GTC October 2025 Keynote with NVIDIA CEO Jensen Huang | Live from GTC 2025 in Washington, D.C.: From "AI Factories" to 6G, NVIDIA Renames the Entire Computing Stack

🧭 Introduction

This GTC in Washington, D.C. (October 28, 2025) was no ordinary product launch. It took the idea that "AI = a new kind of industrial infrastructure" all the way up to the level of industry, supply chain, and national strategy.

Jensen Huang laid out 30 years of the CUDA ecosystem, GPU evolution, and co-design spanning from the server room → the rack → the entire factory:

The AI Factory is not just a metaphor. The goal is to design and operate the building, power, cooling, racks, networking, software, and models as one optimizable "giant computer."

https://www.youtube.com/watch?v=lQHK61IDFH4

🚀 Key Content

1. One Rack Is One "Giant GPU": Grace-Blackwell + NVLink-72

   • Architecture highlight: Grace CPU + Blackwell GPU + NVLink-72 form a single compute domain across the entire rack.

   • Two headline goals: performance density and cost per token (TCO).

   • Conceptual shift: split models into "experts" so more experts can communicate with low latency inside the same NVLink domain.

   • Status: GB300 is in volume production; the next-generation Rubin architecture is being readied in the lab.

💡 This isn't simply chaining GPUs together; it jointly optimizes in-network communication, memory bandwidth, and thermal design. The real KPI becomes token/s × latency × cost per token.

2. DSX: The "Digital Twin Operating System" for AI Factories

   1. Omniverse DSX: an end-to-end digital twin covering civil and MEP engineering, data hall main piping, rack layout, thermal and power simulation, all the way to software scheduling.

   2. Partners: Jacobs, Siemens, Schneider, Trane, Vertiv, Bechtel.

   3. Goals: shorten build time, raise first-time go-live yield, and maximize token output per unit of power.

   4. Application: NVIDIA is building an AI factory research center in Virginia.

💬 "Digital first, physical second": without a digital twin, trial-and-error on large infrastructure is extremely costly.

3. Quantum × GPU: NVQLink + CUDA-Q

   • NVQLink connects QPUs directly to GPU supercomputers, enabling real-time quantum error correction (QEC), control calibration, and hybrid simulation.

   • DOE partnership: the Department of Energy will build seven AI supercomputers with NVIDIA.

   • Industry significance: future scientific computing will be a converged "accelerated computing + AI + quantum" architecture.

⚙️ "Quantum doesn't replace classical computing; it gets embraced by the GPU."

4. 6G/RAN Comes Back to the US: NVIDIA × Nokia + ARC

   • ARC (Aerial RAN Computer): combines Grace CPU, Blackwell GPU, and ConnectX NIC.

   • Partner: Nokia. NVIDIA is investing US$1 billion in Nokia to jointly build AI-RAN and 6G.

   • Significance:

     • AI for RAN: reinforcement learning optimizes beams and spectral efficiency in real time.

     • AI on RAN: turns base stations into edge AI inference platforms.

🇺🇸 Rebuilding communications technology to be AI-native, and "Made in America" again.

5. Enterprise and Security Ecosystem: CrowdStrike, Palantir, and the Big Four Clouds

   • Big four cloud integration: AWS, GCP, Azure, and Oracle fully integrate CUDA-X, NIM, and model libraries.

   • Security × data:

     • Building low-latency AI security agents with CrowdStrike.

     • Integrating with Palantir to accelerate ontology data processing.

🧠 Turning "AI capability" into an auditable enterprise service stack.

6. Physical AI: Factories, Robots, and Autonomous Vehicles

   • Three-computer architecture

     1. GB: training

     2. Omniverse: simulation

     3. Jetson/Thor: edge inference

   • Examples

     • Building an AI factory digital twin in Houston with Foxconn.

     • Developing lifelike robot simulation with Disney Research.

     • Launching the Drive Hyperion "Robo-Taxi Ready" platform, partnering with Uber to connect to its fleet network.

🦾 "Get it working in the digital twin first, then take it into the physical world."

🔍 Three Key Observations

1. Performance iteration shifts from "chip generations" to "factory generations"

   • DSX merges hardware and building; the real unit of performance is the token cost of the whole factory.

2. The communications stack becomes AI-native

   1. AI-RAN maps site density to AI workloads, extending from telecom into private networks and factories.

3. Quantum × GPU becomes a shared industry language

   • Quantum is no longer an island but a pluggable partner in the GPU ecosystem.

Implications for Taiwan and the Optoelectronics Supply Chain

• Data center infrastructure enters a long investment cycle

  • Liquid cooling, power distribution, and thermal interface materials will all grow alongside AI factories.

• Optical interconnect becomes even more important

  • NVLink-72-class bandwidth density makes CPO, OCS, and IPoF critical.

• GPU-based RAN drives edge AI applications

  • Private networks and factory AI-RAN are the next wave of hardware design opportunities.

• New demand for quantum control electronics

  • Testing, cryogenic modules, and superconducting control circuits all gain new applications.

Conclusion

This GTC made NVIDIA's central theme very clear:

Use extreme co-design to turn the entire computing stack into one "factory-scale" computer.

From NVLink-72, DSX AI factories, and AI-RAN to quantum, everything revolves around the same core logic:

Convert power and cooling into more valuable tokens.

For us, the key is:

Build "factory thinking" into every product and project process, starting with the KPIs.