Source: Smart Investor
$NVIDIA (NVDA.US)$ Still firmly at the top of the global market capitalization. However, in the eyes of its founder Jensen Huang, this behemoth continues to maintain the relentless mindset of a 'small company' as it did 30 years ago.
Following the recent conclusion of GTC 2026, Jensen Huang appeared on the podcast of his old friend Ben Thompson for an in-depth review.
The value of this dialogue lies in Jensen Huang's repeated dissection of NVIDIA’s 'trump cards.' He is not only building computing power but also defining a set of survival rules that make everything accelerable.
In this conversation, there are three key takeaways for investors to focus on:
First, from the thriving CPU business (such as Vera) to the critical acquisition of Groq, these moves may seem to deviate from the pure GPU track, but in reality, they represent Jensen Huang’s ultimate completion of 'accelerated computing.' He stated frankly that NVIDIA has never been against CPUs, but in the era of AI agents, expensive computing power must not be wasted due to the mediocrity of CPUs.
Second, why was it difficult for AI to generate revenue in the past? Jensen Huang pointed out that when AI evolves from simple information retrieval to executable tasks like programming and architectural design, its economic value truly explodes. This leap from 'probabilistic correctness' to 'task delivery' marks the watershed moment for AI commercialization.
Moreover, regarding the Chinese market, Jensen Huang demonstrated his consistent perspective and foresight. He specifically praised Chinese indigenous forces such as DeepSeek, Kimi, and Qwen, and warned Washington not to be frightened by the 'doomsday' sci-fi hype. If technology is not allowed to diffuse through competition, the United States risks being left behind by the times.
At the end of the interview, Jensen Huang said,
"We are still sprinting with all our might. We are not behaving like a giant who owns a deep moat and can relax in comfort."
On the contrary, we are starting over—redefining networks, redefining CPUs, and reshaping software stacks. In order to redefine the inference architecture, we are even willing to break the status quo through acquisitions.
This in no way resembles a company that believes it has 'already won everything.'
It is highly admirable.
Smart investors have compiled and shared this with everyone.
01, NVIDIA's CUDA Core
Ben Thompson: In this talk, you spent a lot of time explaining what NVIDIA really is, starting from the history of programmable shaders to the release of CUDA 20 years ago. Why do you think it's necessary to retell this story now? Why redefine CUDA and its significance?
Jensen Huang: Well, because we are entering a wide range of entirely new industries.
Moreover, AI will learn to use tools that were originally created for humans. AI can use Excel, Photoshop, logic synthesis tools, Synopsys tools, and Cadence tools.
These tools must be 'super-accelerated,' and the databases they use must also be super-accelerated because AI operates at an extremely high speed.
Therefore, I believe that in this era, we need to accelerate all software in the world as much as possible and then present them to AI so that it can invoke them as an 'intelligent agent.'
Ben Thompson: So, does this mean that we have already prepared for some industries and are now expanding to more industries?
Jensen Huang: That's correct, much more. For instance, data processing.
Ben Thompson: It was indeed unexpected. I didn’t anticipate you’d start by discussing the partnership with IBM.
Jensen Huang: Yes, it helps everyone see the bigger picture. After all, IBM is the pioneer of all this.
Ben Thompson: You wrote last week that AI is a 'five-layer cake': electricity, chips, infrastructure, models, and applications. Are you concerned that over the past four or five years, NVIDIA might be pigeonholed into just the 'chip' category? Hence, emphasizing that NVIDIA is a vertically integrated company is important.
Jensen Huang: My way of thinking doesn’t begin with 'what we are not,' but rather with 'what we need to become.'
Many years ago, we realized that accelerated computing is a full-stack problem. You must understand the application to accelerate it. We recognized the need to master application logic, possess a developer ecosystem, and excel in algorithm development.
Since those old algorithms developed for CPUs do not perform well on GPUs, we had to rewrite and restructure them to enable acceleration by our GPUs.
But if we do this, we can achieve speedups of 50x, 100x, or even 10x, which is absolutely worth it.
I think from the very beginning, we realized: 'What do we want to achieve? What will it take to get there?'
Today, we are building AI factories and AI infrastructure across the globe. This is far more complex than simply manufacturing chips.
Of course, making chips is crucial; it is the cornerstone.
Ben Thompson: Exactly, it’s a full stack for networking, storage, and now you’ve also ventured into CPUs.
Jensen Huang: Now you have to integrate all of that into these massive systems.
A gigafactory may cost between 50 and 60 billion US dollars. Of this amount, approximately 15 to 17 billion US dollars goes toward infrastructure, including land, electricity, and factory construction. The rest is for computing, networking, storage, and similar components.
Faced with this level of investment, unless you can give customers enough confidence that they will be able to successfully build and operate, you won’t stand a chance. No one will risk 50 billion US dollars.
So I think the core idea is that we must not only help our customers make chips but also help them build systems; after building the systems, we must help them create AI factories.
An AI factory contains a lot of software, not just our software, but also extensive software used for cooling management, power control, and redundancy design.
In the past, due to a lack of communication across various stages, many designs were overly redundant.
When a group of people who don’t communicate integrate a system, by definition, they must over-design their respective parts. But if we collaborate as a team, we can ensure pushing the limits, achieving higher throughput from the existing power supply or saving costs under the same throughput requirements.
02, Software, Inference, and New Architecture
Ben Thompson: Returning to the software aspect, you mentioned that Excel was not originally designed for AI use. Now, models like Claude have introduced new features that can invoke Excel.
When you talk about investing in these areas, is it to make such models perform better? Or is it intended for$Microsoft (MSFT.US)$enterprise clients, meaning that you want to use these tools but don’t want to be constrained by other players within the ecosystem?
Jensen Huang: Well, SQL (Structured Query Language) is a great example.
SQL is designed for human use, and we run our operations on SQL systems just like everyone else. It is the 'source of truth' in the business world.
Now, it’s no longer just humans querying our SQL databases; there will be a large number of intelligent agents performing queries on them.
Ben Thompson: Exactly, they will query much faster.
Jensen Huang: Their demand for speed will be much higher. So, the first thing we need to do is accelerate SQL, which is the simplest logic.
Ben Thompson: That makes perfect sense. Regarding models, you mentioned that language models are just one category. In your article last week, you wrote: 'Some of the most transformative work is happening in areas like protein AI, chemical AI, physics simulations, robotics, and autonomous systems.'
You also touched on this point in other keynote speeches before, and I remember you used a phrase: 'Everything is a token.' Do you think the Transformer architecture is the key to unlocking everything? Or do we need new foundational breakthroughs to support these applications?
Jensen Huang: We need a variety of new models. For example, the attention mechanism capability of Transformers grows quadratically. So how do you achieve ultra-long memory? How do you sustain a long conversation without the KV cache becoming cluttered over time?
Ben Thompson: Or, to avoid filling an entire rack of solid-state drives just for the KV cache.
Jensen Huang: Exactly. Moreover, suppose you could record all our conversations. When you go back to reference a particular dialogue, which part of the reference is the most important?
A new architecture is needed that can handle attention correctly and process information at extremely high speeds.
We have developed a hybrid architecture that combines Transformers and SSM (State Space Models), enabling Nemotron 3 to achieve both ultra-high intelligence and ultra-high efficiency. This is one example.
Another example is developing models with geometric awareness. This means that many things in life and nature are symmetrical.
When generating these models, you don't want them to only produce statistically plausible outputs; they must also conform to physical laws, such as symmetry. For instance, our cuEquivariance technology achieves this.
So we possess all these different technologies. Another example is when generating text tokens, the output is chunked, bit by bit, one token at a time. But when generating motion, it needs to be continuous.
Thus, there is discrete information that you need to generate and understand, as well as continuous information that you want to generate and understand. The Transformer architecture cannot perfectly handle both simultaneously.
03. About Inference and Programming
Ben Thompson still wants to quote a sentence from your article: 'In the past year, AI has crossed a critical threshold. The models have become good enough for large-scale application. Reasoning has improved, hallucinations have decreased, and practical applications have seen significant enhancements. For the first time, AI has started to generate real economic value.'
I would like to discuss specifically what changes have occurred. Given the current momentum, this year is undoubtedly the year of 'agents.' But for last year, was 'reasoning capability' really the most pivotal breakthrough?
Jensen Huang Generative technology is indeed a milestone, but it has a fatal flaw—it tends to produce nonsense with great seriousness (hallucination).
Therefore, we must make it more 'grounded,' and the way to do that is by introducing reasoning, reflection, retrieval, and search.
Without reasoning ability, none of these auxiliary methods can function. It is reasoning that allows generative AI to truly take root in reality.
Once it can be grounded in reality, you can use it to analyze problems, break down steps, and turn complex issues into manageable tasks. Naturally, the next phase evolves into 'tool usage.'
This actually reveals an interesting phenomenon. You see, search used to be a service that no one was willing to pay for because, although access to information is important, people are accustomed to not paying for it.
To get users to pay, the value you provide must far exceed the information itself. Most people will not pay for simple queries like 'Which restaurant is good?'
But now, we have crossed that payment threshold. AI is no longer just chatting with you or feeding you information; it can now genuinely 'get things done' for you.
Programming is a perfect example.
Think carefully, writing code and speaking are actually two different things. You have to teach AI how to handle spaces, indents, and symbols, which has almost become a completely new modality. Moreover, you cannot output tokens piece by piece like squeezing toothpaste; instead, you must refactor the entire block of code.
This code must be logically consistent, optimized, and most importantly—it must compile.
It cannot merely stay at the level of being 'probably correct'; it must be 'executable.'
Ben Thompson: Exactly, code is very honest—it either works or it doesn’t.
Jensen Huang: It must work. Therefore, teaching AI to program in this modality is an extremely significant milestone.
Once we possess this capability, consider this: we used to spend hundreds of thousands of dollars annually hiring engineers to write code, but now they have programming assistants. Engineers can focus on higher-level architecture design.
In the past, writing programs was labor-intensive and exhausting; now they can define software through 'specifications,' which is more abstract and efficient.
They spend their time solving problems and innovating. Currently, 100% of our company's software engineers use programming agents. Many of them haven't manually written a single line of code in a long time, yet their efficiency has skyrocketed, keeping them busy every day.
Ben Thompson: But aren’t you worried that people might overestimate AI’s capabilities because programming is 'verifiable'? After all, agents can repeatedly trial and error until the code works, often without human intervention.
Jensen Huang: The key lies in this ability to 'reflect.' For instance, designing a house or a kitchen used to be the exclusive domain of architects, but now carpenters can do it too.
You have significantly elevated the upper limit of a carpenter's capabilities through AI. A carpenter can use an intelligent agent to conceptualize plans and design styles. Although the intelligent agent does not hold a hammer (physical tool), it can continuously iterate.
For example, you give it a reference: 'This is the aesthetic style I want.'
The intelligent agent will constantly self-reflect, comparing its plans with the reference image. It might say, 'Hey, this part didn’t turn out as expected; I need to redo it.'
It keeps self-iterating. The plans it generates don’t necessarily have to be one hundred percent 'executable' (in terms of physical realization). In fact, in areas involving probability, aesthetics, and subjective judgment, AI performs even more impressively.
Ben Thompson: This is quite fascinating. You’ve almost covered two extremes here: one end is generating images, where there is no definitive answer; the other is programming, where right and wrong are absolutely clear. AI seems to excel at both ends. The suspense now is how quickly it can replicate that success across the vast intermediate zone.
Jensen Huang: At the very least, we can confidently say now that it can handle architectural design and also the decoration of kitchens and living rooms.
Section 04: The Role of the CPU in Accelerated Computing
Ben Thompson: Speaking of which, with the rise of intelligent agents, there’s something interesting going on. You used to always talk about 'accelerated computing,' and I remember you often dismissed CPUs, saying they would eventually be replaced as everything could be accelerated. But just like that, CPUs have made a comeback.
Since you’re now personally involved in selling CPUs, it’s obvious that they have become both highly useful and critical. So, what’s it like being a 'CPU salesperson'?
Jensen Huang: Without a doubt, Moore's Law has reached its end. However, one thing needs clarification: accelerated computing is not synonymous with parallel computing.
Looking back 30 years ago, there were at least two or three dozen companies in the market working on parallel computing, but in the end, only NVIDIA survived.
The reason is that we saw it clearly back then: our goal was never to replace the CPU, but to accelerate applications.
Ben Thompson: So, the 'accusation' I just made against you, although not valid in your case, would certainly apply to other companies.
Jensen Huang: We have never been against the CPU, nor did we want to challenge Amdahl's Law (which states that the speed of a system ultimately depends on the part that cannot be accelerated).
In the logic of accelerated computing, we actually choose to buy the top-tier CPUs within the system, even if they are the most expensive. The reasoning is simple: if the CPU's performance is not at the highest level, it will, in turn, become a bottleneck and hinder the performance of my chips, which are worth millions of dollars.
Ben Thompson: In the past, when dealing with branch prediction, you were concerned about wasting CPU time; now, what pains you is the waste of GPU time.
Jensen Huang: Exactly. You must never let the GPU be idle or left spinning without work.
That's why we've always pursued the strongest CPU, even going so far as to create our own Grace processor, all to achieve the best single-threaded performance and make data transfer extremely fast.
Thus, accelerated computing has never rejected the CPU. My basic view remains unchanged: Amdahl’s Law cannot be bypassed. The old approach of relying on general-purpose computing and stuffing more transistors has completely died out.
But fundamentally, we have no issues with the CPU.
These intelligent agents have now learned to use tools. The problem is that these tools were originally designed for humans. They are roughly divided into two categories: one runs in data centers, mostly related to SQL and databases; the other operates on personal computers.
What we are facing are AI systems capable of learning 'unstructured operations.'
You see, the first category of tools is 'structured,' such as command-line interfaces (CLI) or application programming interfaces (APIs). These tools have clear commands and precise parameters, with rigid rules governing how you interact with applications.
However, the challenge lies in the vast number of applications that were never designed with any interfaces. This requires AI to possess multimodal capabilities to learn those 'unstructured' operations.
It needs to act like a human, navigating web pages, finding buttons, pulling up menus, and figuring things out step by step. Such tasks absolutely require a PC.
Our current strategy is to tackle both fronts: on one hand, we have top-tier data processing systems; on the other hand, as you know, NVIDIA's workstations (PCs) rank among the best in the world in terms of performance.
Ben Thompson: So, what is the difference between a CPU designed with intelligent agents at its core and a regular CPU? From the sound of it, you plan to deploy an entire rack of Vera CPUs.
Jensen Huang: Oh, that’s a very sharp question. Over the past decade, the design philosophy of CPUs has revolved entirely around 'hyperscale cloud.'
Cloud providers monetize based on the number of CPU cores, so the design goal has been to maximize the number of cores, as they are primarily used for renting purposes. Whether single-core performance is strong or not is secondary.
Ben Thompson: After all, the main focus has been handling latency on the web end.
Jensen Huang: Exactly. What you are optimizing is the number of instances. That's why you see CPUs with two to three hundred, or even four hundred cores emerging.
But the trade-off is that their single-core performance is rather mediocre. However, for scenarios like 'agent invoking tools,' when the GPU is just sitting there waiting for the result of the tool operation—
Ben Thompson: And they are connected via NVLink, which is extremely fast.
Jensen Huang: Exactly. What you need most in this case is a computer with single-threaded performance that is as fast as possible.
Ben Thompson: So is this purely a speed race? Or does the CPU itself need to become more 'parallel' to solve problems such as cache misses? Or has the entire design logic of the pipeline changed?
Jensen Huang: Exactly. The most critical aspects are single-threaded performance and unbeatable I/O capabilities.
Because there are massive numbers of single-threaded instances running in data centers today, which puts enormous pressure on the I/O systems and memory controllers.
The bandwidth per core of Vera, and even the total bandwidth of the entire chip, is three times higher than any previous CPU. Its original intention was to ensure that the CPU would never become a bottleneck by leveraging massive I/O and memory bandwidth. Once the CPU slows down, it drags behind an entire group of GPUs.
Ben Thompson: This Vera cabinet you mentioned is tightly integrated with the GPU, but is it still 'decoupled' (meaning the originally tightly bundled components are separated, allowing them to operate and upgrade independently without interfering with each other)? In other words, can the GPU serve multiple different Vera cores simultaneously instead of being rigidly tied to a specific board…
Jensen Huang: Yes, they are decoupled.
Ben Thompson: I understand; the logic makes sense. So, what role does your collaboration with Intel play, and how does NVLink fit into this?
Jensen Huang: Good question. While many people are satisfied with the Arm architecture today, certain areas—especially enterprise computing—are still holding on to a lot of legacy technology stacks that they are reluctant to migrate away from, so x86 remains essential.
Ben Thompson: Does the resilience of x86 code come as a surprise to you?
Jensen Huang: Not at all. NVIDIA’s PCs are still based on the x86 architecture, and all our workstations are x86 as well.
05, Groq’s integration is a “strategic coincidence.”
Ben Thompson: In your presentation today, you referred to yourself as the 'King of Tokens.' You also mentioned in your article that electricity is the 'first principle' of AI infrastructure, directly limiting how much intelligence a system can produce. Given that token production is constrained by power consumption, why do other companies think they stand a chance and dare to challenge you, the 'King of Tokens'?
Jensen Huang: The challenge is significant because it’s no longer realistic to achieve groundbreaking breakthroughs by simply creating a single chip.
Take Groq for instance—unless its technology is paired with our Vera Rubin (NVIDIA's next-generation GPU architecture), Groq cannot achieve that extraordinary level of performance.
Ben Thompson: Tell me more about that. I was just about to ask about Groq.
Jensen Huang: When it comes to inference, there are two extremes: one is where you want to achieve the highest throughput, meaning how many tokens are generated per unit of time; the other is where you aim to deliver the 'smartest' possible tokens. The smarter the tokens, the higher the premium you can charge.
Striking a balance between the two—maximizing throughput while withstanding the pressures of intelligence—is an exceedingly difficult challenge.
Ben Thompson: I must remind you, last year you had a slide specifically discussing the 'Pareto curve.' When you were promoting the Dynamo compiler, you claimed that NVIDIA's GPUs could perfectly cover the entire curve, and users could purchase with confidence because Dynamo would handle everything. But now you're changing your tune to say, 'Well, it can't quite do everything after all.'
Jensen Huang: Our coverage capability remains the strongest in the industry. The reason we are further broadening this 'Pareto frontier,' especially in terms of ultra-high token rates and ultra-low latency, is the emergence of programming agents.
These agents can create tremendous economic value, and they are designed to serve 'humans,' who are the most expensive and valuable resource in the world.
Ben Thompson: Exactly, humans are far more expensive than GPUs.
Jensen Huang: Therefore, I want to provide my software engineers with the highest possible token rate.
If Anthropic releases an advanced version of 'Claude Code' that can increase programming speed by 10 times, I would not hesitate to purchase it.
Ben Thompson: So, are you developing this product primarily for your own use?
Jensen Huang: Most great products originate from pain points you personally experience.
We hope our programming agents can become 10 times faster, but achieving this in a high-throughput system is almost an impossible task.
Therefore, we decided to integrate Groq’s low-latency system to enable both systems to work collaboratively.
Ben Thompson: I see. Is this merely about separating the Decode and Prefill processes?
Jensen Huang: It goes even further. We will specifically isolate and handle the computationally intensive parts of the Decode phase that involve attention mechanisms.
Ben Thompson: So you have achieved 'decoupling' at the Decode level.
Jensen Huang: Exactly. This requires extremely tight coupling and deep software integration.
Ben Thompson: How did you move so quickly? It has only been a few months since the acquisition, and you are already able to ship products this year?
Jensen Huang: In fact, our research on 'Disaggregated Inferencing' began long ago.
When we launched Dynamo, NVIDIA had already revealed its hand. The astute should have realized then that I was already contemplating how to achieve finer-grained inferencing splits on heterogeneous infrastructure.
Groq’s architecture is simply an extreme variant of our architecture; it was indeed challenging for them to operate independently before.
Ben Thompson: It seems you have been planning this for three to five years. Was there any specific event that pushed you to acquire the Groq team this time?
Jensen Huang: Strategy always comes first.
Before the deal was announced, we had already been collaborating for half a year. Integrating Grace Blackwell and Groq was a direction we had set long ago.
As for the collaboration itself, I have great admiration for their team, but I’m not interested in their cloud services. They also have another business segment they are confident about and excel in—let them keep that part.
We acquired the team, obtained the license, and began evolving from the foundational architecture.
Ben Thompson: So this is just a beautiful coincidence?
Jensen Huang: It should be called a 'strategic serendipity.'
Ben Thompson: Because OpenAI just announced a partnership with Cerebras in January.
Jensen Huang: That’s their business. I wasn’t even aware of it at the time, and knowing wouldn’t have changed our decision. Groq’s architecture remains my top choice—it makes more logical sense for us.
Ben Thompson: Is this the first ASIC (Application-Specific Integrated Circuit) solution that made you think, 'This is interesting; it really stands out'?
Jensen Huang: No, the previous one was Mellanox.
Ben Thompson: Indeed, that was a stroke of genius.
Jensen Huang: Right, Mellanox. We embedded part of the compute stack directly into its network stack.
Without this 'in-network computing,' NVLink could never have reached its current scale. NVIDIA excels at decomposing the software stack and placing it where it belongs.
We don't care where the computation happens; we only care about accelerating applications. Remember, NVIDIA is an accelerated computing company, not just a GPU company.
Ben Thompson: Understood. Given power as a hard constraint, should customers believe that this LPU (Language Processing Unit) cabinet will generate more revenue for them compared to traditional GPUs?
Jensen Huang: It depends on their product. If you don’t have enterprise use cases right now, buying Groq doesn’t make much sense because most of your customers are free users, and you’re busy converting them into paying ones.
If the majority of your user base consists of free users, adding Groq will only increase costs and power consumption, making it uneconomical.
Ben Thompson: Not to mention the added complexity and server rack space required.Opportunity cost。
Jensen Huang: Exactly. But if you're like Anthropic or OpenAI’s Codex, with extremely high-value business operations limited only by token output speeds, then adding this accelerator will truly supercharge your revenue.
06, Competition in the Chinese Market and Washington's 'AI Doomsday Advocates'
Ben Thompson: By this day in 2026, what will be the limiting factor for us? Will it be electricity, wafer capacity, or something else? Everyone is shouting about shortages; where is the real 'ceiling'?
Jensen Huang: I think almost every aspect is nearing its limit. At this point, trying to double any one of them is basically impossible.
Ben Thompson: Because as soon as you double one, you immediately hit another bottleneck.
Jensen Huang: Exactly.
Ben Thompson: However, it seems that the U.S. is doing relatively well in terms of power generation, more optimistically than what people thought two years ago. Now, it feels like chip supply has become the bottleneck.
Jensen Huang: Our supply chain plans are actually very tightly scheduled.
You should know, we've been preparing for a 'mega year' for quite some time, and even next year's plans have already been arranged.
Ben Thompson: Understood. We’ve also seen the photos of you drinking soju and eating fried chicken with those supply chain leaders.
Jensen Huang: (Laughing) Yes, exactly. We have hundreds of partners globally, and we’ve built long-standing relationships with all of them. So, I’m confident about our supply situation.
I don’t see electricity or chips being overabundant—nothing is excessive. But from my perspective, we can fully handle the situation at the supply chain level.
My only wish now is that the land, power facilities, and factory shells can be completed even faster.
Ben Thompson: Can I understand it this way? As long as 'scarcity' still exists, NVIDIA is the biggest winner? For instance, if there isn't enough electricity, because your chips are the most energy-efficient, everyone will have to buy yours; or, if production capacity is limited, you've already secured the upstream supply chain. In such a situation, aren't you dominating the market?
Jensen Huang: Let's put it this way, we are the leading player in this field, and our planning started early.
We have positioned ourselves across the entire supply chain, upstream and downstream. Rather than calling us winners, it’s more accurate to say that we’ve ensured all our partners are prepared for this explosive growth.
Ben Thompson: Understood. But at the end of the day, wouldn't losing the Chinese market be a fatal blow? If China eventually manages to secure sufficient electricity and produces enough chips—even if they're just 7nm—they would have the resources to create an ecosystem capable of competing with CUDA. Aren’t you concerned about that?
Jensen Huang: Without a doubt, we must ensure that American technology stacks remain accessible in China. I’ve consistently held this view because open-source software cannot be stopped.
China contributes more to open source than any other country, and half of the world’s top AI talent resides there, demonstrating remarkable creativity. Take DeepSeek, for example—it’s no joke, incredibly impressive. Kimi and Tongyi Qianwen are also exceptionally strong.
They have made original contributions in architecture and AI stacks, which deserve the highest level of attention from us.
As long as the world continues to develop using American technology stacks, when these innovations inevitably flow back from China (due to open source) to the U.S., Southeast Asia, or Europe, our ecosystem will be able to integrate seamlessly.
I have always believed that this represents the core geopolitical strategy for the American tech industry.
Ben Thompson: Indeed. The last time we spoke, the Trump administration had just banned H20. It was quite surprising that you managed to convince them to accept your logic. Ironically, now it's the Chinese government that has put a stop to your progress.
Jensen Huang: I’m not surprised by the setback. Of course, China wants to support its own technology stack. Given their speed in industry development, and with Huawei posting record-breaking results during our absence from that market, it’s no shock.
They also have several AI chip companies lined up for IPOs.
I think when considering American leadership, we need to adopt a longer-term perspective.
AI is not just about one model; that’s a huge misconception. As I’ve said, AI is like a 'five-layer cake.' We need to excel across all five layers—infrastructure, chips, platforms, models, and applications.
Some of the current approaches are self-destructive, undermining our chances of staying ahead in each layer.
The idea that success can be achieved by tightly bundling these five layers into a closed stack, forcing us to compete at the weakest link, is a grave mistake.
We need to let each layer compete and win on its own in the market.
Ben Thompson: Perhaps it’s because companies operating at other layers have been in Washington longer, whereas you chip and platform makers arrived later?
Jensen Huang: Heh, perhaps.
Ben Thompson: So what have you learned? What has been the most eye-opening thing about Washington for you?
Jensen Huang: What surprised me the most was how deeply the 'doomsayers' have penetrated, and their rhetoric has completely taken over the mindset of policymakers.
Ben Thompson: The result is that everyone is afraid, rather than excited.
Jensen Huang: Exactly. This creates two fatal problems.
In this industrial revolution, if we don't allow technology to fully diffuse within the United States and don't quickly utilize it, we will end up like Europe did in the last revolution — left behind by the times.
After all, many of the technologies from the last revolution were actually invented in Europe, but in the end, it was we who truly benefited. I hope we can learn from this historical wisdom and not fall into the trap set by those science fiction stories and 'doomsday theories' designed to scare laypeople.
What worries me most is the decline in AI's popularity as shown in public opinion polls — this is critical. It's no different from when people once opposed electricity, engines, and the internet.
If other countries adopt technology faster than we do, it will integrate more quickly into their societies. We must be wary of these sci-fi gimmicks that 'demonize' technology, which only serve to scare people and are utterly useless.
I don’t like those 'doomsayers' spreading panic. There is a fundamental difference between genuine risk warnings and rhetoric purely intended to create fear.
Ben Thompson: This phenomenon is quite typical. Smart people like to put on their deep-thinking hats and focus on minute details, but they forget that mass communication hinges on the overarching tone. You can't say 'I’m a little afraid of this but not that' while expecting the public not to panic. What you're conveying is fear, not optimism.
Jensen Huang: Yes, and doing so seems to make them appear smarter.
Ben Thompson: Who doesn't like to show off their intelligence?
Jensen Huang: Sometimes it’s to sound smart, sometimes it’s for fundraising, and sometimes it’s even about 'regulatory capture' to protect their turf. There are plenty of reasons.
These people are indeed incredibly intelligent, but I must remind them that these small maneuvers could very well backfire or even boomerang. One day, they will regret what they have done today.
07. The Essence of NVIDIA
Ben Thompson: Finally, let's talk about NVIDIA itself. NVIDIA is now the world's most valuable company by market capitalization. Does it feel real? Are you afraid? Or do you still feel, as you often said before, that the company is 'always 30 days away from bankruptcy'?
Jensen Huang: To be honest, it doesn’t feel real. When you’ve been with a company for 32 years, your perspective on it becomes very fixed.
In my eyes, we are still that small company from back then, and we continue to operate in that 'survival mode' to this day.
I remain constantly concerned about the company’s situation and deeply respectful of the astonishing pace of technological change.
Look, we’re still in full sprint mode. We don’t act like a giant corporation with deep moats and the luxury to rest on our laurels.
On the contrary, we are starting from scratch—rebuilding the network, redesigning the CPU, and reengineering the software stack. We are even willing to break the status quo through acquisitions to redefine the architecture for inference.
This in no way resembles a company that believes it has already 'won everything.'
Ben Thompson: At the end of your speech, you mentioned that you hope NVIDIA will become the 'engine that generates intelligence.' Does this mean that, in your ultimate vision, NVIDIA should not be seen as a hardware company, or even just a software company, but more like a 'utility company'?
Jensen Huang: I prefer to call it an 'enabling company.'
What we provide is the 'Promethean fire' of this era, or perhaps the electricity of this age. Our mission is to ensure that this form of energy is as affordable, widespread, and efficient as possible.
If this 'intelligence' can permeate every industry—from developing new drugs and designing safer cars to boosting the productivity of ordinary people tenfold—then we will have fulfilled our mission.
As for how others define us, I don’t care at all. As long as we continue solving problems that are unsolvable without accelerated computing, we are on the right path.
Editor/Rice