CES 2026 Keynote: Two Platform Shifts at Once as NVIDIA Goes All-In on Physical AI
Watch original →Core Argument
Jensen Huang opens CES 2026 with a sweeping thesis: the computing industry is undergoing two simultaneous platform shifts — not only are applications moving from traditional software to AI-native, but the entire computing stack itself is being reinvented. Software is no longer programmed but trained; it no longer runs on CPUs but GPUs; applications no longer ship as pre-compiled binaries but generate every token and pixel from scratch in real time.
This means roughly $10 trillion of the past decade’s computing infrastructure is being modernized, with hundreds of billions in annual VC funding and R&D budgets from $100 trillion worth of global industries shifting toward AI. Huang uses this macro framework to explain why NVIDIA is simultaneously redesigning six chips, building a full-stack Physical AI platform, and shipping new architectures on a yearly cadence.
Key Takeaways
1. Four AI Milestones from 2025
- Scaling laws hold: From BERT (2015) → Transformers (2017) → ChatGPT (2022) → o1 reasoning models (2023), each generation demands order-of-magnitude more compute
- Agentic AI explosion: Systems capable of reasoning, research, tool use, and planning proliferated in 2025; Huang specifically credits Cursor for revolutionizing software development at NVIDIA
- Physical AI emergence: AI that understands physical-world common sense — object permanence, causality, inertia, friction — began materializing
- Open models reach the frontier: DeepSeek R1 as the first open reasoning model “caught the world by surprise and activated this entire movement”; open models trail closed by ~6 months but closing fast
2. Physical AI: The Three-Computer Architecture
Huang frames Physical AI infrastructure as three computers working together:
| Computer | Function | NVIDIA Product |
|---|---|---|
| Training computer | Train AI models | DGX supercomputers |
| Inference computer | Deploy at edge (cars, robots, factories) | Orin / Thor processors |
| Simulation computer | Simulate physical world feedback, generate synthetic data | Omniverse + Cosmos |
- Cosmos world foundation model: generates physically plausible surround video from traffic simulator output for AV training; language-aligned, understands physical interactions
- Alpamo: NVIDIA’s first reasoning autonomous vehicle AI, trained end-to-end (camera in → actuation out), explains its driving decisions through chain-of-thought reasoning
- Dual AV stack redundancy: Alpamo + classical traceable AV stack run in parallel; a policy safety evaluator decides which controls the vehicle
3. Mercedes-Benz CLA: First Full-Stack NVIDIA AV
- On road Q1 2026 (US), Q2 (Europe), Q3-Q4 (Asia)
- Rated “world’s safest car” by NCAP
- Every line of code, every chip, every system is safety-certified
- Diverse and redundant sensors; dual software stacks
- Huang predicts “in the next 10 years, a very large percentage of the world’s cars will be autonomous or highly autonomous”
4. Vera Rubin Architecture: Extreme Co-Design
Core challenge: Moore’s Law slowing (1.6x transistors/year), but AI models grow 10x/year, token generation 5x/year, token cost drops 10x/year. The solution is extreme co-design — redesigning all 6 chips simultaneously:
- Vera CPU: 88 cores, 176 threads via spatial multithreading; 2x performance per watt
- Rubin GPU: 5x Blackwell floating-point with only 1.6x transistors. Key breakthrough: MVFP4 Tensor Core — a dynamically adaptive precision Transformer engine, not a simple FP4 data format
- NVLink 6 Switch: 400 Gbps SerDes; 240 TB/s bandwidth per rack (2x global internet)
- Spectrum X Photonics: First silicon-photonics-integrated Ethernet switch (TSMC CoUPE process)
- Bluefield 4: Revolutionary KV cache context memory — adds 16 TB context memory per GPU
| Metric | Blackwell → Vera Rubin |
|---|---|
| Peak inference | 5x |
| Peak training | 3.5x |
| Transistor count | 1.6x |
| Liquid cooling temp | 45°C (no chillers needed) |
| Assembly time | 2 hours → 5 minutes |
5. Enterprise AI and Industrial Partnerships
- Enterprise AI: Integrated with Palantir, ServiceNow, Snowflake, CrowdStrike; agentic systems become the new platform interface, replacing Excel and command lines
- Industrial AI: Deep integration with Siemens (CUDA-X, Physical AI, Omniverse across full industrial lifecycle); partnerships with Cadence and Synopsis to bring AI into chip design
Notable Quotes
“You no longer program the software, you train the software. You don’t run it on CPUs, you run it on GPUs.”
“Some $10 trillion of the last decade of computing is now being modernized to this new way of doing computing.”
“DeepSeek R1, the first open model that’s a reasoning system. It caught the world by surprise and it activated literally this entire movement.”
“In the next 10 years, I’m fairly certain a very large percentage of the world’s cars will be autonomous or highly autonomous.”
“It is impossible to keep up with those kind of rates unless we deploy aggressive extreme co-design — basically innovating across all of the chips across the entire stack all at the same time.”
Analysis
This is a quintessential Jensen Huang keynote — extremely high information density with technical specifics woven into industry narrative. The most significant signal isn’t any single product launch, but NVIDIA’s explicit strategic repositioning: from a chip company to a full-stack AI infrastructure company. The decision to simultaneously redesign all six chips speaks to NVIDIA’s conviction that Moore’s Law alone cannot sustain AI’s compute demand curve — only system-level co-design can maintain leadership. Alpamo for autonomous driving and the Siemens industrial AI partnership mark the inflection point where Physical AI moves from concept to production-scale deployment.