GTC Notes from Nvidia's Jensen Huang Keynote
My hot takes from my notes from inside the Nvidia CEO GTC Keynote
Here are my notes of the key points from inside NVIDIA’s GTC 2025 Keynote
Key Takeaways:
Blackwell Architecture represents the biggest leap in AI computing.
AI Factories will replace traditional data centers.
NVIDIA Dynamo is the AI Operating System for large-scale inference.
Enterprise AI adoption accelerating → Full-stack AI solutions, AI-powered workforce.
Networking & Power Efficiency are key scaling challenges.
AI-driven Robotics & Digital Twins will drive the next wave of automation.
Jensen Huang took the stage at NVIDIA’s GTC 2025 keynote with his characteristic blend of technical mastery, visionary ambition, and a touch of humor. This year’s event underscored not just how fast the AI revolution is moving but how NVIDIA continues to redefine computing itself. The keynote was a masterclass in scaling AI, pushing the limits of hardware, and building the future of AI-driven enterprises.
Blackwell in Full Production: Computation Demand and Reasoning
At the heart of the keynote was the Blackwell system, the latest in NVIDIA’s GPU evolution. This isn't just another generational upgrade; it represents the most extreme scale-up of AI computing ever attempted. With the Grace Blackwell NVLink72 rack, NVIDIA has built an architecture that brings inference at scale to new heights. The numbers alone are staggering:
1 exaflop computing in a single rack
600,000 components per data center rack
120 kW fully liquid-cooled infrastructure
The shift from air-cooled to liquid-cooled computing is a necessary adaptation to manage power and efficiency demands. This is not incremental innovation; it’s a wholesale reinvention of AI computing infrastructure.
Blackwell NVL with Dynamo: 40x Better Performance and Scale-Out
Huang emphasized that AI inference at scale is extreme computing, with an unprecedented demand for FLOPS, memory, and processing power. NVIDIA introduced Dynamo, an AI-optimized operating system that enables Blackwell NVL systems to achieve 40x better performance. Dynamo represents a breakthrough in delivering an operating system software to run on the AI Factory engineered hardware systems. This should unleash the agentic wave of applications and new levels of intelligence.
Dynamo manages three key processes:
Pre-fill Phase: Efficiently reading vast amounts of information.
Key-Value Storage: Optimizing memory access for inference.
Decode Phase: Accelerating response time and token generation.
The takeaway? Dynamo and Blackwell together redefine AI performance, making large-scale inference more efficient and scalable than ever before.
Upcoming AI Infrastructure Product Roadmap: Cloud, Enterprise, and Robotics
Jensen made a point to emphasise the importance of NVIDIA laying out a predictable annual rhythm for AI infrastructure product and technology evolution, covering cloud, enterprise computing, and robotics.
The roadmap includes:
Now: Full-scale production of Blackwell GPUs.
2H 2025: Blackwell Ultra NVL72.
2H 2026: Vera Rubin NVL144 (named after the scientist who discovered dark matter).
2H 2027: Rubin Ultra NVL576 (600kW per rack!).
Each milestone is an exponential leap forward, resetting industry KPIs for AI efficiency, power consumption, and compute scale.
Scaling the AI Network: Spectrum-X & Silicon Photonics
Networking is the next bottleneck, and NVIDIA is tackling this head-on:
Spectrum-X: A “supercharged” Ethernet for AI Factories.
Silicon Photonics: 1.6 terabit per second bandwidth, enabling AI at massive scales.
Micro Mirror Technology: A new NVIDIA-developed transceiver that reduces power consumption for massive GPU networks.
As Huang pointed out, datacenters are like stadiums, requiring short-range, high-bandwidth interconnects for intra-factory communication, and long-range optical solutions for AI cloud scale.
Enterprise AI: Redefining the Digital Workforce
Huang predicted that AI will reshape the entire computing stack, from processors to applications. AI agents will become integral to every business process, and NVIDIA is building the infrastructure to support them.
10 billion digital AI agent workers are coming.
100% of NVIDIA’s operations will be AI-assisted by year-end.
AI-powered coding will replace traditional programming.
This isn’t just about replacing humans; it’s about enabling enterprises to scale intelligence like never before.
The Shift from Datacenters to AI Factories
NVIDIA’s ultimate vision is to move from traditional datacenters to AI factories—self-contained, ultra-high-performance computing environments designed to generate AI intelligence at scale. This transformation redefines cloud infrastructure and makes AI an industrial-scale production process.
Huang’s punchline, “The more you buy, the more revenue you get,” was a comedic yet poignant reminder that AI’s value is directly tied to scale. NVIDIA is positioning itself as the architect of this new era, where investing in AI computing power isn’t an option—it’s an economic necessity.
Storage must be completely reinvented to support AI-driven workloads, shifting towards semantic-based retrieval systems that enable smarter, more efficient data access. This transformation will define the future of enterprise storage, ensuring seamless integration with AI and next-generation computing architectures. Look for key ecosystem partners like Dell Technologies, HPE, and others to step up with new products and solutions for the new AI Infrastructure wave. Michael Dell was highlighted by Jensen in showcasing Dell as having a complete Nvidia enabled set of AI products and systems.
Beyond AI: Reinventing Robotics
Finally, NVIDIA is applying its AI leadership to robotics. Huang outlined a future where general-purpose robots will be trained in virtual environments using synthetic data, reinforcement learning, and digital twins before being deployed in the real world. This marks the beginning of AI-driven automation at an industrial scale.
Final Takeaways
NVIDIA’s GTC keynote wasn’t just about the next wave of GPUs—it was about redefining the entire computing industry. The shift from datacenters to AI factories, from programming to AI agents, and from traditional networking to AI-optimized interconnects positions NVIDIA at the forefront of the AI industrial revolution.
Jensen Huang has set the tone for the next decade: AI isn’t just an application—it’s the future of computing itself. As we have been saying on theCUBE Pod AI Infrastructure has to deliver the speeds/feeds and scale to open the floodgates for innovation in the agentic and new AI applications that sit on top.
Appendix - My Raw Notes
Key Announcements & Themes
Geforce 25-Year Anniversary
Growth in San Jose
Super Bowl 2024 Reference
Industry-Specific Innovations
CUDA Libraries by Industry
New Partnerships & Technologies Cisco, NVIDIA, T-Mobile: Development of 6G full-stack radio network. General Motors: New self-driving car fleet.
Compute & Infrastructure Advancements
Transition from Air-Cooled to Liquid-Cooled Systems 60K components per computer, 600K per rack. 120kW fully liquid-cooled system. 1 Exaflop computing in a single rack.
Compute Node Fits in One "Blade"
How to Scale Up: Blackwell System Grace Blackwell NVLink72 Rack → The most extreme scale-up ever.
Inference at Scale = Extreme Computing
Performance Metrics Breakdown: X-Axis: Tokens per second (Response Time/Speed → Smarter AI). Y-Axis: Tokens per second factory (Throughput → Maximizing AI Factory Output). Key Needs: FLOPS, memory, processing.
Model Evolution: Capture: Traditional LLMs (Fewer tokens). Reasoning models (More tokens).
NVLink and AI Workflow Optimization
Pre-Fill Phase (Reading large amounts of information).
Decode Phase (Generating responses, token-heavy).
NVIDIA Dynamo: AI Factory OS Pre-fill → KVS → Decode Open-source. Strong partner support (e.g., Perplexity AI). Named after "Dynamo," the first instrument that started the electric revolution.
Hardware Efficiency & Energy Optimization
Hopper Performance Metric: 1MW = 100 tokens per second
Future Key to Success: Blackwell GPUs with NVLink8. FP4 computing.
Power-Limited Industry: NVIDIA is optimizing for most energy-efficient architecture.
"Can’t give Hopper away" Joke → Emphasizing rapid tech evolution
AI Factories & Data Center Evolution
NVIDIA’s Business Pitch: Last year the phrase was "The more you buy, the more you save." This year it's been revised: "The more you buy, the more revenue you get."
Datacenters-in-a-Box → AI Factories replacing traditional data centers.
Roadmap & Future Hardware Releases
2025 (Now): Blackwell Full Production
2H 2025: Blackwell Ultra NVL72
2H 2026: Vera Rubin NVL144 Named after Vera Rubin, who discovered dark matter.
2H 2027: Rubin Ultra NVL576 600kW per rack. Revolutionary AI computing scale-out.
Scaling Up & Scaling Out
Formula: FLOPS × Bandwidth = AI Factory Progress.
Key Benchmarks: Watts / (FLOPS × Bandwidth).
Scaling Out with Networking
Key Technologies: InfiniBand & Spectrum-X Spectrum-X = "Supercharged" Ethernet Silicon Photonics (1.6 Terabit/sec)
Challenges in Scaling Millions of GPUs: Power envelope limitations. World’s first micro mirror transceiver built by NVIDIA.
Enterprise AI & Future of Computing
Enterprise Computing Stack Reinvented New processors, OS, applications. Different orchestration and data access models.
AI Agents Powering Digital Workforces 10 Billion AI-powered digital workers. 100% of NVIDIA to be AI-assisted by EOY. AI coding agents will replace human coders.
New Line of AI Computing DGX Station (DGX Spark) 20 Petaflops AI PC Available for enterprise partners.
Reinventing Storage for AI
Semantic-Based Retrieval Systems
Future AI Data Storage Needs
Dell Announcement: Complete NVIDIA AI Compute Lineup
Robotics & Synthetic Data
Building General-Purpose Robots
Challenges in Robotics AI: Action & Control Data Scarcity Synthetic Data for Simulation Reinforcement Learning AI Digital Twins
Future Robotics Systems Require: Mega-scale AI Models. Collaboration among multiple robots. NVIDIA's new robotics AI platform.