NVIDIA founder Jensen Huang: AI factories, physical intelligence, and the future of robotics

Writer: Techub News Compilation

Introduction

On the stage of COMPUTEX 2025, NVIDIA founder and CEO Jensen Huang delivered a keynote speech lasting nearly two hours. This speech was not only a highlight of the annual technology event but also a comprehensive strategic declaration from NVIDIA in the era of AI computing. Huang reflected on NVIDIA's transformation over the past thirty years—from a graphics chip company to the creator of the CUDA computing platform and now the core builder of global AI infrastructure. The importance of this speech lies in its clear outline of NVIDIA's vision for the next computing era: a new industry defined by "AI factories," Physical AI, and ubiquitous robots.

Summary

NVIDIA announced a new product line: the GB300 super chip, DGX Spark for developers, enterprise-level RTX Pro servers, and the open ecosystem MVLink Fusion platform.
Introduced the concept of "AI factories," believing future data centers will be factories for producing "smart tokens," with AI infrastructure becoming a societal foundation comparable to the power grid and the internet.
Revealed the evolution path of AI: from perception and generative AI to "Agentic AI," which has reasoning capabilities, ultimately achieving "Physical AI" that understands the physical world.
Announced collaboration with Foxconn, the Taiwan regional government, and TSMC to build the first giant AI supercomputer in Taiwan, strengthening the local AI ecosystem.
Showcased NVIDIA's positioning in the robotics technology stack, including the Isaac Groot open-source platform, Jetson Thor robotics processor, and the "Groot Dreams" blueprint that uses AI to generate synthetic data for training robots.

From Chips to AI Infrastructure: NVIDIA's Paradigm Shift

Huang began by pointing out that NVIDIA's story is a history of reshaping the computer industry and a story of the company's self-reinvention. He recounted the journey from the launch of the revolutionary CUDA platform in 2006 to the realization in 2016 of the need to reshape the entire technology stack (processors, software, systems), resulting in the birth of the DGX1 system, which was donated to OpenAI, igniting the spark of the AI revolution.

Today, NVIDIA's self-positioning has transcended that of a mere technology company. “We realize that NVIDIA is no longer just a technology company. In fact, we are an essential infrastructure company,” Huang emphasized. He likened AI infrastructure to the historical power grid and internet, becoming a new generation of fundamental elements ready to integrate into the fabric of society. This infrastructure specifically takes the form of "AI factories"—no longer traditional data centers, but production facilities that input energy and output highly valuable "smart tokens." In the future, corporate earnings reports might report "token output" like they report production capacity. Huang predicts that the scale of this AI infrastructure industry may reach trillions of dollars.

Core Engine: The Evolution of CUDA and Accelerated Computing Libraries

Huang reaffirmed that the integration of accelerated computing and AI is at the core of NVIDIA, and the key to achieving this is the vast ecosystem of software libraries. He showcased the CUDA-X libraries covering various fields, including Aerial for 5G/6G wireless signal processing, Parabrics for genomics, Monai for medical imaging, Earth-2 for climate prediction, and cuLitho for computational lithography (which can accelerate mask-making calculations by 50-70 times, with partners including TSMC, ASML, and Synopsys).

He emphasized that it is these domain-specific libraries that enable NVIDIA to migrate one scientific and industrial field after another to the accelerated computing platform. For example, in the telecommunications sector, after six years of refinement, NVIDIA can now provide a fully accelerated Radio Access Network (RAN) software stack that rivals the performance (data rate per watt) of the most advanced dedicated chips (ASICs), laying the groundwork for further overlay of AI (AI on 5G/6G).

Next-Generation AI Supercomputers: Grace Blackwell GB300 and MVLink Marvel

One of the highlights of the speech was the announcement of the upgraded version of the Grace Blackwell platform GB300. It adopts the same architecture and physical form but with upgraded chips, delivering 1.5 times the inference performance, 1.5 times the HBM memory capacity, and 2 times the network bandwidth. A single GB300 node can provide 40 Petaflops of performance, equivalent to the Sierra supercomputer of 2018, which comprised 18,000 GPUs, achieving an extreme Moore's Law of "four thousand times performance improvement in six years."

The key to achieving this "scale-up" capability lies in NVIDIA's self-developed MVLink interconnect technology. Huang demonstrated a system composed of 9 MVLink switches (each 7.2 TB/s) and a "MVLink Spine" weighing about 70 pounds and containing 2 miles of cable. This spine can provide 130 TB/s of fully interconnected bandwidth, “exceeding the peak traffic of the entire internet.” This allows 72 GPUs (144 chips) to work together like a giant GPU. Due to the physical limits of signal transmission, such high-performance integration can currently only be achieved within a single rack, which also leads to a power density of up to 120 kilowatts, necessitating all-liquid cooling.

An even more significant ecological strategy is the announcement of MVLink Fusion. This is an open platform that allows partners to integrate their custom ASICs, CPUs, or accelerators into NVIDIA's AI supercomputer ecosystem via MVLink chiplets or IPs. Huang stated clearly, “Your AI infrastructure can have a part that is NVIDIA's, and a large part that is your own,” providing flexible integration pathways for partners, including MediaTek, Fujitsu, and Qualcomm, aimed at expanding the overall AI infrastructure ecosystem.

The Next Wave of AI: Reasoning, Agency, and Physical AI

Huang outlined the stages of evolution of AI capabilities: from early perceptual AI (recognizing images and voices) to generative AI (content generation) over the past five years, now entering the “Reasoning AI” phase. AI needs not only to learn knowledge but also to have the ability to solve new problems, perform step-by-step reasoning, and weigh options (such as through techniques like thinking chains and tree thinking).

Combining multimodal perception, tool usage, and collaboration abilities, reasoning AI will evolve into “Agentic AI,” capable of understanding, thinking, and taking action as a digital entity. Huang likened it to digital forms of robots and believed that they will fulfill the workforce shortage globally (expected to reach a gap of 30-50 million by 2030).

Beyond digital agents, the next phase is “Physical AI,” which can understand the laws of the physical world (such as inertia, friction, object permanence, and causality). This capability is crucial for fields such as robotics and autonomous driving. NVIDIA is leveraging its simulation technology to generate vast amounts of synthetic video scenarios for training autonomous driving systems through prompts.

Democratizing Products: From DGX Spark to Enterprise-Level RTX Pro

To make AI computing power more accessible, NVIDIA launched several new products. Among them, DGX Spark is a desktop device aimed at AI-native developers, researchers, and students, providing 1 Petaflops of performance and 128 GB of memory, equivalent in performance to the DGX-1 from 2016, which weighed 300 pounds, but significantly reduced in size and power consumption, expected to be available in weeks.

For the enterprise market, NVIDIA announced the RTX Pro enterprise-level server. The key design of this server is its ability to seamlessly run traditional x86 enterprise IT environments (virtual machines, Kubernetes, etc.) while efficiently handling various AI agent workloads (text, graphics, video). It features the new Blackwell RTX Pro 6000 GPU and achieves up to 800 Gb/s of east-west traffic bandwidth between GPUs through a new CX8 interconnect chip (which combines switching and networking functions). Huang cited performance curve graphs showing that the server's performance on the Llama 70B model is 1.7 times that of the H100, while on the more optimized DeepSeek R1 model, performance can reach up to 4 times that of the H100.

Huang also outlined the other two major pillars of the enterprise AI stack: AI Storage and AI Operations (AI Ops). Future enterprise storage will require GPU upfront to handle semantic searches and indexing of unstructured data, with NVIDIA's NeMo Retriever and related blueprints working with storage vendors like VAST, Dell, Hitachi, IBM, and NetApp. AI Ops involves model fine-tuning, evaluation, safeguards, and security, with NVIDIA co-building an ecosystem with partners like CrowdStrike, DataRobot, and Elastic.

The Future of Robotics: Simulation, Learning, and Digital Twins

Robotics was another core aspect of Huang's speech. He announced that the complete platform for robot development and operations, Isaac Groot N1.5, is now open source and has been downloaded over 6,000 times. This platform includes AI supercomputers for training (GB200/300), Omniverse for simulation, and Jetson Thor robotics processors and Isaac OS for deployment.

To address the challenge of scarce training data for robots, NVIDIA introduced the “Groot Dreams” blueprint. This solution utilizes NVIDIA's Cosmos physical AI foundational model to augment a small amount of human teaching (teleoperation) data, generating large quantities of high-quality synthetic trajectory data, thereby achieving the effect of thousands of demonstrations with the work of a small team.

Huang emphasized the importance of humanoid robots, viewing them as the only robot form capable of seamlessly integrating into existing "brownfield" environments (designed for humans), and due to their potential for massive scale, they are most likely to become the next trillion-dollar industry.

None of this would be possible without digital twins. Huang demonstrated digital twins of factories, data centers, and even cities built using Omniverse by partners in Taiwan, such as TSMC, Foxconn, Quanta, and Wistron. These highly realistic simulation environments are not only tools for planning and optimization but also act as "robot gyms" for training robots and simulating multi-agent collaboration. He revealed that there are approximately $5 trillion in new factory investment plans globally over the next three years, making digital twin technology crucial.

Deepening Ties with Taiwan: Announcing Local AI Supercomputing and New Headquarters

As an affirmation and investment in the Taiwan ecosystem, Huang announced a significant collaboration: NVIDIA will work with Foxconn, the Taiwan regional government, and TSMC to build the first giant AI supercomputer in Taiwan, providing world-class AI infrastructure for local researchers, students, startups, and large companies.

At the end of the speech, Huang heightened the atmosphere by announcing a "new product"—due to the expanding employee base in Taiwan, NVIDIA plans to build a new headquarters building named “NVIDIA Constellation” in the Beetho Sheiling area of Taipei, symbolizing a closer and longer-term bond between NVIDIA and the Taiwanese technology industry.

Huang concluded that this is a "once-in-a-lifetime" opportunity; NVIDIA and its partners are not just creating the next generation of information technology but are also pioneering an entirely new industry. He looks forward to collaborating with ecosystem partners to build AI factories, enterprise agents, and robots, while expanding the ecosystem with a unified architecture.

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。