), we watched the rise of agentic systems. AI is no longer just a chatbot but an agent that can reason, plan and use tools. That changes the nature of software development; NVIDIA itself already makes heavy internal use of agentic tools (such as Cursor) to assist with coding.
2. Software Defines the Future: Agentic AI and Open Models
The Agentic AI Blueprint
Future application architectures will be agentic. NVIDIA showed a "blueprint": an architecture that combines multimodal, multi-model and hybrid cloud.
Router: based on the user's intent, the system automatically decides which model is best to call, whether a frontier model in the cloud or an edge model on-premises.
Live demo: Jensen showed a personal-assistant demo that handled a calendar, controlled a robot (Reachy) via API, generated architectural renderings, and even told the dog to get off the couch. A few years ago this was science fiction; now it is "utterly trivial."
The Triumph of Open Models
Jensen highlighted how the arrival of DeepSeek R1 in 2025 shocked the world by proving that open models can deliver strong reasoning. NVIDIA pledged to remain one of the largest contributors to open models, releasing a series of open projects including Cosmos (world models) and Nemotron (hybrid Transformer-SSM models).
3. Physical AI: AI in the Real World and Robotics
AI's next big battlefield is physical AI: AI that understands the common sense of the physical world (gravity, inertia, cause and effect). That requires three computers working together:
Training computer: trains the models.
Inference computer: runs on the robot or vehicle.
Simulation computer: the key piece, using Omniverse to simulate the physical world for data generation and validation.
NVIDIA Cosmos: World Foundation Models
NVIDIA released Cosmos, a foundation model for physical AI. Instead of generating text, it generates video and data that obey the laws of physics. It can turn rough traffic simulations into photorealistic surround-view video, addressing the scarcity of real-world data, especially long-tail edge cases.
Alpamayo: A Self-Driving Brain That Thinks
NVIDIA announced Alpamayo, billed as the world's first self-driving AI that "thinks and reasons."
End-to-end training: from camera input directly to steering output.
Reasoning: it doesn't just drive; it tells you "why" it is doing what it does and predicts future trajectories.
Dual safety stack: Alpamayo drives, but beneath it sits a traditional, traceable AV stack acting as a guardrail, each backing up the other. This is the architecture Mercedes-Benz has adopted.
4. The Hardware Headliner: The Vera Rubin Platform
Named in honor of Vera Rubin, the astronomer whose work revealed dark matter, NVIDIA's next-generation architecture is called Rubin. With model parameters growing 10x a year, Moore's Law can no longer keep up, so NVIDIA took an "extreme co-design" approach and redesigned six chips at once.
Key Specs Decoded
Vera CPU: a CPU designed for AI collaboration, with 88 physical cores (176 threads) and 2x the performance per watt of the previous generation.
Rubin GPU:
Packaged from two dies.
Transistor count rises only 1.6x, but thanks to architectural innovation, floating-point performance rises 5x.
NVFP4 Tensor Core: the key technology, dynamically adjusting precision inside the processor to maximize throughput.
Connectivity and cooling:
NVLink 6 Switch: up to 240 TB/s of bandwidth (about twice global internet traffic).
100% liquid cooling: inlet water temperature up to 45°C (warm-water cooling), meaning data centers no longer need chillers, saving a large amount of energy.
Cable-free design: the new rack design eliminates traditional cables and hoses entirely, cutting assembly time from 2 hours to 5 minutes.
Solving the KV Cache Bottleneck: BlueField-4 and Massive Memory
This was one of the most hardcore technical highlights of the keynote. As AI conversations get longer (larger context windows), GPU HBM can no longer hold the huge KV cache (key-value cache). NVIDIA proposed a revolutionary fix: pairing the BlueField-4 DPU with an in-rack 150TB shared memory pool. This gives each GPU an extra 16TB of fast-access memory dedicated to large context caches, freeing up HBM entirely.
Networking: Spectrum-X and Silicon Photonics
NVIDIA showed a new Spectrum-X switch chip built on TSMC's COUPE process. It is a giant chip integrating silicon photonics, with fiber connecting directly to the chip ("lasers come in directly"), providing 512 ports of 200Gbps connectivity designed for AI east-west traffic.
5. Conclusion: The Triumph of the Full Stack
From the extreme hardware performance of Vera Rubin, to physical AI applications in Cosmos and Alpamayo, to partnerships with Siemens, Cadence and Synopsys bringing AI into industrial design and chip manufacturing, NVIDIA is building a closed loop.
The numbers say it all:
Rubin trains a 10-trillion-parameter model 4x faster than Blackwell.
Factory throughput per watt is up 10x.
Cost per token falls to 1/10.
Jensen closed by stressing that this is not just about chips but about building the infrastructure of the future. Starting in 2026, every car, factory and piece of software will be redefined by AI.
It's a crazy time, and we're just getting started.
Comments