NVIDIA founder Jensen Huang talks about AI factories, sovereign AI, and future computing paradigms.

Written by: Techub News Compilation

Introduction

Recently, NVIDIA founder and CEO Jensen Huang was invited to attend an event hosted by Citadel Securities, where he engaged in an in-depth dialogue with Sequoia Capital partner Konstantine Buhler. During this global exchange aimed at top institutional investors, Huang reflected on NVIDIA's thirty-year journey from a graphics chip company to an AI infrastructure giant, sharing profound insights on the future development trends of AI, industry landscape, technological challenges, and geopolitical impacts. As a core architect of the current AI revolution, his viewpoints are crucial for understanding the future investment direction in AI, which may reach tens of trillions of dollars.

Summary

AI factories are facilities that produce "intelligence": Huang redefined data centers, emphasizing that their core value lies in the "token" (intelligence) throughput generated per unit of energy consumption, which directly determines customers' revenue capabilities.
From accelerated computing to general-purpose platforms: NVIDIA's success stems from early predictions about the limits of Moore's Law and a commitment to transforming dedicated accelerators into general-purpose computing platforms through frameworks like CUDA.
Generative AI will reshape all computation: The future computation paradigm will shift from retrieval-based to 100% generative, generating content in real-time like human conversation, which will drive a massive demand for AI factory infrastructure.
Sovereign AI as a national strategy: Every country should develop capabilities to produce national intelligence based on its own data, rather than relying completely on external sources, accelerating the global build-out of sovereign AI.
Robotics and physical AI are the next frontier: General AI models will be able to "embody" in various physical forms, from cars to humanoid robots, with their development relying on three sets of computer systems for training, simulation, and deployment.

From Graphics Acceleration to AI Revolution: NVIDIA's "First Principle" Path

The dialogue began with the founding origin of NVIDIA. In 1993, the personal computer and CPU revolution was in full swing, and Moore's Law served as the investment bible of Silicon Valley. However, at only 30 years old, Huang saw things differently. He observed that general-purpose CPUs were insufficiently efficient for solving certain complex problems, while dedicated accelerators could provide better solutions. More importantly, he and his co-founders foresaw the eventual physical limits of transistor miniaturization, and the scale of computational problems the world needed to solve was nearly infinite. Thus, they decided to establish NVIDIA, focusing on accelerated computing as a complement and enhancement to general computing.

However, creating a brand-new computing architecture faced the "which came first, the chicken or the egg" dilemma: without a large market, a new platform couldn't be sustained; and without a new platform, the market couldn't even be discussed. Huang recalled an early anecdote from seeking funding from Sequoia's Don Valentine, where he pitched Electronic Arts as the "killer app," unaware that Sequoia had just invested in the company and that its CTO was only 14 years old. Despite this, NVIDIA ultimately succeeded in pioneering and driving the modern 3D graphics gaming ecosystem, which became the foundation for everything that followed.

3D graphics essentially simulate reality, with a mathematical foundation in physical simulation and linear algebra. NVIDIA realized that gradually generalizing highly specialized technology was a key growth path for the company. The invention of CUDA was not only a technological breakthrough but also a comprehensive innovation in product strategy, market expansion, and ecosystem building. Huang emphasized that creating a widely-used new computing platform like ARM or x86 is extremely rare, and it took NVIDIA nearly 30 years to achieve this.

The Birth of Deep Learning, CUDA, and AI Factories

Fast forward to the early 2010s, deep learning was still on the academic fringes. The breakthrough performance of AlexNet in the ImageNet competition in 2012 ran entirely on NVIDIA GPUs. Huang shared the opportunities and insights from that time: he was struggling with stagnation in computer vision technology, and during his outreach to global university researchers through the "CUDA Everywhere" strategy, he met top scholars like Geoff Hinton, Andrew Ng, and Yann LeCun who were tackling the same problem.

NVIDIA developed the key library—Q DNN—to enable researchers to utilize CUDA more efficiently. But more important than witnessing the leap in computer vision performance was the deeper reasoning by the NVIDIA team: due to the properties of layered training and backpropagation, deep neural networks can learn almost any function, becoming a universal function approximator. They concluded that most problems awaiting resolution could incorporate deep learning components. Thus, NVIDIA decided to go all-in and rethink every layer of the computing stack—from chips, systems to software.

In 2016, NVIDIA launched the world's first AI supercomputer DGX-1. Huang humorously recalled that the audience's reaction was tepid during the GTC conference release until he invited Elon Musk to the stage. Musk immediately stated that OpenAI, this "non-profit organization," needed one, so Huang personally acted as a "delivery man," bringing the massive device to San Francisco. Today, the DGX has evolved into a rack-level GPU worth millions of dollars, weighing two tons, and consuming significant power, with larger-scale AI factory GPU clusters worth hundreds of billions to even trillions of dollars.

Huang explained why he referred to it as a "factory" rather than a data center: "They are making money with it." The intelligence output per unit energy consumption (token generation rate) directly determines customers' revenue. Therefore, NVIDIA's innovation speed is incredibly fast; through full-stack collaborative design (algorithms, software, networks, CPUs, GPUs), performance improves by about 10 times each year while driving costs down rapidly, allowing customers to generate more revenue from the same factory.

Trillion-Dollar Market: AI's ROI and Future Investment Directions

Regarding the market potential and ROI that investors are concerned about, Huang provided a clear picture. He first dismissed the analogy between the current AI craze and the 2000 internet bubble. He pointed out that today’s AI revolution is primarily a transformation of existing trillion-dollar-scale industries. For example, when Meta faced the impact of Apple's privacy policy at the end of 2022, it restored its ad attribution capabilities using the AI recommendation system powered by NVIDIA GPUs, resulting in a market value rebound of over a trillion dollars. Google's, Amazon's, and TikTok's search, recommendation, and personalized advertising systems are all fully transforming towards AI. This constitutes the first wave of transformation investments worth hundreds of billions.

The second wave involves emerging AI model manufacturers (such as OpenAI, Anthropic, xAI, etc.) building their own AI factories. The third wave, which is the one with the greatest potential, is the AI-native application layer, especially "agent AI" and physical AI. Agent AI will create a digital workforce, such as software engineers, accountants, lawyers, and marketers, tapping into a multi-trillion-dollar corporate market. Huang revealed that NVIDIA internally uses AI coding assistants (like Cursor) 100% to enhance all engineers' work.

Physical AI refers to robotics. Huang explained its inevitability through a thought experiment: since AI can generate a video of "Jensen Huang picking up a bottle and drinking water," then it should be able to control a robot to perform that action. An autonomous vehicle is essentially a "digital driver," and its technology can be generalized to manipulate robotic arms, humanoid robots, or any physical entity. This requires three sets of computer systems: an AI factory for training models, a "laboratory" computer for simulating learning in the virtual world (Omniverse), and a "brain" computer within the robot's body. NVIDIA provides all three and collaborates with nearly all robotics and autonomous driving companies.

Huang predicted that the future computation paradigm will completely shift to 100% generative. He cited Perplexity AI and Sora as examples: the answers generated by the former are not pre-stored, but synthesized in real-time; every pixel and action from the latter is created by AI. Future computers will resemble real-time collaborating CEOs, artists, or poets, dynamically generating unique content based on context. Supporting all this will be AI factories. He believes that infrastructure worth hundreds of billions has only been built globally so far, while the future may require trillions of dollars in investment each year.

Sovereign AI, Geopolitics, and AI Security

Regarding the increasingly important topic of sovereign AI, Huang expressed a clear viewpoint: "No single country can afford to outsource all its national data and then import its own intelligence." He believes this contradicts first principles. Every country should leverage its own data to produce national intelligence, but it can also import or purchase some technologies. With open-source capabilities and tools maturing, various countries are advancing sovereign AI initiatives, like France's Mistral, the UK's Nscale, and many companies in Japan.

When discussing export controls on China, Huang displayed a pragmatic and nuanced stance. He pointed out that while the U.S. certainly wants to win the AI race, policies need to strike a balance. China has about 50% of the world’s AI researchers and is a vibrant, huge market. He thinks allowing Chinese researchers to develop AI based on U.S. technology stacks is in America's interest, and a complete separation would be a lose-lose situation. He revealed that due to current policies, NVIDIA's market share in China has dropped from 95% to zero, with the company assuming zero revenue from China in all forecasts. He hopes for a policy adjustment because "it's hard to imagine any policymaker thinking that letting the U.S. lose one of the world's largest markets is a good thing."

Regarding AI security, Huang believes its future will be akin to cybersecurity, requiring the global community to share vulnerabilities and threat intelligence. Furthermore, as the marginal cost of AI trends towards zero, the marginal cost of secure AI will also approach zero. Therefore, in the future, every AI system may be surrounded by thousands of "guardian AIs," forming multilayered protections. In the digital world, the ratio of security personnel may far exceed that in the physical world.

Lightning Q&A: Undervalued KPIs, Technologies, and Leadership

In the fast-paced Q&A segment at the end of the dialogue, Huang shared some incisive insights:

The most undervalued KPI on Wall Street: The "token generation rate per unit energy consumption" of AI factories, which directly determines customers' revenue capabilities.
The most undervalued aspect of the NVIDIA platform: It’s not CUDA, but rather the over 350 libraries built on top, such as Q DNN and cuLitho (for semiconductor manufacturing lithography), which are NVIDIA's "treasures."
Severely undervalued technology: Omniverse (the virtual world simulation platform), which is crucial for training physical AI and is sweeping the robotics industry.
The most impactful business books: All works by Clayton Christensen, Al Ries's "Positioning," and Geoffrey Moore's "Crossing the Chasm."
$10 billion investment advice for CIOs: Start experimenting immediately to build their own company-specific AI. The future IT department will become the "HR department for agent AI," responsible for hiring, training, and integrating digital employees.

The entire dialogue outlined Huang's clear line of thinking: anticipating trends from first principles, building a moat through full-stack innovation, and driving the evolution of the entire industry ecosystem with platform thinking. In his view, the AI revolution is far from over; rather, it has just opened the curtain on the trillion-dollar infrastructure buildup.

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。