1:57:53
NVIDIA GTC Taipei 2026 Keynote | Full Replay
NVIDIA
Overview
NVIDIA CEO Jensen Huang's GTC Taipei 2026 keynote announces the arrival of 'agentic AI,' a new era where AI agents perform complex tasks autonomously. The presentation highlights NVIDIA's advancements in hardware and software, including the Vera Rubin supercomputer designed for agentic AI, the new Vera CPU optimized for AI workloads, and the NVIDIA Agent Toolkit for enterprise AI. Huang emphasizes the transformative impact of these technologies on industries, from chip design to personal computing, underscoring NVIDIA's commitment to building the infrastructure for the AI revolution and its deep ties with Taiwan's manufacturing ecosystem.
How was this?
Save this permanently with flashcards, quizzes, and AI chat
Chapters
- Agentic AI, a more useful and capable form of AI, has arrived, moving beyond generative AI.
- This new wave of AI is significantly boosting productivity, exemplified by a tripling of software commits on GitHub.
- AI is now a profit and GDP generator, leading to increased demand for compute power and a booming ecosystem, particularly in Taiwan.
- The core of agentic AI is the 'agent'—a system combining large language models with a 'harness' for orchestration, memory management, and tool use.
Understanding the shift to agentic AI is crucial as it represents the next major evolution in computing, driving unprecedented productivity and economic growth.
The example of GitHub showing a near tripling of commits in early 2026 illustrates the dramatic increase in software development output enabled by agentic AI.
- An agent application consists of a large language model within a harness that orchestrates tasks, manages memory (working and long-term), and utilizes tools.
- Tool use is a key breakthrough, enabling agents to interact with spreadsheets, web browsers, and databases.
- NVIDIA's CUDA-X libraries are presented as essential 'tools' for agents, allowing them to perform complex scientific and engineering tasks.
- The computing model is becoming disaggregated and distributed, with agents running across various components in a data center.
This chapter explains the fundamental architecture of agentic AI and how NVIDIA's existing software libraries are being repurposed as powerful tools for these new AI agents.
Examples of agent capabilities include generating code, creating a CAD file for 3D printing from a description, and generating a GIF animation based on specific instructions.
- Vera Rubin is NVIDIA's most ambitious endeavor, a multi-rack, pod-scale system designed specifically for processing agentic AI.
- It represents a complete system, integrating GPUs, Vera CPUs, advanced storage, networking, and security processors.
- The system is built using extreme co-design principles, involving hundreds of supply chain partners, primarily in Taiwan.
- Vera Rubin aims to maximize efficiency and profitability in AI factories by optimizing power, cooling, and compute utilization.
Vera Rubin is NVIDIA's flagship hardware solution for the agentic AI era, designed from the ground up to handle the immense computational demands of these advanced AI systems.
The Vera Rubin NVL72 is highlighted as the thinking component, while modular compute trays, ConnectX-9, SuperNICs, and BlueField-4 DPUs handle various tasks, all integrated into a highly efficient, liquid-cooled rack system.
- Traditional CPUs were designed for human interaction (seconds-based), but agents require nanosecond-level responsiveness.
- The new Vera CPU is built from the ground up for agents, focusing on high single-threaded performance, massive bandwidth, and energy efficiency.
- Key features include high Instructions Per Clock (IPC), superior bandwidth per core and overall chip bandwidth, and a novel fabric connecting cores at near light speed.
- Vera CPUs are designed to be highly energy-efficient, allowing for greater density in AI factories without compromising power for token generation.
This section introduces a fundamental shift in CPU design, moving from human-centric to agent-centric processing to overcome bottlenecks in AI workloads.
Vera CPU demonstrates significant speedups in real-world applications like SQL (3x faster) and real-time stream processing for the New York Stock Exchange (6x faster), showcasing its revolutionary performance.
- The NVIDIA Agent Toolkit provides the essential components for enterprises to build and deploy AI agents: models, harnesses, tools/skills, and runtimes.
- OpenShell is introduced as a secure, open-source harness for running agents within enterprises, ensuring privacy and controlled access.
- NVIDIA's Nemotron models, like Nemotron 3 Ultra, are open, powerful, and cost-effective, trained on extensive reasoning and tool-use datasets.
- The toolkit enables companies to create 'super agents,' exemplified by the partnership with Cadence for accelerating chip design.
This toolkit empowers businesses to leverage agentic AI by providing the necessary software infrastructure, models, and frameworks to build custom AI solutions.
The partnership with Cadence to create a chip design agent that reduces verification cycles from weeks to hours demonstrates the practical application and immense efficiency gains offered by the Agent Toolkit.
- NVIDIA and Microsoft are collaborating to reinvent the PC for the age of AI, integrating agents directly into the operating system.
- The new PC experience will feature autonomous agents that understand users, can communicate, and perform tasks using local or cloud-based models.
- Introducing RTX Spark: a new chip combining a Blackwell RTX GPU, a custom Grace CPU, and unified memory, designed for AI workloads on personal devices.
- RTX Spark, along with a new Windows platform for agents, aims to create a new personal computing revolution, enabling complex tasks like 3D house design on a laptop.
This chapter signals a paradigm shift in personal computing, transforming the PC from a tool into an intelligent assistant powered by AI agents.
The demonstration of an agent running on RTX Spark assisting in the design of a house, using tools like Rhino and Blender, and generating photorealistic renders, showcases the future of personal computing.
Key takeaways
- Agentic AI represents a fundamental leap in AI capability, moving from generating content to autonomously performing complex tasks and driving significant economic growth.
- NVIDIA is positioning itself as the infrastructure provider for the AI revolution, offering integrated hardware (Vera Rubin) and software solutions (Agent Toolkit).
- The development of specialized hardware like the Vera CPU is critical to meet the unique, high-speed demands of AI agents.
- NVIDIA's CUDA-X libraries are evolving into essential 'tools' that AI agents will leverage to perform sophisticated tasks across various industries.
- The concept of 'AI factories' is central to scaling AI, requiring end-to-end design and optimization from chip to data center infrastructure.
- The personal computer is being reimagined with integrated AI agents, promising a more intuitive and powerful user experience.
- Taiwan's role in NVIDIA's supply chain and manufacturing ecosystem is highlighted as indispensable to bringing these advanced technologies to market.
Key terms
Agentic AIAI FactoryTokensHarness (AI)Large Language Model (LLM)CUDA-X LibrariesVera RubinVera CPUNVIDIA Agent ToolkitRTX SparkExtreme Co-designDisaggregated Computing
Test your understanding
- What is agentic AI, and how does it differ from previous forms of AI like generative AI?
- How does NVIDIA's Vera Rubin system address the computational demands of agentic AI?
- Why was the Vera CPU specifically designed for agents, and what are its key performance advantages over traditional CPUs?
- What are the core components of the NVIDIA Agent Toolkit, and how do they enable enterprises to build AI agents?
- How is the personal computer being reinvented for the age of AI, and what role does NVIDIA RTX Spark play in this transformation?