Written by: Techub News Compilation

Introduction

On June 1, 2026, NVIDIA founder and CEO Jensen Huang delivered a two-hour keynote speech at the COMPUTEX exhibition in Taipei. This speech is seen as a "weather vane" guiding the direction of AI and computing industries in the coming years. Huang not only returned to his hometown in Taiwan but, in front of his parents and tens of thousands of live audiences, systematically elaborated on the key transition of AI from "generative" to "useful," while unveiling a series of hardware and software products aimed at reshaping the landscape from data centers to personal computers.

Currently, the world is in a frenzy of building AI factories, where computational power equals profit. The core of Huang's speech was to announce that a brand-new computing paradigm—agent-centric computing—has matured and begun to create immense economic value. He detailed the technology stack supporting this paradigm, including the fully operational Vera Rubin system, the Vera CPU designed specifically for agents, and the RTX Spark personal computer, reinvented in collaboration with Microsoft after 40 years.

Summary

• The era of Agentic AI has truly arrived, where AI has transitioned from generating content to executing useful work, becoming a direct creator of profits and GDP.
• NVIDIA launched the Vera Rubin multi-rack supercomputer, the first end-to-end system designed specifically for handling agent workloads, which is now fully operational.
• Introduced the revolutionary Vera CPU, which shifts its design focus from serving humans to serving agents, pursuing extreme single-thread performance, bandwidth, and energy efficiency to eliminate bottlenecks in agent workflows.
• NVIDIA is cooperating closely with Microsoft to launch the RTX Spark PC product line, reinventing the personal computer to natively run agents, initiating the largest PC revolution since Windows 95.
• Released the enterprise AI agent toolkit, physical AI foundational model Cosmos 3, autonomous driving model Alpamo 2, and humanoid robot reference platform Isaac Groot, extending the agent paradigm across various industries.

The Arrival of the Agent Era: From Generative to Useful, Computational Power Equals Income

Huang opened by pointing out that the tide of AI development has transitioned from generative AI to the Agentic AI phase. "Useful AI" has arrived. Using global software development as an example, he demonstrated the productivity revolution brought by agents: the volume of code submissions on GitHub nearly tripled in the first few months of 2026. This means that the economic output created by approximately 30 million software developers (representing about $30 trillion in annual salaries) could approach $9 trillion. AI has not replaced jobs but has greatly enhanced the output value of individual engineers, prompting companies to hire more engineers.

The core of this transition lies in AI-generated "Tokens," which are no longer merely information units, but profitable units of income. Agents can observe, reason, plan, and use tools (such as browsers, databases, CAD software) to accomplish real tasks, from writing code and creating GIF animations to designing 3D printed parts. This has created massive demand for Tokens, subsequently driving frenzied investments in computational power—that is, AI factories.

Huang emphasized, "Computation equals income, computation equals profit." In this new paradigm, every watt of power in a data center or AI factory is directly linked to Token output and income. Therefore, the key metrics to measure system performance are no longer merely peak computational power but "Tokens per watt" and the overall reliability and lifecycle of the system. This fundamentally changes the design philosophy and economics of computational infrastructure.

Vera Rubin: A Supercomputer Built for Agents

To address the extreme complexity of agent workloads, NVIDIA introduced Vera Rubin. Huang clarified that Vera Rubin is not a single GPU chip, but rather an end-to-end designed, multi-rack scalable agent processing system. The working mode of agents is "decentralized and distributed": large language models are responsible for "thinking," tools run on CPUs or accelerators, memory management systems handle working memory and long-term memory, and secure processors ensure full-process encryption, all coordinated by orchestration software. This is completely different from the traditional application running within a single operating system.

The Vera Rubin system integrates at least seven new chips, including the Vera Rubin GPU (MVLink 72), Vera CPU, Bluefield-4 DPU, ConnectX-9 network chip, and Spectrum-X Ethernet switches. Its design employs "extreme collaborative design," completing the entire system simulation, validation, and optimization in the Omniverse through digital twins before physical construction. Huang specifically showcased its revolutionary rack design: cable-free, hose-free, and fan-free, connecting via a center-mounted PCB backplane, reducing rack assembly time from 2 hours during the Grace Blackwell era to just 5 minutes, significantly enhancing reliability and maintainability.

Huang announced that Vera Rubin has now entered full production, with a supply chain scale twice that of Grace Blackwell. This system is designed for the agent era, capable of efficiently handling the complex tasks required by agents, such as observation, reasoning, planning, tool usage, and memory management, marking a landmark product in NVIDIA's transition from a GPU company, to a systems company, to an AI infrastructure company.

Vera CPU: Redefining the Computing Core for Impatient Agents

Huang pointed out that traditional CPUs are designed for human users, with their usage and rental patterns measured in "seconds." However, agents are "impatient" and live in a "nanosecond" world. In agent workflows, CPUs are used for orchestration, managing tool calls, accessing databases, and any waiting will degrade the entire agent's execution efficiency, thus affecting valuable Token output.

To address this, NVIDIA designed the completely new Vera CPU, the first CPU built for the agent era. Its design goals focus on four points: extremely high single-thread performance (IPC), extremely high per-core bandwidth, extremely high total chip throughput, and extreme energy efficiency. The Vera CPU uses a self-developed core named "Olympus," a second-generation scalable coherent interconnect fabric, and supports LPDDR5X memory for the first time, achieving several times the performance improvement compared to existing x86 CPUs. For example, on key workloads like SQL queries and real-time streaming processing (such as at the New York Stock Exchange), the Vera CPU provides a 3x to 6x speedup.

Huang predicts that the number of agents will far exceed that of humans in the future, and their thirst for low latency and high efficiency will fuel a brand new, enormous CPU market. The Vera CPU will be tightly coupled with the Vera Rubin system, becoming a key link in optimizing AI factory output (Tokens).

Reinventing the PC: The RTX Spark Revolution by NVIDIA and Microsoft

Huang shifted the topic from data centers to personal computing, announcing a collaboration with Microsoft to reinvent the personal computer after 40 years. The new computing paradigm—agents—will not only run in the cloud but will also operate on everyone’s devices. To this end, NVIDIA launched the RTX Spark platform.

The core of RTX Spark is a chip developed in partnership with MediaTek, the N1X chip, which integrates an RTX GPU based on the Blackwell architecture, a customized 20-core Grace CPU, and, through NVLink interconnect, possesses 128GB of unified memory. Huang emphasized that this is a "100% compatible with Windows, 100% running CUDA, and 100% equipped with NVIDIA AI Tensor Cores" platform that can seamlessly run all applications and AI workloads accumulated over the past decades.

A live demonstration showcased how an agent running on the RTX Spark laptop helped users design houses: users expressed intentions through natural language and sketches, and the agent could invoke local professional software like Rhino and Blender for modeling and rendering, even leveraging cloud models to generate visual effects. This indicates that the future role of PCs will shift from passive tools to proactive assistants with their own autonomous agents.

Additionally, NVIDIA announced a brand new product line, including RTX Spark laptops, desktops, and workstations. Huang envisions that in the future, households might have an always-on "AI supercomputer" connected to all devices, serving as a personal agent hub, akin to the evolution from mobile phones to smartphones.

Empowering Smart Everything: From Enterprise Toolkits to Physical AI

To enable industries to build their own agents, NVIDIA launched the "Enterprise AI Agent Toolkit." This toolkit consists of four core components: models (like the open-source Neotron, with a smarter, more efficient Neotron 3 Ultra to be released at the conference), agent orchestration frameworks (like Open Shell), tools and skills (such as the CUDA-X library), and runtime environments. Huang used the example of a "chip design super-agent" developed in collaboration with Cadence to demonstrate how this toolkit reduced the chip verification cycle from several weeks to a few hours.

Finally, Huang expanded the scope of agents from the digital world to the physical world. He announced the physical AI foundational model Cosmos 3, a multimodal model capable of understanding, reasoning, and generating scenarios in the physical world, applicable for robotic training, simulation, and more. Additionally, he introduced the reasoning-enabled autonomous driving model Alpamo 2, and an open-source reference platform Isaac Groot designed to lower the barrier for humanoid robot development (including hardware, software, and reference robot designs).

Huang summed up that whether it is cloud agents, PC agents, robots, or autonomous vehicles, their core computational model is the same: models, orchestration frameworks, tool skills, and runtimes. Through Vera Rubin, Vera CPU, RTX Spark, and a series of software toolkits, NVIDIA is building a complete infrastructure stack for this new agent era. He thanked Taiwan's supply chain partners for their great contributions and firmly believes that close cooperation with the ecosystem will continue to drive this unprecedented technological revolution.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。