Jensen Huang's Podcast Transcript: NVIDIA's Moat Runs Much Deeper Than Chips

Video Title: Jensen Huang: – Will Nvidia's Moat Persist?

Video Author: Dwarkesh Patel

Compiled by: Peggy, BlockBeats

Editor's Note: While the outside world continues to debate whether "Nvidia's moat comes from its supply chain," this conversation posits that what is truly difficult to replicate is not the chips themselves, but the entire system capability of "transforming electrons into Tokens"—that is, the synergistic operation from computing architecture and software systems to the developer ecosystem.

This article is compiled from a conversation between Dwarkesh Patel and Jensen Huang. Dwarkesh Patel is one of the most-watched tech podcast hosts in Silicon Valley today, running the YouTube channel Dwarkesh Podcast, known for in-depth research-oriented interviews and long-term dialogues with AI researchers and key figures in the tech industry.

Dwarkesh Patel on the right, Jensen Huang on the left

Centered around this core idea, this conversation can be understood from three levels.

First, the changes in technology and industry structure.

Nvidia's advantage extends beyond hardware performance to the developer ecosystem underpinned by CUDA and the path dependency formed around the computing stack. In this system, computing power is no longer the sole variable; algorithms, systems engineering, networking, and energy efficiency collectively determine the pace of AI advancement. This leads to an important judgment: software will not be simply "commoditized" by AI. On the contrary, with the proliferation of Agents, tool invocation will grow exponentially, further amplifying the value of software.

Second, business boundaries and strategic choices.

Facing the ever-expanding AI industry chain, Nvidia chooses to "do everything necessary, but not everything." It does not enter cloud computing, nor does it engage in excessive vertical integration. Instead, it amplifies the overall market size through investment and ecosystem support. This restraint allows it to maintain critical control while avoiding becoming a substitute for the ecosystem, thereby integrating more participants into its technological system.

Third, the divergence regarding technology diffusion and industry structure.

The most tense part of the conversation lies not in specific conclusions, but in how to understand "risk" itself. One view emphasizes the first-mover advantage brought by computing power leadership, while another focuses more on the long-term ownership of ecosystems and standards during the process of technology diffusion. Compared to short-term capability gaps, perhaps the more critical question is: which technological system will future AI models and developers ultimately run on?

In other words, the endgame of this competition is not just "who builds a stronger model first," but "who defines the infrastructure on which models run."

In this sense, Nvidia's role is no longer just that of a chip company, but more akin to the "underlying operating system provider" for the AI era—it seeks to ensure that, regardless of how computing power diffuses, the path of value generation still revolves around itself.

The following is the original content (edited for readability):

TL;DR

· Nvidia's moat lies not in "chips," but in the "entire system capability from electrons to Tokens." The core is not hardware performance, but the full-stack ability to transform computation into value (architecture + software + ecosystem).

· The essential advantage of CUDA is not the tool, but the world's largest AI developer ecosystem. Developers, frameworks, and models are all bound to the same technology stack, creating a hard-to-replace path dependency.

· The key to AI competition is not just computing power, but the combination of "computing stack × algorithms × systems engineering." The improvements brought by the synergy of architecture, networking, energy efficiency, and software far exceed those from process node advancements alone.

· Computing power bottlenecks are a short-term issue; supply will catch up with demand signals within 2–3 years. The real long-term constraints are not chips, but energy and infrastructure.

· AI software will not be commoditized; instead, the explosion of Agents will lead to exponential growth in tool usage. The future is not about software becoming cheaper, but about a massive surge in software invocations.

· Not entering the cloud is Nvidia's core strategy: do "everything necessary," but don't swallow the entire value chain. Amplify the overall market size through investment and ecosystem support, not vertical integration.

· The real strategic risk is not competitors gaining computing power, but the global AI ecosystem no longer being based on the US technology stack. Once models and developers migrate, long-term technical standards and industry dominance will follow.

Interview Content

Where is Nvidia's Moat: In the Supply Chain, or in Control from "Electrons to Tokens"?

Dwarkesh Patel (Host):

We've seen valuations of many software companies decline because there's an expectation that AI will turn software into a standardized commodity. There's also a somewhat naive way of understanding it, roughly like this: Look, you hand the design files (GDS2) to TSMC, TSMC manufactures the logic chips, wafers, builds the switching circuits, then packages them together with HBM produced by SK Hynix, Micron, Samsung, and finally sends them to ODMs to assemble into complete server racks.

Note: HBM (High Bandwidth Memory) is an advanced memory technology specifically designed for high-performance computing and AI; ODM (Original Design Manufacturer) refers to contract manufacturers responsible not only for production but also for product design.

So, from this perspective, Nvidia is essentially doing software, while manufacturing is done by others. If software gets commoditized, then Nvidia would be commoditized too.

Jensen Huang (Nvidia CEO):

But ultimately, there has to be a process that turns electrons into tokens. This transformation from electrons to tokens, and making those tokens more valuable over time, I believe is very difficult to fully commoditize.

The transformation from electrons to tokens is itself an extraordinary process. And making one token more valuable, like making one molecule more valuable than another, is about making one token more valuable than another.

This process involves a great deal of art, engineering, science, and invention to imbue this token with value.

Clearly, we are observing this happening in real-time. So this transformation process, the manufacturing process, and the various signals involved are far from being fully understood, and this journey is far from over. So I don't think that scenario will happen.

Of course, we will make it more efficient. In fact, the way you just described the problem is actually my mental model of Nvidia: the input is electrons, the output is tokens, and the part in between is Nvidia.

Our job is to "do as much as necessary, and as little as unnecessary," to achieve this transformation and make it highly capable.

When I say "as little as possible," I mean for any part we don't need to do ourselves, we partner with others and bring it into our ecosystem. If you look at Nvidia today, we probably have one of the largest partner ecosystems, both upstream and downstream in the supply chain. From computer manufacturers, application developers, to model developers—you can think of AI as a "five-layer cake." And we have ecosystem presence across all five layers.

Related Reading: "Nvidia's Jensen Huang's Latest Article: AI's 'Five-Layer Cake'"

So we try to do as little as possible, but the part we must do is extremely difficult. And I don't think that part will be commoditized.

In fact, I also don't think enterprise software companies are essentially in the business of "tool making." But the reality is, most software companies today are indeed tool providers.

Of course, there are exceptions, some are encoding and solidifying workflow systems, but many companies are essentially tool companies.

For example, Excel is a tool, PowerPoint is a tool, Cadence makes tools, Synopsys is also a tool.

Jensen Huang:

And the trend I see is precisely the opposite of what many people think. I believe the number of agents will grow exponentially, and the number of tool users will also grow exponentially.

The number of invocation instances for various tools is also likely to surge. For example, the number of instances using Synopsys Design Compiler is likely to increase significantly.

There will be a large number of agents using floor planners, layout tools, design rule checking tools.

Today, we are limited by the number of engineers; tomorrow, these engineers will be supported by a large number of agents, and we will explore the design space in ways never before possible. When you use these tools today, this change will be very apparent.

The use of tools will drive explosive growth for these software companies. The reason this hasn't happened yet is because current agents aren't proficient enough at using tools.

So, either these companies build agents themselves, or agents themselves become capable enough to use these tools. I think ultimately it will be a combination of both.

Dwarkesh Patel

I recall in your latest disclosures, you have nearly $100 billion in purchase commitments for boundary components, memory, packaging, etc. And SemiAnalysis reports suggest this number could reach $250 billion.

One interpretation is that Nvidia's moat lies in you locking down the supply of these scarce components for many years to come. That is, others might also be able to make accelerators, but can they get enough memory? Enough logic chips?

Is this Nvidia's core advantage for the next few years?

Jensen Huang:

This is something we can do, but others find difficult. The reason we can make such massive commitments upstream is partly explicit, the purchase commitments you mentioned; and partly implicit.

For example, many upstream investments are actually made by our supply chain partners, because I would tell their CEOs: let me tell you how big this industry will be, let me explain why, let me walk through it with you, let me tell you what I see.

Through this process—conveying information, inspiring vision, building consensus—I align with CEOs across different upstream industries, and they are willing to make these investments.

Why are they willing to invest for me, and not for others? Because they know I have the ability to buy their capacity and absorb it through my downstream. It is precisely because Nvidia's downstream demand and supply chain scale are so large that they are willing to invest upstream.

Look at GTC, the scale of the conference shocks many people. It's essentially a 360-degree AI universe, bringing the entire industry together. People gather because they need to see each other. I bring them together, letting upstream see downstream, downstream see upstream, while letting everyone see the progress of AI.

More importantly, they can access AI-native companies and startups, see the various innovations happening, and thus personally verify the judgments I've been making.

So I spend a lot of time, directly or indirectly, explaining the opportunities ahead to our supply chain and ecosystem partners. Many people say my keynote doesn't announce products one after another like a traditional launch event; part of it sounds like a "lecture." And that is actually my intention.

I need to ensure the entire supply chain—both upstream and downstream—understands: what will happen next, why it will happen, when it will happen, how big the scale will be, and can reason about these issues systematically like I do.

So the "moat" you mentioned does exist. If this market reaches a trillion-dollar scale in the coming years, we have the ability to build the supply chain to support it. Like cash flow, supply chains also have flow and turnover. If an architecture's business turnover isn't fast enough, no one will build a supply chain for it. The reason we can sustain this scale is because downstream demand is extremely strong, and everyone can see that.

This is precisely what allows us to do things at the scale we do now.

Dwarkesh Patel

I still want to understand more concretely whether upstream can keep up. For many years, your revenue has basically doubled year over year, and the computing power supplied globally has even tripled.

Jensen Huang:

And it's doubling at this scale.

Dwarkesh Patel

Right. So if you look at logic chips, for example, you are one of TSMC's largest N3 customers, and also a major customer for N2.

According to some analysis, AI might account for 60% of N3 capacity this year, and could even reach 86% next year.

Note: N3 refers to TSMC's 3-nanometer (3nm) process node, which can be understood as one of TSMC's current-generation most advanced chip manufacturing processes.

Given you already occupy such a large share, how can you continue to double? And double every year? Have we entered a phase where AI computing power growth must slow down due to upstream constraints? Is there a way around these constraints? How exactly do we build twice as many fabs every year?

Jensen Huang:

At certain moments, instantaneous demand does exceed the entire industry's supply, both upstream and downstream. And in some cases, we've even been limited by the number of plumbers—that actually happened.

Dwarkesh Patel :

Then next year's GTC should invite plumbers.

Jensen Huang:

Yes, that's actually a good sign. You want to be in a market where instantaneous demand is greater than the total supply of the entire industry. The opposite, of course, is not good.

If the gap is too large, and a specific link, a specific component becomes a clear bottleneck, the entire industry will rush to solve it. For example, I notice people don't talk much about CoWoS anymore. The reason is, over the past two years, we've made massive investments and expansions in it, increasing it several times over.

Now I think the overall situation is in a good state. TSMC has also realized that CoWoS supply must keep pace with the demand growth for logic chips and memory. So they are expanding CoWoS, while also expanding future advanced packaging technologies, and at the same pace as logic chips.

This is very important because in the past, CoWoS and HBM memory were more like "special capabilities," but not anymore. Now everyone realizes they are part of mainstream computing technology.

At the same time, we are now more capable of influencing a broader range of the supply chain. In the past, at the beginning of the AI revolution, the judgments I'm talking about now were actually being made five years ago.

Some believed and invested back then, like Sanjay's team at Micron. I still remember that meeting; I explained very clearly what would happen, why, and predicted today's results. They chose to significantly ramp up, and we established a partnership with them. They invested in multiple directions like LPDDR, HBM, which obviously brought them great returns. Some companies followed later, but now everyone is at this stage.

So I believe every generation of technology, every bottleneck, will receive a lot of attention. And now, we are already "pre-fetching" these bottlenecks years in advance. For example, our collaborations with Lumentum, Coherent, and the entire silicon photonics ecosystem. Over the past few years, we have essentially reshaped the entire ecosystem and supply chain.

In silicon photonics, we built a complete supply chain around TSMC, collaborated with them on technology development, invented many new technologies, licensed these patents to the supply chain, keeping the ecosystem open. We prepared the supply chain by inventing new technologies, new workflows, new test equipment (including double-sided probing, etc.), investing in related companies, and helping them expand capacity.

So you can see we are actively shaping this ecosystem, enabling the supply chain to support future scale.

Dwarkesh Patel:

It sounds like some bottlenecks are easier to solve than others. For example, compared to those harder than expanding CoWoS.

Jensen Huang:

I actually gave the hardest example earlier.

Dwarkesh Patel:

Which one?

Jensen Huang:

Plumbers. Yes, really. I mentioned the hardest one earlier—plumbers and electricians. The reason is, this also makes me a bit concerned about the "doomsayers" who always talk about jobs disappearing, positions being replaced. If we discourage people from becoming software engineers because of this, we will truly lack software engineers in the future.

Similar predictions appeared a decade ago. Back then, people said: "Whatever you do, don't become a radiologist." You can still find those videos online, saying radiology would be the first profession to be eliminated, the world would no longer need radiologists. But the reality is, we now have a shortage of radiologists.

Dwarkesh Patel:

Okay, back to the earlier question: some links can be scaled, some cannot. Specifically, how do you double logic chip capacity? After all, the real bottleneck is here, both memory and logic are limiting factors. What about EUV lithography machines? How do you double their number every year?

Jensen Huang:

These are not impossible things. True, rapid expansion is not easy, but getting these things done within two to three years is actually not difficult. The key is having clear demand signals. Once you can build one, you can build ten; once you can build ten, you can build a million. So these things are essentially not difficult to replicate.

Dwarkesh Patel:

How deep do you communicate this judgment into the supply chain? For example, would you go to ASML and say: Looking three years ahead, for Nvidia