NVIDIA GTC 2026: In the era of Intelligent Agent AI, the full-stack strategy has been officially implemented.

On June 1, NVIDIA Founder and CEO Jensen Huang delivered a keynote address at GTC 2026 in Taipei, Taiwan. In this presentation, Jensen Huang highlighted the launch of several new products and platforms, including the Vera Rubin architecture, the CPU Vera for AIAgents, the AI model Nemotron 3 Ultra, and the AI factory platform DSX.

He explicitly pointed out that the generative model sown two years ago… AI The seeds have already borne abundant fruit. ——“ Useful Artificial Intelligence ” It has already arrived, with Intelligent agents as its core hallmark. AI ( AgenticAI ) of the full-scale outbreak. In this speech, Jensen Huang not only showcased a super-powerful hardware cluster designed for mass production of Intelligent agents, but also redefined the entire computing ecosystem—from chips and infrastructure to personal computers.

This talk is not a product launch; it is a systematic declaration of a new paradigm: AI From “ Can be used ” to “ Useful ” , from “ Generated content ” to “ Complete the work ” , from GPU The company arrives at AI In the infrastructure sector, NVIDIA is undergoing a profound transformation.

Intelligent agents drive AI Enter the commercialization closed loop

At the outset, Jensen Huang offered a groundbreaking industry assessment: large language models ( LLM ) is rapidly evolving into one with the capacity to understand, observe, reason, and plan. “ Intelligent System ” . Regarding “AI” Reduce employment ” Regarding that line of argument, he stated frankly. “ It’s complete nonsense. ”。

GitHub The number of code commits has decreased from 2023 year's 3 The number of trillions has been increasing year by year, reaching… 2026 At the beginning of the year, it had already nearly tripled, reaching approximately 9 trillions of times. The underlying business logic behind this is: similarly 3 The trillion-dollar investment in engineer compensation, amplified by Intelligent agents, can now generate nearly 9 Trillions of dollars in productivity value. AI It’s not about抢ing jobs; it’s about exponentially amplifying each engineer’s output. When computing power itself becomes a high‑output economic unit, Tokens Its production subsequently became profitable, which directly boosted global demand for… AI The explosive growth in computing power demand from factories.

VeraRubin Full-scale mass production:

A supercomputer designed specifically for Intelligent agents

The most significant hardware announcement of this presentation is…   VeraRubin The architecture has officially entered full-scale mass production. . This is NVIDIA’s most complex end-to-end rack-level cluster system to date: it integrates RubinGPU ( NVL72 ), in-house development VeraCPU 、ConnectX-9 Network, BlueField-4 Security processor, and the world’s first 200GbCPO (Optoelectronic Co-Packaging) Optical Ethernet Spectrum-X Switches. Its supply chain scale is GraceBlackwell Twice as much, thanks to innovative cable-free technology. midplane Design, the single-rack assembly time from 2 Hours have been significantly shortened to 5 Minute.

VeraCPU : NVIDIA enters the fray CPU The market’s strategic iron fist

NVIDIA has launched a product specifically designed for AI In-house development by Intelligent   VeraCPU , based on the brand-new Olympus The core architecture—undoubtedly the most strategically significant announcement of the entire presentation. Huang Renxun stated: “ In the future, the number of Intelligent agents will far exceed that of humans, and they will be extremely impatient with latency when invoking tools or accessing databases. For ‘ Humanity ’ The Tradition of Design CPU Natural incompatibility with nanosecond‑level heterogeneous computing—this is an entirely new, massive market that NVIDIA has conjured out of thin air. ”

In real-world workloads, VeraCPU By eliminating Chiplet Lossy monolithic mesh interconnection ( 3.6TB/s Core bandwidth) and powerful single-threaded performance IPC Performance, let   SQL Database processing speedup 3 times , on the New York Stock Exchange's It’s even soaring in real-time stream processing tasks. 6 times , directly shattering the performance ceiling of traditional processors.

Open-source toolkit and 5 Speed-up Nemotron-3Ultra：

Lowering the threshold for enterprise Intelligent agents

To enable global enterprises to deploy Intelligent agents safely and in a customized manner, just like assembling software, NVIDIA Launch Intelligent enterprise AI Toolkit ( Agentic Toolkit for Enterprise AI ) , including open-source OpenShell Security sandbox and orchestration framework.

Jensen Huang demonstrated on-site with Cadence The chip design super-intelligent agent jointly developed by (Cadence) leverages automation. RTL Verification and debugging have reduced the original multi-week iteration cycle to just a few hours, dramatically boosting efficiency. 40 times. As the foundational cornerstone of this ecosystem, NVIDIA Officially released  Nemotron-3Ultra Open-source model . This model is based on a completely new SSM (State-space model) and Ministry of Education (Mixture of Experts Model) Hybrid architecture, with improved inference speed. 5 times, total cost reduction 30% , and has fully open-sourced the training scripts and the datasets used for tool evaluation, jointly CrowdStrike 、 Palantir 、 SAP and ServiceNow Let the industry giants jointly expand the enterprise Intelligent ecosystem.

40 Redefined for the first time in years PC

On the client side, NVIDIA Jointly with Microsoft, we have rearchitected the long-standing… 40 The evolving PC ecosystem over the years introduces a brand-new… RTXSpark Intelligent agent computing platform 。

Its core driving force stems from NVIDIA and MediaTek ( MediaTek ) jointly developed by  N1X chip : Integration Blackwell Architecture GPU (Providing 1PetaFLOP computing power), 20 Nuclear customization GraceCPU and also 128GB Unified memory, manufactured by TSMC.  3nm Craftsmanship.

The presentation featured a striking scene: in RTXSpark The agent runs locally on the notebook, via the integrated MCP (Model Context Protocol) The server interacts directly with local software and connects to the cloud. ClaudeSonnet , completed the transition from conceptual sketches to 3D The end-to-end design of architectural renderings. The Intelligent agent autonomously invokes Rhino 、 Blender and Flux2 Collaboratively completed, with automatic error correction throughout the process—no manual intervention required.

New PC The product line covers laptops, desktops, and DGXStation Three forms (the latter equipped with 768GB Memory, enabling the execution of trillion-parameter models on desktops. Jensen Huang predicts that in ten years, households… AI Supercomputers will become as commonplace as modern household appliances, this… PC It is on the same scale as the invention of the smartphone.

Physics AI ：Cosmos3 With Embodied Intelligence

The conclusion of the speech focuses on physics. AI and embodied intelligence. Huang Renxun believes that physical AI The critical pain point lies in the severe scarcity of first-person, machine‑perceived data; in response to this… NVIDIA Unleash the core weapon —— Cosmos3 Open-Source Cutting-Edge World Model 。

Cosmos3 It integrates autoregressive and diffusion models. Transformer The architecture can unify the processing of pixels, motion, sound, and language, directly generating synthetic videos that fully adhere to real-world physical laws through computation, thereby becoming a robot’s… “ Virtual Tutor ” . Based on this, NVIDIA It has released a reasoning-level Autonomous driving model with the capacity for self-talk and thinking.   Alphamayo , as well as highly integrated  IsaacGR00T Humanoid robot reference development platform . This robot is equipped with 31 one degree of freedom (including 25 A dexterous hand with degrees of freedom), by JetsonThor Chip-level drivers and a complete software stack are ready to use out of the box within hours for research institutions worldwide.

Core signal

Looking back on the entire speech, Jensen Huang repeatedly emphasized one core principle: Computing power equals revenue, per watt. Token Number equals profit margin. . NVIDIA’s competitive moat is no longer just a single… GPU Performance—not just at the chip level, but across the entire stack, from the chip to the rack, from software to operations, and from the cloud to the edge. PC Its full-stack integration and extreme collaborative design capacity.

When monomer AI The construction costs of factories are approaching the hundred-billion-dollar mark; whoever can help customers ramp up production faster, operate more efficiently, and sustain profitability over a longer period will emerge as the winner in this era. Huang Renxun stated in… Computex The single most crucial signal being conveyed is: NVIDIA is no longer a purely… GPU A company, rather than an Irreplaceable Infrastructure Monopolist in the Age of Intelligent Agents.