黄仁勳最新訪談：晶片業還要再擴張5到10倍，中國模型對所有人都是利多

Compiled & Translated: TechFlow

Podcast: Axios Behind the Curtain

Guest: Jensen Huang, Co-founder and CEO of NVIDIA

Duration: 70 minutes

Recording Location: NVIDIA Factory in Fort Worth, Texas

Conflict of Interest Disclosure: Jensen Huang is the CEO and Co-founder of NVIDIA, holding approximately 3.5% of the company's shares, with a personal net worth of about $181 billion. This content touches on topics such as AI industry expansion, chip demand, and Chinese government regulations, which are directly related to his personal financial interests. His stance clearly leans towards promoting AI industry expansion and reducing regulation.

Summary

The CEO of the world's most valuable company, seated in his newly built Texas factory, addressed nearly every hot-button controversy in the AI industry. Huang was blunt: Wall Street's panic over China's AI models is a misunderstanding, AI doomsday scenarios are "nonsense," the chip industry needs to expand 5 to 10 times in the next decade, and a bubble is unlikely within five years. He also publicly stated that Anthropic's Mythos model should be open to everyone, and that OpenAI and Anthropic would be "the most successful IPOs in human history." Regarding Trump, he said the President has an excellent memory, capable of recalling chip model numbers like H20, H200, Blackwell, and Rubin, with his only concern being that the government might be misled by "science fiction narratives" into over-regulating.

Key Quotes

"The market misunderstood the impact of DeepSeek, and now it's misunderstanding Kimi. Great models lead to more usage, more usage means selling more Nvidia computers and building more data centers. The starting point is: good models lead to good applications, and good applications lead to growth."

"Saying AI will destroy humanity is complete nonsense. Saying AI will wipe out half the jobs in America is complete nonsense. All the facts and evidence point to the opposite."

"China has more AI researchers than the rest of the world combined. China will become extraordinary in this field; it is destined. Trying to stop China? That's a stupid idea, and it's simply not possible."

"The semiconductor industry needs to expand 5 to 10 times. Everything is in short supply today: chips, memory, land, electricity, and construction workers. This shortage is actually a good thing; it gives us time to build infrastructure."

"OpenAI and Anthropic will become the most successful IPOs in human history. How can a company be worth a trillion dollars just a few years after its founding? Because AI is useful, and if AI is useful, it can make money."

1. China's AI Shockwave: Wall Street Misunderstood Again

Host Mike Allen kicked off with the most sensitive topic: the FT reported that China is considering tightening export controls on AI models and semiconductors, that Kimi, a Chinese model, just caused Nvidia's stock to plummet, and chip stocks fell 18% in a month. What did Wall Street get wrong?

Huang's response was crisp and decisive. He said the market got it wrong when DeepSeek came out, and now it's happening again. "Good models lead to more usage, more usage means selling more Nvidia computers. You need to build more data centers, provide more services, and technology will permeate more industries." He repeatedly emphasized this logical chain: good models lead to good applications, good applications lead to growth, and growth requires more computing power.

Regarding whether to ban Chinese models like Kimi, his answer was equally direct: "Of course you should use them, it's smart." He explained that after downloading a model, you can fine-tune, enhance, and add guardrails. The model runs in a so-called "harness," and the harness is within a "sandbox." The sandbox is secure, with privacy protection, security controls, and access control. He compared it to an operating system: Linux is open source, reviewed, tested, and hardened by millions worldwide, so we can trust it. The same principle applies to open-source AI models.

He also mentioned an easily overlooked point: open models and closed models are not opposites. "The people most likely to upgrade to a good model like Anthropic or OpenAI are those already using AI." Free models lower the barrier to trying AI. Once users get the hang of it, they will naturally want better services and will then pay closed-model vendors. So, the more open-source models there are, the greater the opportunity for closed models.

When asked about NVIDIA's sales in China, Huang gave an answer many didn't expect: "Our sales in China are approximately zero today." He said he had already told investors not to expect any revenue from China. If the Chinese government and market welcome them back, it would be a "great honor," but until then, treat sales as zero.

2. Where the AI Doomsayers Are Wrong

This was the most confrontational part of the conversation. The host asked: Are some of your tech peers overstating the risks of AI?

Huang's response was blunt. "It's okay to warn people, it's better to warn with solutions, but fabricating facts is absolutely unacceptable." He then named several popular AI fear narratives: "Saying AI will end humanity is complete nonsense. Saying AI will eliminate half of American jobs is complete nonsense. All the facts and evidence point to the opposite."

He presented several specific cases to support his argument. The number of radiologists has increased by about 20% because, after AI automated scan analysis, doctors can see more patients, and there are too many people needing care, so actually, more radiologists are needed. The number of paralegals has increased by about 10%, based on the same logic. Manufacturing jobs have grown by about 50% in the past few years because AI data centers need to be built and chips need to be produced.

"Intuition and common sense tell you that AI will increase productivity, and increased productivity creates opportunities. Throughout history, technology has made society more efficient, creating more jobs, not fewer. If it were otherwise, America would only have about 100,000 jobs today."

He also pointed a finger at AI leaders. "Doomsayers spend too much time theorizing about sci-fi endings; maybe it makes them look smart." When pressed on whether he was referring to CEOs of specific AI companies, he didn't deny it but said: "If your goal was to make the world aware of this technology's incredible capabilities, that goal has been achieved. We should now spend our time on how to make the technology safe. That's the responsibility of technology leaders."

The host mentioned that Asia's attitude towards AI is completely different from that of the US, with Huang being mobbed by fans and asked for autographs in Asia. His explanation was thought-provoking: "Maybe it's because doomsayers spend too much time fabricating sci-fi endings." In Asia, people embrace AI as a tool and opportunity, not fear it as a threat.

3. The Chip Industry Needs to Expand 5-10x, a Bubble in 5 Years Is Unlikely

The host asked a sharp question: Every industrial revolution experiences a bubble. Where is the bubble risk for this generation?

Huang's answer was cautious but clear. "A bubble will come eventually, but not today. We are at the very beginning of construction." He gave a timeframe: unlikely within five years, but depends on circumstances between five and ten years. The reason is the overall supply-side constraints. "The industry can build faster now, but there aren't enough chips, enough memory, enough land, enough electricity, or even enough construction workers. We are constrained in every direction, in every aspect."

This constraint is precisely a good thing, he said. "This constraint holds the system back, giving us ample time to build infrastructure." Demand is strong, but the ability to translate that demand into productive supercomputers is delayed due to all these physical limitations. This pushes the point of oversupply further out.

More critically, his judgment on the overall size of the semiconductor industry. "I believe the semiconductor industry needs to expand by 5 to 10 times." The timeframe? "Over the next decade." This means today's chip industry is still far too small, far from sufficient to support the construction needs of AI infrastructure.

He explained why this cycle is different from past semiconductor cycles. "This time it's different because it's not demand-driven, not seasonal, not consumer-driven. It's industrial infrastructure-driven." Just as the world needs energy, the internet, highways, and railroads, it now needs a layer of AI intelligence infrastructure, built on top of all existing infrastructure. This layer needs chips.

Regarding the risk of customers borrowing money to buy NVIDIA products, he said he wasn't too worried. "These companies are exceptional; they generate massive amounts of cash." He specifically mentioned the startup of the AI profit flywheel: AI is useful, so AI can make money. Coding Agents are extremely profitable; they do useful work in high-paying roles, and many companies are willing to pay hundreds of millions of dollars annually to enhance their coding capabilities. "This flywheel has already started."

4. Token Economics: Smarter AI is More Valuable

This was the most technically profound part of the conversation. The host asked: What makes you confident that tokens will become increasingly profitable?

Huang's explanation was vivid. "A token is an embedding. It embeds knowledge and intelligence. This number is not static, like pi. The intelligence encoded by this number gets smarter over time." When intelligence becomes smarter, it becomes more useful, more useful means more valuable, and more valuable means people are willing to pay more.

He compared the AI industry to the past software industry. The old software industry was "asset-light," which is why software companies had high gross margins. However, the software industry in the AI era will become more "asset-heavy" because producing modern software requires machines like the supercomputer in front of him to generate intelligence. Every industry will become more capital-intensive, but the rewards are incredible intelligence, productivity, and growth.

"We are laying the foundation and building the infrastructure. This is the largest-scale industrial infrastructure construction in human history."

Regarding doubts about whether AI has already peaked, his answer was: "It can't have peaked because AI's penetration into society and industries is just beginning." He also mentioned that in the past six months alone, $300 billion has been poured into venture capital and startups in the US, creating new jobs and new companies.

5. Trump, Regulation, and Government Equity Stake

Huang's assessment of Trump was surprisingly specific. "He's smart, remembers everything. My god, he does know numbers." He mentioned that Trump is the only president who can remember NVIDIA chip model numbers like H20, H200, and Blackwell, and even knew the next generation is called Rubin.

He recounted their first meeting: "He said he wanted to restore America's manufacturing capability, wanted to re-industrialize America, wanted a secure and resilient supply chain, and wanted to bring semiconductor manufacturing back to the US." He said the Fort Worth factory where they were being interviewed was a direct result of that conversation.

When asked what mistake the government should avoid, Huang showed rare anxiety. "I worry about over-regulation, over-correction." He said some companies want the government to help create regulations favorable to them. "I think we should compete openly." He dismissed the comparison of the AI race to a 100-meter sprint as "nonsense": "Whoever reaches the finish line first wins forever? That's nonsense. Ultimate victory depends on whether society uses the technology, regardless of who invented it. We didn't invent electricity, we didn't invent manufacturing, but America applied it faster and with more passion, which is why it became the America of today."

The host pressed: If Trump called asking for an equity stake in Nvidia, what would you say? Huang's answer was clever: "It's not necessary. America already has an equity stake in NVIDIA. We paid $10 billion in taxes last year, and we'll pay more this year. We create massive amounts of jobs and tax revenue. Moreover, don't forget, most Americans are in the stock market today; when the market rises, everyone benefits."

Regarding whether Anthropic's Mythos model should be open to everyone (currently only open to specific institutions), Huang's stance was very clear: "Of course it should be open to everyone." He said it's Anthropic's responsibility to ensure the technology is safe, just like any software—fix vulnerabilities as soon as they are found. He mentioned the Mythos jailbreak incident, saying "Everything is fine, you and I are still here chatting."

Regarding the issue of open-model companies distilling closed models and then reselling them, his attitude was balanced: "It depends on the terms of service. If the service provider is unhappy, they should contact that company. We have many traditional legal means to handle it." However, he emphasized that AI learning from other sources is a fundamental characteristic of intelligence. "AI has to learn from something. The original AI also scraped all existing knowledge on the internet. Now, AI generates more content than humans, and in a few years, 99% of the internet's content will be AI-generated. You are already constantly distilling the intelligence of other AIs."

6. The Robot ChatGPT Moment and the Era of Trillions of Agents

The host asked: When will the ChatGPT moment for robots arrive?

Huang gave a surprising answer: "The ChatGPT moment for robots has already arrived." He explained that the ChatGPT moment wasn't when AI became useful. When ChatGPT came out in 2022, it was just "interesting and surprising." True usefulness came four years later. Robots are currently in that "interesting and surprising" phase. You can tell a robot, "put the apple in the drawer," and it will reason through the task sequence, including opening the drawer before putting the apple in. "When you first see a physical robot truly do these things, it opens up your imagination for the future of robotics."

As for when they will become useful, he gave a timeframe of "three to four years" and said, "I wouldn't be surprised."

Regarding the era of Agents, he painted a stunning picture. Today, a billion people use computers, but most of the time they are idle. In the future, everyone will be assisted by a multitude of Agents running 24/7. "We will have 100 billion, a trillion Agents running all day and night. Smart Agents, less smart Agents, specialized Agents, super Agents, all kinds of Agents running around the clock. So, the number of computers we need will grow dramatically."

7. The CEO's Craft and Philosophy of Suffering

The conversation concluded by returning to Huang personally. The host mentioned he founded NVIDIA at 30, and today the company is worth $5 trillion but has only 50,000 employees, and he said it might only have 75,000 in ten years, aiming to be "as small as possible."

Huang defined the CEO's job as "strategy." "Strategy is using limited resources to achieve a future vision as efficiently as possible." He said he started this job at 30, making him probably the longest-serving CEO in tech history. "This is my craft. This is my kung fu."

Regarding the theory that "pain and suffering are the key to greatness," he corrected that it's not related to a specific painful event, but more of a continuous state. "No great athlete became great by accident. It's a lot of practice when no one is watching, a