GTC DC 2025 Keynote (Oct: Sovereign AI / Nokia / Foxconn / Uber)

Category: DC Keynote · Year: 2025 · ▶ Watch

Speakers: Jensen Huang - Founder and CEO, NVIDIA

Switch language → 中文

Segments (12)

  • 00:00 · Introduction & History of Computing
    • A video montage highlighting the history of American innovation, leading up to the AI era.
  • 04:30 · Welcome & Accelerated Computing
    • Jensen Huang welcomes the audience and explains the fundamental shift from general-purpose to accelerated computing.
  • 15:10 · Telecommunications & 6G
    • NVIDIA announces partnerships to build AI-native 6G networks using the new Aerial RAN Computer (ARC).
  • 21:38 · Quantum Computing
    • Introduction of NVQLink and CUDA-Q to integrate quantum processors with GPU supercomputers for error correction and simulation.
  • 30:36 · AI as the New Industrial Revolution
    • Jensen describes how AI is transitioning from being a tool to performing actual work as AI agents.
  • 42:00 · The Three Scaling Laws of AI
    • Explanation of pre-training, post-training, and test-time scaling driving the exponential need for compute.
  • 50:00 · Extreme Co-Design & Blackwell
    • Deep dive into the Blackwell NVL72 architecture and how co-designing the entire stack overcomes the limits of Moore’s Law.
  • 70:50 · Next Generation: Vera Rubin
    • Sneak peek at the upcoming Vera Rubin architecture, slated for 2026, including new compute trays and networking.
  • 81:18 · Open Source Models & Enterprise AI
    • Highlighting the importance of open-source models and NVIDIA’s partnerships with major cloud and enterprise software providers.
  • 88:40 · Physical AI & Omniverse
    • Using Omniverse digital twins to design, simulate, and operate gigascale AI factories and robotic systems.
  • 100:00 · Manufacturing Blackwell in America
    • A video showcasing the complex, automated manufacturing process of Blackwell systems in the US.
  • 113:30 · Humanoid Robotics & Autonomous Vehicles
    • Showcasing advancements in robotics, including a Disney robot, humanoid robots, and the DRIVE Hyperion platform for autonomous vehicles.

Product Announcements (6)

  • [17:30] NVIDIA ARC (Aerial RAN Computer)
    • An AI-native 6G base station computer.
    • specs: Built with Grace CPU, Blackwell GPU, and ConnectX networking to run software-defined wireless and AI processing simultaneously.
    • availability: N/A
  • [26:24] NVIDIA NVQLink
    • A Quantum-GPU Interconnect.
    • specs: Directly connects quantum processors to GPU supercomputers to enable real-time quantum error correction and hybrid computing.
    • availability: N/A
  • [52:00] Grace Blackwell NVL72
    • A rack-scale AI supercomputer.
    • specs: 72 GPUs connected via an all-to-all NVLink switch, acting as a single massive GPU for training and inference.
    • availability: In full production
  • [71:00] Vera Rubin Superchip
    • Next-generation processor for AI factories.
    • specs: 100 PF AI performance, 88 Custom Arm Cores, 2 TB Fast Memory, 6 Trillion Transistors.
    • availability: 2026
  • [73:10] NVIDIA BlueField-4
    • 800G SmartNIC for AI Factories.
    • specs: 64 Core Grace CPU with ConnectX-9, designed for AI data storage acceleration.
    • availability: N/A
  • [91:15] NVIDIA Omniverse DSX
    • A blueprint for Gigascale AI Factories.
    • specs: Allows co-designing of building, power, cooling, and IT infrastructure using digital twins.
    • availability: N/A

Specific Numbers (6)

Timestamp Metric Value Context
52:40 Bandwidth 14.4 TB/s Per-GPU All-to-All bandwidth in the NVLink Switch Tray.
62:20 Performance Gain 10X Performance, throughput, and cost reduction for GB200 NVL72 compared to H200 NVL8.
68:00 CAPEX $632 Billion Projected CSP CAPEX Spend by 2027.
71:00 Transistors 6 Trillion Number of transistors in the Vera Rubin Superchip.
101:00 Components 1.2 Million Number of components in a single GB300 NVL72 rack.
101:00 Transistors 130 Trillion Total transistors in a GB300 NVL72 rack.

Benchmark Claims (1)

  • [62:20] InferenceMAX: 10X Performance
    • vs: H200 NVL8
    • gain: Delivers 10x the performance, throughput, and cost efficiency for Mixture-of-Experts models.

Customer Stories (4)

  • [16:50] Nokia
    • Partnered with NVIDIA to build AI-Native 6G infrastructure.
    • outcome: Developing the Aerial RAN Computer to upgrade millions of base stations globally.
  • [28:45] U.S. Department of Energy
    • Partnered with NVIDIA to build 7 new AI supercomputers.
    • outcome: Advancing national science initiatives using Blackwell GPUs.
  • [86:40] CrowdStrike
    • Built cybersecurity AI agents on NVIDIA platforms.
    • outcome: Created a faster, more responsive defense system against cyber threats.
  • [106:50] Foxconn
    • Used Omniverse digital twins to design a new factory in Texas.
    • outcome: Building a state-of-the-art facility to manufacture Blackwell systems in America.

Key Technologies (3)

  • CUDA: NVIDIA’s parallel computing platform and programming model that enables accelerated computing across various domains.
  • NVLink: A high-speed interconnect technology that allows multiple GPUs to communicate with each other at massive bandwidths.
  • Omniverse: A computing platform that enables developers to build and operate physically based digital twins for simulation.

Demos Shown (3)

  • [11:15] A montage of simulations powered by CUDA-X, including Earth-2 weather modeling and aerodynamics.
    • True
  • [114:00] A lineup of humanoid robots powered by NVIDIA technology standing on stage.
    • True
  • [134:00] A small Disney Research robot walking in a simulated environment and then shown in real life.
    • True

Predictions / Commitments (2)

  • [70:50, 2026] The Vera Rubin architecture will be in full production.
  • [138:30, Future] The future will see a trillion miles a year driven by autonomous vehicles and millions of robotaxis.

Companies Mentioned (5)

Nokia · Google Cloud, Oracle, SAP · Palantir · TSMC, SK hynix, Foxconn · Uber

Notable Quotes (3)

AI is not a tool. AI is work. — Jensen Huang @ 35:50

You can’t just design chips and hope that things on top of it is going to go faster. — Jensen Huang @ 50:10

The age of AI has begun. Made in America. Made for the world. — Jensen Huang @ 141:26

Key Topics

Accelerated Computing · Generative AI · Telecommunications · 6G Networks · Quantum Computing · AI Factories · Blackwell Architecture · Vera Rubin Architecture · Digital Twins · Omniverse · Humanoid Robotics · Autonomous Vehicles · US Manufacturing

Takeaways

  • Accelerated computing has fundamentally replaced general-purpose computing, overcoming the physical limits of Moore’s Law.
  • AI is evolving from software tools into autonomous agents capable of performing complex work and reasoning.
  • The telecommunications industry is moving toward AI-native 6G networks, utilizing software-defined infrastructure like NVIDIA ARC.
  • NVIDIA is bridging the gap between quantum and classical computing with NVQLink, enabling real-time error correction.
  • Sustaining exponential AI performance requires extreme co-design across chips, systems, networking, and software.
  • NVIDIA maintains a rapid one-year innovation rhythm, with the Vera Rubin architecture scheduled for 2026.
  • Physical AI and robotics represent the next major wave, heavily dependent on Omniverse digital twins for simulation and training.
  • NVIDIA is actively supporting US reindustrialization by partnering with manufacturers to build massive AI factories domestically.