Source: All-In Podcast
Compiled by: Felix, PANews
This week's NVIDIA GTC 2026 conference, a global technology event, attracted participants from almost every industry, technology company, and AI company. NVIDIA founder and CEO Jensen Huang also delivereda keynote speech .
During the four-day conference, Jensen Huang gave an exclusive interview to the All-In Podcast at the GTC conference, covering topics such as NVIDIA's future, physical AI, the rise of intelligent agents, the explosive growth of inference capabilities, and AI public relations crises. PANews has compiled the interview, and the following are some highlights.
Host: One of the best announcements of the past year was the acquisition of Groq. Did you realize at the time that Chamath (one of the podcast hosts and CEO of Social Capital) would be extremely frustrated?
(PANews note: Social Capital was an early investor in Groq. Their frustration wasn't due to losing money, but rather stemmed from their personality and investment style: they were happiest before milestone events, not after.)
Jensen Huang: I had a premonition. After all, we're friends with Chamath, and we deal with him every week. The two weeks after the acquisition were definitely not comfortable. In fact, many of our strategies were publicized years ago. Two and a half years ago, I introduced Dynamo, the operating system for AI factories. Dynamo is a machine invented by Siemens that converts water into electricity, powering factories of the last industrial revolution. I thought it was the perfect name for the operating system for the factories of the next industrial revolution. Inside Dynamo, the underlying technology is "decoupled inference." Today's inference processing is an extremely complex computational problem, involving large-scale mathematical operations of all shapes and sizes. Our idea is to decouple the processing, allowing one part to run on some GPUs and another part on other GPUs, thus achieving heterogeneous computing. Today, NVIDIA's computing is distributed across multiple parts such as GPUs, CPUs, switches, and network processors, and now with the addition of Groq, the goal is to place the right workload on the right chip. We have evolved from a GPU company into an AI factory company.
Host: You've said on stage that 25% of data center space should be allocated to Groq and processors like this. How does the industry view this idea? How do you think people will react to this?
Jensen Huang: When we added this technology, the industry was shifting from large language model processing to agent processing. Running agents requires access to working memory, long-term memory, and various tools, which puts immense strain on storage. Data centers contain various models, such as very large models, small models, diffusion models, autoregressive models, etc. We developed the Vera Rubin architecture to run this extremely diverse workload. Our potential market size (TAM) has therefore increased by approximately 33% to 50%. A large portion of this increase will come from storage processors (Blue Field), Groq processors, CPUs, and network processors. All of these will work together to run the “agent” computer that drives the AI revolution.
Host: What about embedded applications? For example, if my daughter's teddy bear wants to talk to her, would it contain a custom ASIC, or would there be different development tools for edge and embedded applications?
Jensen Huang: Overall, solving this problem requires three computers working together. The first is used to train and develop AI models. The second is used to evaluate robots (such as cars, robots, etc.) in a virtual environment that follows the laws of physics. The third is an edge computer, or robot computer. It could be a self-driving car, a robot, or even a small computer in a teddy bear. In addition, we are working to transform the $2 trillion telecommunications base station industry into part of the AI infrastructure, with future radio base stations becoming edge devices. So these three basic computers are all necessary.
Host: You once predicted that demand for inference technology would explode by 1000 times or even 1 billion times. Now, there are voices saying that your inference factory cost as much as $40-50 billion, while competitors only need $25-30 billion. Do you think customers will be willing to pay this double premium? Will this affect your market share?
Jensen Huang: Never equate the cost of building a data center with the cost of generating tokens. I can prove that a $50 billion factory can generate the lowest-cost tokens for you because our production efficiency is extremely high. Even the basic costs—land, electricity, storage, network, servers, and cooling—are fixed (around $20 billion). If you factor that in, the difference in GPU prices, spread across the overall cost, might only be a difference between $50 billion and $40 billion, which isn't a large percentage. But our $50 billion data center provides 10 times the throughput of other solutions. In this industry, if your technology doesn't keep up with the pace of development, even if chips are given away for free, they won't be cheap enough.
Host: As the CEO of the world's most valuable company (with projected revenue of $350 billion next year), how do you make decisions, gather information, and determine which areas to invest in or exit?
Jensen Huang: Defining vision and strategy is the CEO's primary responsibility. We primarily rely on top computer scientists and engineers both inside and outside the company for information, but it is we who must shape the future. Our criterion for evaluating a new direction is: Is it unprecedented and extremely difficult? If something is easy to do, there will be countless competitors; if it is extremely difficult and happens to align with our company's unique "superpower," that's the intersection we're looking for. Because it's unprecedented and extremely challenging, the process will inevitably involve a lot of pain and hardship, and you'd better enjoy the process.
Host: Regarding long-tail businesses, could you talk about the long-term viability and growth potential of areas such as space data centers, automotive ADAS, or biology?
Jensen Huang: First, there's physics-based AI. For the first time, the tech industry has the opportunity to solve a $50 trillion traditional industry that has been largely untouched by technology. We started 10 years ago, and now it's exploding. It's already a nearly $10 billion business for us, growing exponentially. Second, there's digital biology. We're very close to the "ChatGPT moment" in digital biology. In the next 2 to 5 years, we will be able to use AI to represent and understand the dynamics of genes, proteins, and cells, which will completely transform the healthcare industry. The agricultural sector is also experiencing explosive growth.
Host: We see many enthusiasts and innovators obsessed with desktop open-source agent systems like "OpenClaw". What does this grassroots-driven open-source agent movement mean for you and the industry?
Jensen Huang: There have been three major turning points in the past two years: First, ChatGPT democratized generative AI; second, models with reasoning capabilities enabled AI to not only answer questions but also provide more practical answers; and third, revolutionary agent systems like "Claude Code" emerged within the industry. However, Claude Code was initially only geared towards enterprises, and it wasn't until OpenClaw appeared that the public truly realized what AI agents could do. More importantly, this type of system fundamentally reshapes the computing paradigm. It possesses short-term memory (scratch disk), can manage resources, perform task scheduling, create sub-agents to solve problems, and can run various applications (skills) through APIs. These elements define a computer. This means that for the first time, we have a blueprint for an open-source personal AI computer that can run anywhere, which will become the operating system of modern computing. Of course, software with such high privileges needs to be well-governed. We have invested a lot of engineers working with Peter Steinberger (founder of OpenClaw) and others to ensure that agents are subject to good security governance and privacy protection.
Host: Does the speed of this paradigm shift in AI render recent AI regulatory bills meaningless? Regarding the panic caused by AI, such as the PR turmoil surrounding Anthropic, if you were a member of Anthropic's board of directors, what advice would you give their team to change public perception?
Jensen Huang: We need to continuously educate policymakers about the current state of technology: AI is just computer software; it's not an alien life form, it has no consciousness, and it's not, as some people say, "something we know nothing about." We cannot allow doomsday theories and extremism to dictate policy. But policy also cannot outpace technology too quickly. The biggest national security concern right now is that while we hesitate to use AI out of fear, anger, or paranoia, other countries are actively adopting the technology.
As for Anthropic, their technology is superb, and their focus on security and preparedness is admirable. Warning people about the potential of technology is good, but “scare tactics” are not. As technology leaders, because our industry is crucial to national security and societal fabric, our voices carry immense weight. Therefore, when predicting the future, we need to remain humble, more balanced, moderate, and thoughtful, avoiding making extremely catastrophic statements without evidence.
Host: We really need to be more proactive in promoting AI. Regarding the productivity gains brought about by the explosion of intelligent agents, there's currently debate about whether AI has a return on investment (ROI). Seeing the explosive growth of OpenAI and Anthropic, do you think our revenue scale can keep up with the expansion of intelligence levels?
Jensen Huang: Looking around, you'll see representatives from Anthropico and OpenAI, but in reality, 99% of AI companies are involved, and Anthropico and OpenAI are not just one of them. Currently, OpenAI is the most popular, followed by open-source models, with Anthropico ranking third. This illustrates the vast and diverse nature of the AI ecosystem. From generative AI to inference AI, the computational load increases 100-fold; from inference to agent AI, the computational load may increase another 100-fold. People are willing to pay for information, but they are even more willing to pay for "getting the job done." Agent systems can genuinely help software engineers get the job done. So, when your computational load increases 10,000-fold, the energy consumption may increase 100-fold. Our current scaling is just beginning.
Host: In your speech, you mentioned that Nvidia pays a lot of token fees for its engineering teams. Roughly estimating, each engineer needs about 75,000 tokens. Are you currently spending $1-2 billion annually on tokens for your engineering teams? What will the efficiency of these engineers be in two or three years?
Jensen Huang: Let's do a thought experiment: Suppose you pay a top software engineer or AI researcher a $500,000 annual salary. If at the end of the year they tell you they only spent $5,000 in tokens, I'd be furious. If this $500,000-salary engineer didn't consume at least $250,000 in tokens, I'd be deeply shocked and worried. It's like a chip designer refusing to use CAD tools, insisting on using paper and pen. This is also about equipping these outstanding knowledge workers with "superhuman" abilities, just like James spending $1 million a year on health maintenance.
The paradigm shift of the future will see the complete disappearance of ideas such as "This is too difficult," "This takes too much time," and "This requires too many people." Bottlenecks in work will depend solely on your creativity. Future programming will no longer be about writing code, but about writing ideas, architectures, and specifications. We will organize teams, define what constitutes a good outcome, guide how to evaluate it, and brainstorm and iterate with intelligent agents. I believe every engineer will have hundreds of intelligent agent assistants.
Host: We've seen incredible efficiency on many technical levels. For example, a CEO spent 90 minutes over the weekend replacing an entire software stack with Claude and intelligent agents, or Auto Research completed a doctoral dissertation-level study that would normally take seven years in 30 minutes. Does this mean the enterprise IT software industry is going to be destroyed?
Jensen Huang: OpenClaw is so amazing because its timing perfectly aligns with the breakthrough in large-scale language models and the new capabilities of models using tools. Some say the enterprise IT software industry will be destroyed, but there's another perspective: in the past, enterprise software was limited by the number of employees, but in the future, there will be hundreds of times more intelligent agents using these tools. They will use tools like SQL, vector databases, Blender, Photoshop, or CAD because these tools perform well and are the "pipelines" connecting humans to work results. I need AI to put the work results back into tools like Synopsis or Cadence because that's how I can control and verify them.
Host: Recently, the crypto project Bit Tensor successfully trained a 4 billion-parameter LLaMA model using a distributed approach. What are your thoughts on the ultimate form of open-source models? Will decentralized computing power and fully open-source methods be the mainstream in the future?
Jensen Huang: We need both proprietary models as best-in-class products and open-source models; they coexist. Because a model is a technology, not a product or service. For the average consumer, the experience of using ChatGPT, Claude, or Gemini—which offer different services—is great. However, all industries worldwide need to embed domain-specific expertise into models they can fully control, and this can only be achieved through open-source models. Almost all the startups we invest in now adopt an "open-source first" strategy before gradually transitioning to proprietary models.
Host: Last year, the Biden administration's policies restricted the global spread of AI. Now that the new president has taken office, how would you assess our performance in promoting American AI technology globally?
Jensen Huang: President Trump wants American industry and technology to remain at the forefront of the world, to win, and to become the richest nation. Currently, Nvidia, which once relinquished 95% of its market share in China, now has 0%. President Trump wants us to return to that market. We have applied for and obtained licenses to sell to relevant companies through Secretary Lutnick and have received purchase orders; we are currently restarting the supply chain.
On a national security level, national security is compromised when we lose control of micromotors, rare earth minerals, telecommunications networks, or energy. I don't want the AI industry to follow in their footsteps. We can't expect the whole world to use only one universal AI model, but we can allow the "American technology stack," which includes chips, systems, and platforms, to occupy 90% of the global market share, enabling countries around the world to build public or private AI applications that suit their societies. That's the outcome we want.
Host: You have many partners in the field of autonomous driving, such as Mercedes-Benz and Uber. Are you planning to create an open-source platform like Android, or a closed ecosystem like Tesla's iOS?
Jensen Huang: We believe that all moving objects will be automated to varying degrees in the future. We don't want to build cars ourselves; instead, we want to empower every car company in the world to manufacture autonomous vehicles. Therefore, we've built training computers, simulation computers, onboard evaluation computers, and the world's first autonomous driving operating system with inference capabilities, capable of breaking down complex scenarios into simpler ones, just like a human. Through vertical optimization and horizontal innovation, we let our customers decide: someone like Musk (Tesla) might only buy our training computer, while other customers might want to buy our complete hardware and software system. Our attitude is problem-solving; we welcome any choice our customers make in partnering with us.
Host: Many major customers, such as Google and Amazon, are also developing their own AI chips to compete with you. Meanwhile, Wall Street analysts predict that your growth rate will drop to 7% by 2029 and you will lose market share. What do you think?
Jensen Huang: We are an AI company that builds the entire stack and the underlying model, and the only company in the world that collaborates with all AI companies globally. They never show me what they're building, but I always show them what I'm building. As long as we move fast, buying NVIDIA products remains the most economical option. NVIDIA is the only architecture that can be deployed in any cloud, on-premises server, car, or even space. About 40% of our business customers don't want to buy chips; they want to build AI infrastructure and need the full CUDA stack, which we have. So NVIDIA's market share has actually increased rather than decreased. For example, AWS just announced it will purchase 1 million chips over the next few years, and Meta and Anthropic are also turning to NVIDIA.
As for Wall Street analysts, they simply don't understand the sheer scale and breadth of AI. Based on stereotypes, they don't believe a market can grow from $5 trillion to $15 trillion. Most people think AI is concentrated in the hands of the top five cloud service providers, but the impact of AI will be far greater than what OpenAI or Anthropic currently demonstrates. Nvidia is no longer just solving chip manufacturing problems, but the extremely complex AI infrastructure issues.
Host: Can you explain your business in space data centers to people who aren't professionals?
Jensen Huang: We certainly need to prioritize solving problems on Earth, but we should also prepare for space, because space has abundant energy resources. The main challenge lies in heat dissipation: heat dissipation in space cannot be achieved through conduction or convection, relying solely on radiation, which requires a huge surface area. However, space is the one thing we have no shortage of in space. We have already entered space, with radiation-shielded CUDA equipment performing AI image processing on satellites around the world. Much image work can be completed directly in space without transmitting it back to Earth. Exploring the architecture of space data centers takes time, but we have ample time to do so.
Host: The systems in the healthcare field are extremely bloated. How can AI overcome regulations and play a real role in this field?
Jensen Huang: In the medical field, we primarily focus on three areas: First, AI in biology: used for drug discovery, predicting and understanding biological behavior through AI. Second, AI agents: companies like Hippocratic are developing assistants to aid in diagnosis, which is dramatically changing how we interact with doctors. Third, physics AI, AI that understands the laws of physics, used in robotic surgery. In the future, every instrument in hospitals (ultrasound, CT, etc.) will have a built-in secure OpenClaw agent, interacting with patients, nurses, and doctors in entirely new ways.
Host: The humanoid robot industry experienced a "lost decade," but now we are seeing the amazing performance of Musk's Optimus and Chinese companies. How far are we from robots entering our lives?
Jensen Huang: The US actually invented this industry a long time ago, but we entered the market too early. We were already exhausted five years before the enabling technology (AI brain) appeared. But now the brain technology is in place. From the emergence of proof of high functionality to the actual launch of a reasonable product, it usually only takes two to three cycles, that is, about three to five years, and we will see robots everywhere.
China is incredibly powerful; its microelectronics, electrical engineering, rare earth, and magnet technologies are world-leading, forming the foundation of the robotics industry. To a large extent, the global robotics industry will be deeply reliant on China's ecosystem and supply chain. Robots will solve the labor shortage problem; we could even control robots in our homes via virtual reality to help with housework, or serve as advance labor for our interstellar migrations (such as to the moon and Mars).
Host: Anthropic CEO Dario predicts that revenue from non-infrastructure AI applications (models and agents) will reach trillions of dollars by 2030. In the future, how can companies at the software application layer build a competitive moat? Facing inevitable unemployment (such as drivers), what learning advice do you have for young people about to enter the workforce?
Jensen Huang: I think Dario's prediction is very conservative; they will do better. He didn't consider that in the future, every enterprise software company will become a value-added reseller (VAR) of these underlying big models (such as Anthropic and OpenAI), which will greatly expand the market.
The real moat at the application layer is deep specialization . General cloud models connect to software companies' agent systems, but you must train domain-specific sub-agents with your own data. Connecting your agents to customers early on will accelerate the flywheel effect within specialized domains, giving software platforms the opportunity to become experts in their respective verticals. Regarding employment, jobs will indeed change; some will be eliminated, but many new jobs will be created. For example, with the widespread adoption of autonomous driving, existing drivers may become in-car "travel assistants," handling luggage or arranging hotel itineraries, just as autopilots on airplanes have created more demand for pilots.
My advice to young people is: delve deeply into science and mathematics, and language skills are also crucial. Because language is the ultimate programming language for AI. Furthermore, regardless of your education, you must become an expert in using AI proficiently. When deep learning was just emerging, some experts predicted radiologists would lose their jobs. But 10 years later, with computer vision 100% integrated into medical platforms, the demand for radiologists has surged. This is because faster scanning speeds allow hospitals to see more patients, increasing revenue and thus hiring more doctors. Similarly, increased productivity makes countries wealthier, enabling them to have more teachers in classrooms and, with AI, create personalized courses for each student. Every student needs excellent teachers. Regarding AI, we don't need to spread doomsday fears; we can choose how to use this technology to create a better future.
Related reading:The AI industry also has its own Satoshi Nakamoto – Jensen Huang
