NVIDIA founder Jensen Huang talks about AI factories, the inference revolution, and the future of robotics.

Written by: Techub News Compilation

Introduction

In March 2025, NVIDIA's annual technology conference GTC grandly opened in San Jose, California. During a theme speech lasting over two hours, the company's founder and CEO Jensen Huang took the stage again in his signature leather jacket, showcasing the latest advancements and future blueprint under the AI wave to global developers, partners, and industry observers. This speech was acclaimed as the "AI Super Bowl" due to its significance; it not only announced the full-scale production of GPUs based on the Blackwell architecture but also systematically explained that AI development has entered a new paradigm centered around "inference" and "embodied intelligence," revealing an explosion in computation demand by orders of magnitude.

Huang emphasized in his speech that AI is evolving from generative AI to agentic AI, and then to physical AI and robotics. Currently, AI models supporting the "inference" capability of agents require computation far exceeding expectations from a year ago. To address this, NVIDIA introduced the concept of the "AI Factory" and released an operating system named "Dynamo" aimed at managing this new computing paradigm. At the same time, in the robotics field, NVIDIA announced the open-source of its general-purpose robotics foundation model Groot N1 and launched a new generation of physics engine "Newton" in collaboration with DeepMind and Disney's research division, marking a significant acceleration in the industrialization process of embodied intelligence.

Summary

Explosion of inference computing demand: Due to the complex thinking required for agent AI, such as step-by-step reasoning and consistency checking, the number of tokens generated may exceed that of traditional generative AI by hundreds of times, creating unprecedented demands on computing infrastructure.
"AI Factory" and operating system: Future data centers will transform into AI factories focused on "generating tokens." NVIDIA introduced the open-source operating system "Dynamo" to dynamically schedule and manage complex inference workloads on massive GPU clusters.
Blackwell platform fully ready: Blackwell architecture GPUs are now in full-speed production; their MVLink 72 rack-level system has improved inference performance up to 40 times compared to the previous generation Hopper under the same power consumption, designed to meet extreme demands in inference computing.
Era of robotics begins: Through the Omniverse digital twin platform, generative model Cosmos, and new physics engine Newton, NVIDIA has built a complete stack for robotic development and training, and open-sourced the Groot N1 model to accelerate the development of the general robotics industry.
Full-stack AI infrastructure in place: From cloud AI factories, enterprise AI workstations (DGX Spark/Station), to edge communication networks (in cooperation with T-Mobile and Cisco), and to robotic computing platforms, NVIDIA is building AI infrastructure covering all scenarios.

Paradigm Shift in AI: From Generation to Inference, to the Physical World

Huang began by revisiting the development of AI over the past decade: from perceptual AI (computer vision, speech recognition) to generative AI, and now to agentic AI and physical AI. He pointed out that generative AI fundamentally changed the computing model, shifting from "retrieval computing" to "generative computing." The current key breakthrough is agentic AI, which focuses on endowing AI with "inference" capabilities. AI can perceive context, plan steps, use tools, and gradually solve problems through techniques like "thinking chains," rather than outputting answers in one go.

This transformation has led to an exponential increase in computation demand. Huang explained that when an AI performs step-by-step reasoning, it needs to generate a series of tokens representing the steps of thought. For example, to solve a complex banquet seating arrangement problem, a traditional model might quickly generate 500 tokens but provide an incorrect answer, whereas an inference model would need to generate nearly 9000 tokens to deduce the correct answer step by step. This is not just an increase in the number of tokens; to maintain interaction in real-time, computational speed must also increase correspondingly. "With both factors combined, the amount of computation we need can easily exceed our expectations from a year ago by over a hundred times," emphasized Huang.

The training methods supporting agentic AI have also undergone transformations. By combining reinforcement learning with verifiable outcomes (such as mathematical theorems, physical laws), AI can self-train on vast amounts of synthetic data without fully relying on limited human-annotated data. This addresses the fundamental issues of data and training scalability, but it also presents the challenge of "extreme computing" to the entire industry.

AI Factory: Fundamental Restructuring of Computing Infrastructure

In the face of a hundredfold increase in computation demand, Huang proposed the concept of the "AI Factory." Future data centers will no longer be mere locations for running traditional software; they will be factories specifically "producing" intelligent tokens. These tokens are restructured into text, images, videos, protein structures, or business insights, directly creating value. Therefore, the operational efficiency of the AI factory is directly related to service quality and economic returns.

The operation of the AI factory faces a fundamental contradiction: it must provide fast (low-latency) intelligent responses for individual users while simultaneously serving massive numbers of users (high throughput). Huang illustrated this challenge with a graph, where the horizontal axis represents the token generation speed for user experience, and the vertical axis represents the overall factory's token output throughput, with the ideal state maximally maximizing both. This requires infrastructure to have extremely high computational performance (FLOPS) and memory bandwidth.

To this end, NVIDIA launched the groundbreaking Blackwell platform. Its top configuration MVLink 72 integrates all GPUs within a cabinet into a colossal computing device through high-speed interconnects, providing one Exaflops of computing power per second and 576 TB/s of memory bandwidth. Data presented by Huang indicated that under the same power consumption, the Blackwell MVLink 72 system can achieve performance on inference workloads up to 25 times that of the previous generation Hopper system, and for models requiring complex inference, the performance advantage can reach 40 times. He candidly stated, "When Blackwell starts shipping in large quantities, Hopper may become obsolete."

However, hardware is just the foundation. Efficiently running trillion-parameter models on such a large and heterogeneous GPU cluster and dynamically allocating resources to meet the demands of different computation phases such as "context filling" and "token decoding" require extremely complex software scheduling. The Dynamo released by NVIDIA today is precisely created to solve this problem. Huang referred to it as "the operating system of the AI factory." Dynamo can dynamically manage complex distribution strategies such as tensor parallelism, pipeline parallelism, and expert parallelism of models across tens of thousands of GPUs and flexibly allocate computing resources between "thinking" and "output" tasks. NVIDIA announced it will open-source Dynamo and collaborate with partners like Perplexity to advance its ecosystem.

Breaking Communication Bottlenecks: Silicon Photonics and Next-Generation Networks

To scale the AI factory to hundreds of thousands or even millions of GPUs, traditional networking methods have become unsustainable in terms of power consumption and cost. Huang pointed out that under the current architecture, each GPU may require six optical modules, each consuming 30 watts, with a cost of about $1000. For a data center with a million GPUs, the power consumption of just the optical modules would reach 180 megawatts, resulting in a cost of $6 billion.

NVIDIA's solution is silicon photonics technology. Huang demonstrated the world's first 1.6 Tb/s co-packaged optical (CPO) switch chip, which utilizes micro-ring resonator modulator technology at its core. This technology integrates photonic integrated circuits and electronic integrated circuits through 3D packaging, achieving extremely high density and energy efficiency. Compared to traditional solutions, this new technology can save data centers tens of megawatts in power consumption, allowing that precious electricity to be redirected to increase additional computing units. The Spectrum-X Ethernet platform equipped with this technology is expected to launch in the second half of 2026, paving the way for ultra-large-scale AI clusters.

Robotics: The Ultimate Expression of Physical AI and Industrial Explosion

In the second half of the speech, the focus shifted to physical AI and robotics. Huang believes that robots are the key to solving the global labor shortage and will become a massive industry in the future. NVIDIA provides a complete computing platform that covers simulation, training, and deployment.

The core challenges remain data and training. NVIDIA generates highly realistic and diverse synthetic training data via the Omniverse digital twin platform and offers verifiable reward signals through the physics engine to train robotic AI. Huang announced two significant advancements: the first is the new generation GPU-accelerated physics engine Newton, launched in collaboration with DeepMind and Disney's research division, focused on high-fidelity, ultra-real-time simulation of rigid bodies, soft bodies, and haptic feedback, providing a foundation for fine skill training for robots. The second is the open-sourcing of its general humanoid robotics foundation model, Groot N1. This model employs a dual-system architecture of "fast thinking" and "slow thinking," enabling it to perceive the environment, plan reasoning, and control robots to execute complex task sequences.

"The speed of development in physical AI and robotics is so rapid," Huang warned, "this could very well become the largest industry of the future."

From Cloud to Edge: Comprehensive Implementation of Full-stack AI Infrastructure

Huang emphasized that AI will be everywhere, not just limited to the cloud. NVIDIA is building AI infrastructure covering all scenarios:

Cloud and AI Factory: The Blackwell platform and Dynamo operating system serve ultra-large-scale CSPs and AI companies.
Enterprise: A new AI workstation product line is launched, including the portable "DGX Spark" (in collaboration with MediaTek) and desktop-level "DGX Station," bringing powerful AI computing capabilities to every developer and data scientist. At the same time, collaborations with OEM manufacturers such as Dell and Huayou offer full-stack enterprise AI solutions, even redefining the future "semantic storage" system.
Edge and Communication: Announced collaborations with T-Mobile, Cisco, and CUS-ODC to build a telecom edge computing full stack in the United States, aimed at revolutionizing wireless communication networks with AI.
Autonomous Driving: Announced that General Motors chose NVIDIA as a partner to jointly build its future autonomous driving fleet. Huang also specifically introduced NVIDIA's deep investment in functional safety (DRIVE Halos).

Finally, Huang announced a clear product roadmap, pledging to update the architecture every year, launch new product series every two years, and preview the next-generation platform to be named after renowned physicists "Vera Rubin" and "Richard Feynman." Before concluding the speech, he again expressed gratitude to the global developer ecosystem and played a stunning film generated by AI, paying tribute to the human spirit of exploration, bringing the atmosphere of the event to its peak.

The entire speech indicated that NVIDIA is no longer just a chip company; as the "chief architect" of the AI revolution, through full-stack innovations from silicon, systems, software to ecosystems, it is defining and building the core infrastructure of the intelligent age.

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。