NVIDIA Live with CEO Jensen Huang

Welcome to the Stage: Jensen Huang [03:57]

• The computer industry is experiencing a platform shift every 10 to 15 years, moving from mainframe to PC, PC to internet, internet to cloud, and cloud to mobile.
• Currently, there are two simultaneous platform shifts: the move to AI and the fundamental reinvention of the entire five-layer stack of the computer industry.
• AI is no longer just an application; applications are now built on top of AI, requiring a new way of programming, training, and running software on GPUs.
• Computing has been fundamentally reshaped by accelerated computing and artificial intelligence, leading to a modernization of approximately $10 trillion of the last decade's computing.
• This shift involves hundreds of billions of dollars in VC funding and a significant redirection of R&D budgets towards AI.

"Every time the world of applications targets a new platform, that's why it's called a platform shift."

The Last Year in AI: Scaling Laws and Agentic Systems [07:24]

• 2025 was an incredible year for AI, with multiple breakthroughs occurring simultaneously.
• Key advancements include:
  • Scaling Laws: The evolution from early language models (BERT in 2015) to Transformers (2017), the ChatGPT moment (2022), and the introduction of test-time scaling and reinforcement learning for reasoning and skill acquisition.
  • Agentic Systems: Emerged in 2024 and proliferated in 2025, these models can reason, use tools, research, plan, and simulate outcomes, exemplified by systems like Cursor.
  • Physical AI: AI that understands the laws of nature and can interact with the physical world, including AI physics (understanding physical laws) and AI interacting with the physical world.
  • Open Models: The advancement of open-source AI models, exemplified by DeepSeek R1, democratized AI innovation and led to an explosion in downloads and usage.

"Agentic systems are going to really take off from here."

NVIDIA's Frontier AI Model Building and Open Sourcing [12:00]

• NVIDIA has been building and operating its own AI supercomputers (DGX Clouds) for internal use to develop frontier AI models.
• These models are developed "in the open" to enable every company, industry, and country to participate in the AI revolution.
• NVIDIA's contributions span various domains, including digital biology (protein synthesis and structure), weather prediction (Earth 2), and robotics (Groot).
• The company has open-sourced its models and the data used to train them, aiming for trust and derivative development.
• Libraries like Nemo (Physics, Clara, Biono) provide full lifecycle management for AI development, from data processing to deployment, all open-sourced.
• NVIDIA's models are not only open but also world-class, topping leaderboards in intelligence, multimodality (PDF understanding), speech recognition, and retrieval/semantic search.

"NVIDIA is a frontier AI model builder, and we build it in a very special way. We build it completely in the open so that we can enable every company, every industry, every country to be part of this AI revolution."

Building AI Agents and the Future of Applications [16:18]

• AI agents are a groundbreaking area of development, designed to overcome the limitations of early AI models like hallucination.
• The capability for AI to reason, perform research, use tools, and break down problems into steps is crucial.
• Agentic reasoning opens doors to new applications, as AI models no longer need to know everything upfront but can reason about how to solve problems.
• Techniques like reinforcement learning, chain-of-thought, search, and planning are fundamental to this reasoning capability.
• The concept of using multiple AI models simultaneously for complex tasks is highlighted as genius, enabling multi-modal (speech, images, text, 3D) and multi-model capabilities.
• This framework supports multi-cloud and hybrid cloud deployments, allowing AI to operate at the edge, in enterprises, or wherever data is needed in real-time.
• Future AI applications are built on this framework of agentic, multi-modal AIs, which has turbocharged AI startups.

"The ability to reason about do I have to do research? Do I have to use tools? How do I break up a problem into steps?"

Enterprise AI Revolutionized: Agentic Systems as the Interface [20:44]

• The architecture of agentic AIs allows for customizable, domain-specific AI models that are always at the frontier of technology.
• Blueprints integrate this framework into enterprise SaaS platforms, simplifying the creation of personal assistants and other applications.
• The agentic system is becoming the user interface for enterprise platforms, moving beyond traditional interfaces like spreadsheets or command lines.
• NVIDIA is working with leading enterprise platform companies such as Palantir, Service Now, Snowflake, Code Rabbit, CrowdStrike, and NetApp to integrate these AI capabilities.
• This shift means interacting with platforms will become more intuitive, similar to interacting with people.

"This is the basic framework of future applications."

Physical AI: Simulating and Interacting with the World [25:58]

• Physical AI aims to move AI from screens to the physical world, enabling interaction with common sense understanding of how the world works (object permanence, causality, friction, gravity, inertia).
• This requires AI systems to learn the common sense of the physical world, its laws, and to do so from scarce data.
• Simulation is at the heart of developing physical AI, requiring three types of computers: one for training AI, one for inference (at the edge or in robots), and one specifically for simulation.
• NVIDIA's Omniverse provides a digital twin, physically-based simulation world. Cosmos is a world foundation model that aligns language, images, 3D, and action, enabling physical AI skills like generation, reasoning, and trajectory prediction.
• Synthetic data generation, grounded by physics and ground truth, is crucial for training AI when real-world data is insufficient.
• Cosmos generates physically plausible video and motion from simulations, bringing edge cases to life and training AVs and robots for every scenario.

"The chat GPT moment for physical AI is nearly here, but the challenge is clear. The physical world is diverse and unpredictable."

Alpamo: The World's First Reasoning Autonomous Vehicle AI [34:14]

• Alpamo is introduced as the world's first "thinking, reasoning" autonomous vehicle AI.
• It is trained end-to-end, from camera input to actuation, using a combination of human demonstrations and Cosmos-generated data.
• Alpamo not only controls steering, brakes, and acceleration but also reasons about its actions, explaining the rationale and trajectory.
• This reasoning capability is crucial for handling the "long tail" of driving scenarios, where collecting every possible real-world example is impossible.
• The AI decomposes complex scenarios into understandable, normal circumstances it knows how to handle.
• The demonstration shows Alpamo driving in complex, one-shot scenarios with no human intervention.

"Alpo Mayo does something that's really special. Not only does it take sensor input and activates steering wheel, brakes and and acceleration, it also reasons about what action it is about to take."

The Future of Robotics and Industrial AI [46:10]

• Robots are seen as the next major era in robotic systems, coming in various sizes, including humanoid forms.
• NVIDIA's Isaac Sim and Isaac Lab provide robot simulators within Omniverse for training robots.
• Numerous companies are building robots with NVIDIA's technology, spanning industrial, mobile, and humanoid forms.
• Physical AI and AI physics are revolutionizing large physical industries, starting with NVIDIA's core industry: chip design.
• NVIDIA is integrating its technologies (CUDA X, physical AI, agentic AI, Nemo, Neotron) into leading design tools from Cadence, Synopsys, and Siemens.
• This integration aims to create agentic chip and system designers, mirroring the development of agentic software engineers.
• Manufacturing plants will essentially become gigantic robots, with everything from design to manufacturing occurring in computers.

"The industry that made NVIDIA possible. We're I'm just so happy that that now the technology that we're creating is at a level of sophistication and capability that we can now help them revolutionize their industry."

Vera Rubin: The Next Frontier in AI Computation [55:32]

• The computation necessary for AI is skyrocketing, with models increasing by an order of magnitude annually.
• AI inference has become a thinking process, requiring significant computation for reinforcement learning and test-time scaling.
• The demand for NVIDIA GPUs is increasing due to these computational demands and the evolution of AI.
• The Vera Rubin AI supercomputer is named after astronomer Vera Rubin, who discovered dark matter, symbolizing the uncovering of unseen computational needs.
• Vera Rubin is designed to address the fundamental challenge of escalating AI computational requirements, with production underway.
• The system features an architecture of six co-designed chips, including the Vera CPU and Reuben GPU, engineered to work as one for extreme performance.
• It delivers 100 petaFLOPS of AI performance, five times that of its predecessor, with enhanced networking (ConnectX9, Spectrum X Ethernet), DPUs (Bluefield 4), and a revolutionary liquid-cooled design.

"The amount of computation necessary for AI is skyrocketing. The demand for NVIDIA GPUs is skyrocketing."

Vera Rubin Architecture and Performance [59:23]

• The Vera Rubin system is built on extreme chip co-design, featuring a custom Vera CPU and Reuben GPUs that share data faster with lower latency.
• It integrates 17,000 components on a compute board, utilizing high-speed robots for assembly.
• The system delivers 100 petaFLOPS of AI, a five-fold increase over previous generations.
• Key components include ConnectX9 for high bandwidth, Bluefield 4 DPU for offloading, and MVLink 6 switches for massive GPU interconnectivity.
• The design aims for significant improvements in training and inference performance, with Reuben offering approximately 10x higher throughput than Blackwell for training and inference.
• The system emphasizes energy efficiency, using hot water cooling and achieving twice the energy efficiency of previous generations, saving significant data center power.
• It also incorporates confidential computing capabilities, encrypting data at rest, in transit, and during computation.

"This is the building block of the Vera Rubin AI supercomputer."