Regarding investment in OpenAI, AI bubble, and competition in ASIC... Jensen Huang answered it all

Author: Li Xiaoyin

Source: Hard AI

Recently, Jensen Huang, founder and CEO of NVIDIA, appeared on the "Bg2 Pod" bi-weekly dialogue program, engaging in an extensive conversation with hosts Brad Gerstner and Clark Tang.

During the discussion, Huang talked about the $100 billion collaboration with OpenAI and shared his views on the AI competition landscape and the prospects of sovereign AI.

Huang stated that the current AI competition is more intense than ever, the market has evolved from simple "GPU" to complex, continuously evolving "AI factories," which need to handle diverse workloads and exponentially growing inference tasks.

He predicted that if AI brings $10 trillion in added value to global GDP in the future, then the capital expenditure behind the AI factories would need to reach the level of $5 trillion annually.

Regarding the collaboration with OpenAI, Huang mentioned that OpenAI is likely to become the next trillion-dollar-scale mega company, and the only regret is not investing more earlier, "I should have given them all the money."

On the prospects of AI commercialization, Huang anticipated that within the next five years, AI-driven revenue will increase from $100 billion to the trillion-dollar level.

Concerning ASIC competition, NVIDIA stated that even if competitors set chip prices to zero, customers will still choose NVIDIA because their system operating costs are lower.

Here are the highlights from the discussion:

• OpenAI wants to establish a "direct relationship" with NVIDIA similar to that of Musk and X, including direct working relationships and direct procurement relationships.
• Assuming AI brings $10 trillion in added value to global GDP, and this $10 trillion in token generation has a 50% gross margin, then $5 trillion of it requires a factory, needs AI infrastructure, so the reasonable annual capital expenditure for this factory is about $5 trillion.
• 10 gigawatts would require an investment of about $400 billion, and this $400 billion largely needs to be funded through OpenAI's procurement agreements, which means their exponentially growing revenue. This must be funded through their capital, equity financing, and debt that can be raised.
• The probability of AI-driven revenue increasing from $100 billion to $1 trillion within the next five years is almost certain, and it is nearly already achieved.
• The global computing power shortage is not due to a shortage of GPUs, but because cloud service providers often underestimate future demand, leading NVIDIA to be in a "crisis production mode" for a long time.
• Huang stated that NVIDIA's only regret is that OpenAI invited us to invest early on, but we were too poor at the time and did not invest enough; I should have given them all the money.
• NVIDIA is likely to become the first company at the trillion-dollar level. Ten years ago, people said it would be impossible to have a trillion-dollar company. Now there are ten
• The current AI competition is more intense than ever, but it is also more difficult than ever. Jensen Huang stated that this is due to the increasing cost of wafers, which means that unless you engage in extreme-scale collaborative design, you cannot achieve an X-fold growth factor.
• The advantage that Google has is foresight. They launched TPU1 before everything started. When TPU becomes a big business, customer-owned tools will become the mainstream trend.
• The competitive advantage of NVIDIA chips lies in total cost of ownership (TCO). Jensen Huang stated that its competitors are building cheaper ASICs, and they can even set the price to zero. Our goal is that even if they set the chip price to zero, you will still buy NVIDIA systems because the total cost of operating that system is still more cost-effective than buying the chips (land, electricity, and infrastructure are already worth $15 billion).
• The performance of NVIDIA chips or tokens per watt is twice that of other chips, although the performance per unit of energy consumption is also much higher, customers can generate twice the revenue from their data centers. Who wouldn't want double the revenue?
• Every country must build sovereign AI. No one needs an atomic bomb, but everyone needs AI .
• Just as motors replaced labor and physical activities in the past, we now have AI. AI supercomputers and AI factories will generate tokens to enhance human intelligence, in the future, artificial intelligence will account for about 55-65% of global GDP, which is about $50 trillion.
• Artificial intelligence is not a zero-sum game. Jensen Huang stated, "The more ideas I have, the more problems I imagine we can solve, the more jobs we create, and the more employment opportunities we generate."
• One of the really cool things that will be solved in the next 5 years is the integration of artificial intelligence and robotics.
• Even without considering the new opportunities created by AI, the mere fact that AI changes the way we do things has immense value. It's like no longer using kerosene lamps but switching to electricity, or no longer using propeller planes but switching to jet planes.

The following is a summary of Jensen Huang's full conversation, assisted by AI tools for translation, with some edits from Wall Street Insight:

Host:

Jensen, it's great to be back again, and of course, my partner Clark Tang is here. I can hardly believe it—welcome to NVIDIA.

Jensen Huang:

Oh, and nice glasses.

Host:

They really look good on you. The question is, now everyone will want you to wear them all the time. They will say, "Where are the red glasses?" I can attest to that.

Jensen Huang:

It has been over a year since we last recorded the show. Now over 40% of revenue comes from inference, but inference is about to change due to the emergence of the inference chain. It is about to grow a billion times, right? A million times, a billion times

Host:

That's right. This is exactly the part that most people have not fully internalized yet. This is the industry we've been talking about. This is the industrial revolution. To be honest, it feels like you and I have been continuing this show every day since then. In AI time, it's been about a hundred years.

I recently rewatched the show, and many of the things we discussed stood out. The most profound point for me was when you slammed the table to emphasize—remember back then when pre-training was in some sort of low tide, and people were saying, "Oh my gosh, pre-training is over, right? Pre-training is over. We're not going to continue. We overbuilt." That was about a year and a half ago. You said reasoning wouldn't grow by 100 times, 1000 times, but rather by a billion times.

This brings us to the current situation. You announced this huge deal. We should start from there.

Jensen Huang:

I underestimated it. Let me officially state that. I estimate we now have three scaling laws, right? We have pre-training scaling laws. We have post-training scaling laws. Post-training is basically like AI practice.

Yes, practicing a skill until you master it. So it tries various different methods, and to do that, you have to reason. So now training and reasoning are integrated in reinforcement learning. Very complex. This is what we call post-training. Then the third is reasoning. The old way of reasoning was one-off, right? But the new way of reasoning we recognize is thinking. So think before answering.

Now you have three scaling laws. The longer you think, the better the quality of the answer you get. During the thinking process, you do research, check some basic facts. You learn something, think some more, learn some more, and then generate an answer. Don't just generate it directly at the beginning. So thinking, post-training, pre-training, we now have three scaling laws instead of one.

Host:

You knew this last year, but how confident are you this year that reasoning will grow by a billion times and what level of intelligence this will bring? Are you more confident this year than last year?

Jensen Huang:

I am more confident this year, the reason being that now looking at agent systems, AI is no longer just a language model; AI is a language model system, and they are all running concurrently, possibly using tools. Some of us use tools, some do research.

Host:

There has been a lot of development in the industry right now, all about multimodal, look at all the video content being generated, it's truly amazing technology. This indeed brings us to the key moment everyone has been discussing this week, which is your announcement a few days ago about the massive collaboration with OpenAI Stargate, where you will become a priority partner and invest $100 billion in the company for a period of time. They will build 10 gigawatt facilities, and if they use NVIDIA's products in these 10 gigawatt facilities, it could bring NVIDIA up to $400 billion in revenue. Please help us understand this partnership, what it means for you, and why this investment makes so much sense for NVIDIA

Jensen Huang:

First, I will answer the last question, and then elaborate. I believe OpenAI is very likely to become the next trillion-dollar-scale super-large company.

Host:

Why do you refer to it as a super-large company? Super-large companies are like Meta and Google; they will have consumer and enterprise services, and they are likely to become the next trillion-dollar-scale super-large company in the world. I think you would agree with this view.

Jensen Huang:

I agree. If that is the case, the investment opportunity before they reach that goal is one of the smartest investments we can imagine. You must invest in areas you understand, and we happen to understand this area. Therefore, the return on this investment opportunity will be fantastic. So we really like this investment opportunity. We don't have to invest; it's not a requirement for our investment, but they have given us the opportunity to invest, which is a great thing.

Now let me start from the beginning. We are collaborating with OpenAI on several projects. The first project is the construction of Microsoft Azure. We will continue to do this, and the collaboration is progressing very well. We have several years of construction work to do, with hundreds of billions of dollars of work just there.

The second is the construction of OCI, which I believe is about five, six, or seven thousand megawatts about to be built. So we are working with OCI, OpenAI, and SoftBank on this construction, and these projects are all signed, with a lot of work to do. The third is Core Weave, all of which are still in the context of OpenAI.

So the question is, what is this new partnership? This new partnership is about helping OpenAI, working with OpenAI to build their own AI infrastructure for the first time. This is our direct collaboration with OpenAI at the chip level, software level, system level, and AI factory level, helping them become a fully operational super-large company. This will take some time, and it will complement the two exponential growths they are experiencing. The first exponential growth is the exponential growth in the number of customers, because AI is getting better, use cases are improving, and almost every application is now connected to OpenAI, so they are experiencing exponential growth in usage. The second exponential growth is the exponential growth in computing per use. Now it’s not just a single inference, but thinking before answering.

Host:

These two exponential growths compound their computing demands. So we have to build all these different projects. The last project is a supplement to everything they have already announced, a supplement to everything we have already collaborated on, and it will support this incredible exponential growth.

One interesting point you made is that, in your view, they are likely to become trillion-dollar companies, which I think is a great investment. At the same time, they are building on their own, and you are helping them build their own data centers. So far, they have been outsourcing to Microsoft to build data centers, and now they want to build a full-stack factory themselves.

Jensen Huang:

They want to... they want to... they basically want to establish a relationship with us like Elon and X.

Host:

I think this is a very important thing. When you consider the advantages that Colossus has, they are building full-stack. This is hyperscale because if they don't use that capacity, they can sell it to others. Similarly, Stargate is building huge capacity. They believe they will use most of that capacity, but it also allows them to sell it to others. This sounds a lot like AWS, GCP, or Azure.

Jensen Huang:

They want to establish the same direct relationship with us, including direct working relationships and direct procurement relationships. Just like Zuckerberg and Meta's relationship with us, our direct collaboration with Sundar and Google, and our direct collaboration with Satya and Azure. They have reached a large enough scale and believe it's time to start building these direct relationships. I'm happy to support this, Satya knows about it, Larry knows about it, everyone is aware of the situation, and everyone is very supportive.

Host:

Regarding the market prospects for Nvidia's accelerated computing, I found an interesting phenomenon. Oracle is building the $300 billion Colossus project, we know what governments are building, we know what hyperscale cloud service providers are building, and Sam is talking about trillion-dollar investments. But the 25 sell-side analysts covering our stock on Wall Street basically believe that our growth will flatten out starting in 2027, with only 8% growth from 2027 to 2030. The only job of these 25 people is to forecast Nvidia's growth rate.

Jensen Huang:

Frankly, we are quite calm about this. We regularly exceed market expectations without any issues. But there is indeed this interesting divergence. I hear these views every day on CNBC and Bloomberg. I think it involves some skepticism about shortages leading to oversupply; they don't believe in sustained growth. They say, "Well, we acknowledge your performance in 2026, but in 2027, there might be oversupply, and you won't need that much."

Host:

Interestingly, the consensus expectation is that this growth will not happen. We have also made company forecasts considering all this data. This shows me that even though we have been in the AI era for two and a half years, there is a huge divergence between what Sam Altman, I, Sundar, and Satya are saying and what Wall Street still believes

Jensen Huang:

I believe this is not contradictory. Let me explain with three points, hoping to help everyone have more confidence in Nvidia's future.

The first point is from the perspective of physical laws, which is the most important point: the era of general computing is over, and the future is accelerated computing and AI computing. What needs to be considered is how many trillions of dollars worth of computing infrastructure around the world needs to be updated. When they are updated, they will turn to accelerated computing. Everyone agrees on this, saying "yes, we completely agree, the era of general computing is over, and Moore's Law is dead." This means that general computing will shift to accelerated computing. Our collaboration with Intel recognizes that general computing needs to integrate with accelerated computing to create opportunities for them.

The second point is that the first application scenario of AI is actually ubiquitous, in areas like search recommendation engines and shopping. The foundational hyperscale computing infrastructure used to make recommendations with CPUs, and is now shifting to using GPUs for AI. This represents the transition from classic computing to accelerated computing AI. Hyperscale computing is moving from CPUs to accelerated computing and AI, which involves providing services to companies like Meta, Google, ByteDance, and Amazon, transforming their traditional hyperscale computing methods to AI, representing a market worth hundreds of billions of dollars.

Host:

Considering platforms like Douyin, Meta, and Google, there may be a demand workload driven by accelerated computing for 4 billion people globally.

Jensen Huang:

So even without considering the new opportunities created by AI, the mere fact that AI changes the way we do things has immense value. It's like moving from kerosene lamps to electricity, or from propeller planes to jet planes.

So far, I have been discussing these fundamental transformations. And when you turn to AI and accelerated computing, what new applications will emerge? This is all the AI-related opportunities we are discussing.

Simply put, just as motors replaced labor and physical activities in the past, we now have AI. The AI supercomputers and AI factories I mentioned will generate tokens to enhance human intelligence. Artificial intelligence accounts for about 55-65% of global GDP, which is about $50 trillion. This $50 trillion will be enhanced.

Let's start with an individual. Suppose I hire an employee with a salary of $100,000, and then equip him with $10,000 worth of AI. If this $10,000 AI can double or triple the productivity of this $100,000 employee, would I do it? Without hesitation. I am doing this with everyone in the company now; every software engineer, every chip designer has AI collaborating with them, 100% coverage. The result is that the quality of the chips we manufacture is better, the quantity is increasing, and the speed is improving. Our company is growing faster, hiring more people, with higher productivity, higher revenue, and stronger profitability.

Now apply Nvidia's story to global GDP. What is likely to happen is that the $50 trillion will be enhanced by $10 trillion. This $10 trillion needs to run on machines. **The difference with AI compared to the past is that software was pre-written and then run on CPUs, operated by humans But in the future, for AI to generate tokens, machines must generate tokens and think, so the software must keep running. In the past, software only needed to be written once, but now software is essentially being written continuously and is always thinking. To enable AI to think, it needs a factory.

Assuming that the generation of 10 trillion tokens has a gross margin of 50%, of which 5 trillion requires factories and AI infrastructure. If you tell me that global annual capital expenditure is about 5 trillion dollars, I would say this mathematical calculation is reasonable. This is the future: shifting from general computing in Excel to accelerated computing, replacing all hyperscale servers with AI, and then enhancing human intelligence in global GDP.

Host:

Currently, we estimate the annual revenue of this market to be about 400 billion dollars. So the TAM is expected to grow 4 to 5 times from now.

Jensen Huang:

Last night, Alibaba's Eddie Wu said that from now until the end of this decade, they will increase the power of data centers by ten times. NVIDIA's revenue is almost related to power. He also mentioned that token generation doubles every few months. This means that performance per watt must continue to grow exponentially. This is why NVIDIA keeps breaking through in performance per watt, because in this future, revenue per watt is essentially revenue.

This assumption contains a very interesting historical context. For the past 20 years, GDP has basically not grown. Then the Industrial Revolution came, and GDP accelerated growth. The Digital Revolution came, and GDP accelerated growth again. What you are saying, as Scott Besson mentioned, is that he believes we will have 4% GDP growth next year. You mean that global GDP growth will accelerate because now we are providing billions of colleagues working for us. If GDP is the output of fixed labor and capital, it must accelerate growth.

Look at the changes brought by AI technology. AI technology, including large language models and all AI agents, is creating a new AI agent industry. OpenAI is the fastest-growing company in history, experiencing exponential growth. AI itself is a rapidly growing industry because AI requires factories and infrastructure behind it. This industry is growing, my industry is growing, and the industries beneath my industry are also growing. Energy is growing, which is a revival for the energy industry, and all companies in the infrastructure ecosystem, such as nuclear energy and gas turbines, are performing well.

These numbers have everyone talking about excess and bubbles. Zuckerberg said last week on a podcast that he believes there will likely be a vacuum period at some point, and Meta may overspend by 10 billion dollars, but that doesn't matter because it is so important for the future of their business, and it is a risk they must take. This sounds a bit like a prisoner's dilemma, but these are very happy prisoners.

Host:

We estimate that there will be 100 billion dollars in AI revenue by 2026, excluding Meta, excluding GPUs running recommendation engines, and excluding search and other functions. The hyperscale server industry is already in the trillions, and this industry will shift to AI. Skeptics will say we need to grow from 100 billion dollars in AI revenue in 2026 to at least 1 trillion dollars in AI revenue by 2030 You just mentioned a $5 trillion global GDP. If we analyze from the bottom up, can you see AI-driven revenue growing from $100 billion to $1 trillion in the next five years? Are we growing that fast?

Jensen Huang:

Yes, I would say we have already reached that. Because hyperscale service providers have shifted from CPUs to AI, their entire revenue base is now AI-driven.

Without artificial intelligence, TikTok cannot operate, and YouTube shorts cannot be made; these things cannot happen without AI. The amazing work that Meta has done in customizing and personalizing content cannot be separated from artificial intelligence. Previously, this work was done by humans, that is, creating four options in advance and then selecting through a recommendation engine. Now it has turned into AI generating an infinite number of choices.

But these things are already happening; we have transitioned from CPUs to GPUs, primarily for these recommendation engines, which is quite new in the past three to four years. When I met with Zuckerberg at Siggraph, he would tell you that they indeed started using GPUs relatively late. Meta has been using GPUs for about two and a half years, which is quite new. Searching on GPUs is absolutely brand new.

Host:

So your argument is that by 2030, the probability of having a trillion-dollar AI revenue is almost certain because we have almost reached that point. Let's talk about the incremental part from now on. When you do bottom-up or top-down analysis, I just heard your top-down analysis regarding the percentage of global GDP. What do you think is the probability of encountering oversupply in the next three, four, or five years?

Jensen Huang:

It's a distribution; we don't know the future. I think the likelihood of that is very low until we fully transition all general computing to accelerated computing and AI. That will take a few years.

Until all recommendation engines are based on AI, until all content generation is based on AI, because consumer-facing content generation largely consists of recommendation systems and so on, all of this will be generated by AI. Until all classic hyperscale businesses shift to AI, from shopping to e-commerce and everything else, until everything transitions over.

Host:

But all these new constructions, when we talk about a trillion dollars, we are investing for the future. Even if you see a slowdown or some kind of oversupply coming, are you obligated to invest that money? Or is it that you are waving a flag to the ecosystem, telling them to build, and if we see a slowdown, we can always reduce the level of investment?

Jensen Huang:

In fact, the opposite is true, because we are at the end of the supply chain, responding to demand. Now all the venture capitalists will tell you that demand exceeds supply, there is a computing shortage in the world, not because there is a shortage of GPUs in the world. If they give me orders, I will produce. In the past few years, we have really streamlined the supply chain, from wafers to packaging, HBM memory, and all technologies; we are ready If we need to double, we will double. So the supply chain is already prepared.

Right now, we are just waiting for demand signals. When cloud service providers, hyperscalers, and our customers make their annual plans and give us forecasts, we will respond and produce accordingly. Of course, what is happening now is that every forecast they provide us is wrong because they are underestimating, so we are always in emergency mode. We have been in emergency mode for years, and whatever forecast we receive shows significant growth compared to last year, but it is still not enough.

Host:

Satya seemed to retreat last year, and some called him the adult in the room, tempering those expectations. A few weeks ago, he said we built 2 exabytes this year, and we will accelerate in the future. Do you see some traditional hyperscalers who might be moving slower than CoreWeave or Elon X, perhaps slower than Stargate? It sounds like they are all investing more aggressively now.

Jensen Huang:

Because of the second exponential growth.

We have experienced an exponential growth, which is the adoption and engagement of AI growing exponentially. The second exponential growth is reasoning capability. This was our conversation a year ago. A year ago, we said that when you shift AI from one-off, memory answers to memory and generalization (which is basically pre-training), like remembering that 8 times 8 equals what, that is one-off AI. A year ago, reasoning emerged, tool usage emerged, and now you have thinking AI, with a computational load that is a billion times greater.

Host:

Some hyperscale customers have internal workloads that they must migrate from general computing to accelerated computing anyway, so they are continuously building. I think perhaps some hyperscalers have different workloads, so they are uncertain how quickly they can digest, but now everyone concludes that they are severely underbuilding.

Jensen Huang:

One of my favorite applications is traditional data processing, including structured and unstructured data. We will soon announce a very large accelerated data processing program. Data processing represents the vast majority of CPU usage in today's world. It still runs entirely on CPUs. If you go to Databricks, it is mainly CPU; go to Snowflake, it is mainly CPU; Oracle's SQL processing is mainly CPU. Everyone is using CPUs for SQL structured data processing. In the future, all of this will shift to AI data processing.

This is a huge market we are going to enter. But you need everything NVIDIA is doing, you need the acceleration layer and specific domain data processing recipes, and we have to go build these.

Host:

When I opened CNBC yesterday, they were talking about the oversupply bubble; when I opened Bloomberg, they were discussing cyclical revenue issues.

Jensen Huang:

When someone questions our investment and business relationships with companies like OpenAI, I need to clarify a few points. First, recurring revenue arrangements typically refer to companies entering misleading transactions, artificially inflating revenues without any potential economic substance. In other words, this is growth driven by financial engineering rather than customer demand. A classic example that everyone cites is Cisco and Nortel from the last bubble 25 years ago.

When we or Microsoft and Amazon invest in companies that are also our major clients, such as our investment in OpenAI while OpenAI purchases tens of billions of chips, I want to remind everyone that the analysts on platforms like Bloomberg who are overly concerned about recurring revenue or round-trip transactions are mistaken.

10 gigawatts requires about $400 billion in investment, and this $400 billion largely needs to be funded through their purchase agreements, which is their exponential growth revenue. This must be funded through their capital, equity financing, and the debt they can raise. These are three financing tools. The equity and debt they can raise are related to their confidence in maintaining revenue. Smart investors and lenders will consider all these factors. Fundamentally, this is what they need to do; this is their business, not mine. Of course, we must stay closely connected with them to ensure our construction supports their continued growth.

There is no relationship between revenue and investment. The investment aspect is not tied to anything; it is an opportunity to invest in them. As we mentioned earlier, this is likely to become the next multi-trillion-dollar mega company. Who wouldn’t want to be an investor in that? My only regret is that they invited us to invest early on. I remember those conversations; we were too poor at the time and didn’t invest enough. I should have given them all my money.

The reality is that if we don’t do our job well and keep up, if Rubin doesn’t become a good chip, they can procure other chips for these data centers. They are not obligated to use our chips. As mentioned earlier, we see this as an opportunistic equity investment.

We have also made some great investments, and I must mention them; we invested in XAI and Corewave, both of which are fantastic and very wise.

Back to the fundamental issue, we are openly and transparently stating what we are doing. There is potential economic substance here; we are not simply sending revenue back and forth between two companies. People are paying for ChatGPT every month, with 1.5 billion monthly active users using the product. Every enterprise either adopts this technology or will die. Every sovereign nation sees this as a survival threat to its national security and economic security, just like nuclear energy. Which person, company, or country would say that intelligence is essentially optional for us? This is fundamental for them. The automation of intelligence, I have fully discussed the demand issue.

Host:

Speaking of system design, in 2024 we will shift to an annual release cycle, starting with Hopper. Then we will undergo a massive upgrade requiring significant data center renovations, launching Grace Blackwell. In the second half of 2025 and 2026, we will launch Vera Rubin The Ultra will be launched in 2027, and Fineman will be launched in 2028.

How is the transition to an annual release cycle progressing? What are the main goals? Has AI helped us execute the annual release cycle? The answer is yes. Without AI, NVIDIA's speed, rhythm, and scale would be limited.

Jensen Huang:

The answer to the last question is yes. It is simply impossible to build the products we have built today without AI.

Why do this? Remember what Eddie said in the earnings call, what Satya said, and what Sam also said, the token generation rate is growing exponentially, and customer usage is also growing exponentially. I believe they have about 800 million weekly active users. This is less than two years since the release of ChatGPT. Each user is generating more tokens as they use reasoning time.

First, because the token generation rate is growing at an incredible speed, with two exponentials stacking together, unless we improve performance at an incredible rate, the cost of token generation will continue to rise, as Moore's Law is dead, and the cost of transistors remains basically the same each year, as does power consumption. Between these two fundamental laws, unless we propose new technologies to reduce costs, even with slight differences in growth, how do we compensate for two exponential growths by giving someone a few percentage points of discount?

Therefore, we must improve performance every year to keep up with exponential growth. From Kepler to Hopper, it could be a 100,000 times growth, which is the beginning of NVIDIA's AI journey, 100,000 times in 10 years. Between Hopper and Blackwell, due to NVLink 72, we achieved a 30 times growth in one year, and then Rubin will again achieve an X-fold factor, and Fineman will also get another X-fold factor.

We do this because transistors are not much help to us; Moore's Law is basically density is increasing but performance is not. If that is the case, one of the challenges we face is that we must decompose the entire problem at the system level while changing every chip and all software stacks and all systems. This is the ultimate extreme co-design; no one has done co-design at this level before. We are innovating CPU, GPU, network chips, NVLink expansion, and Spectrum X horizontal expansion.

Some say, "Oh yes, it's just Ethernet." Spectrum X Ethernet is not just Ethernet. People are starting to realize, oh my gosh, the X-fold factor is quite incredible. NVIDIA's Ethernet business, just the Ethernet business, is the fastest-growing Ethernet business in the world.

We now need to scale up, and of course, build larger systems. We are scaling across multiple AI factories, connecting them together. We are doing this work on an annual cycle. Therefore, we are now achieving exponential exponential growth in technology. This enables our customers to reduce token costs, making these tokens smarter through pre-training, post-training, and reasoning. The result is that as AI becomes smarter, their usage will increase. When usage increases, it will grow exponentially. **

For those who may not be familiar, what is extreme collaborative design? Extreme collaborative design means you have to optimize models, algorithms, systems, and chips simultaneously. You have to innovate outside the box. Because Moore's Law states that you only need to make the CPU faster, and everything will speed up. You are innovating inside the box, just making the chip faster. But what if that chip can’t get any faster? You innovate outside the box. NVIDIA really changed the game because we did two things. We invented CUDA, we invented the GPU, and we invented the concept of large-scale collaborative design. That’s why we are involved in all these industries. We create all these libraries and collaborative designs.

First, the full-stack extreme even goes beyond software and GPUs. Now it’s about switches and networks at the data center level, optimizing software, network interface cards, expansion, and horizontal scaling in all these switches and networks. Therefore, Blackwell has a 30 times improvement over Hopper, which no Moore's Law can achieve. This is the extreme, and it comes from extreme collaborative design. This is why NVIDIA is entering networking and switching, expansion and horizontal scaling, cross-system expansion, building CPUs and GPUs, and network interface cards.

This is why NVIDIA is so rich in software and talent. The amount of open-source software we contribute to the world exceeds almost any company, except for one, I think it’s something like AI2. We have such a vast software resource, and that’s just in AI. Don’t forget about computer graphics, digital biology, autonomous vehicles; the amount of software our company produces is incredible, enabling us to perform deep and extreme collaborative design.

Host:

I heard from one of your competitors that, yes, they do this because it helps reduce token generation costs, but at the same time, your annual release cycle makes it almost impossible for your competitors to keep up. The supply chain is locked in more because you provide three years of visibility to the supply chain. Now the supply chain has confidence in what they can build. Have you considered this?

Jensen Huang:

Before you ask this question, think about this. To enable us to build hundreds of billions of dollars in AI infrastructure each year, think about how much capacity we had to start preparing a year ago. We are talking about building hundreds of billions of dollars in wafer starts and DRAM procurement.

This has now reached a scale that almost no company can keep up with.

Host:

So do you think your competitive moat is larger than it was three years ago?

Jensen Huang:

Yes. First, the competition is more intense than ever, but it is also more difficult than ever. The reason I say this is that wafer costs are rising, which means unless you engage in extreme-scale collaborative design, you cannot achieve an X-fold growth factor. Unless you are developing six, seven, or eight chips every year, that’s remarkable. It’s not about building an ASIC; it’s about building an AI factory system. This system has many chips, all of which are collaboratively designed, providing us with the 10x factor we almost regularly achieve

First, collaborative design is extreme. Second, scale is extreme. When your customer deploys one terawatt-hour, that's 400,000 to 500,000 GPUs. Getting 500,000 GPUs to work together is a miracle. Your customer is taking on enormous risk by purchasing all of this from you. You have to ask yourself, what customer would place a $50 billion purchase order on a single architecture? On an unproven architecture, a new architecture. A brand new chip that you are excited about, and everyone is excited for you, and you just showcased the first silicon. Who would give you a $50 billion purchase order? Why would you launch a $50 billion wafer for a chip that has just been taped out?

But for NVIDIA, we can do this because our architecture has been validated. The scale of our customers is incredibly impressive. Now the scale of our supply chain is also incredibly impressive. Who would launch all of this for a company, pre-build all of this, unless they know NVIDIA can deliver? They believe we can deliver to customers around the world. They are willing to launch hundreds of billions at a time. This scale is unbelievable.

Host:

One of the biggest key debates and controversies in the world regarding this is the issue of GPUs versus ASICs, with Google's TPU, Amazon's Trainium, and rumors from ARM to OpenAI to Anthropic seemingly building ASICs. Last year you said we are building systems, not chips, and you drive performance through every part of the stack. You also mentioned that many of these projects may never reach production scale. But considering most projects, and the apparent success of Google's TPU, how do you view the evolving landscape today?

Jensen Huang:

First, the advantage that Google has is foresight. Remember, they launched TPU1 before everything started. This is no different from a startup. You should create a startup before the market grows. When the market reaches a trillion-dollar scale, you shouldn't emerge as a startup. This is a fallacy that all venture capitalists know, that if you can capture a few percentage points of market share in a large market, you can become a giant company. This is fundamentally wrong.

You should capture 100% market share in a small industry, just like Nvidia and TPU did. At that time, there were only the two of us, but you have to hope that this industry can really grow. You are creating an industry.

The story of Nvidia illustrates this. For companies building ASICs now, this is a challenge. While the market may seem enticing now, remember that this enticing market has evolved from a chip called GPU into the AI factory I just described. You just saw me announce a chip called CPX for context processing and diffusion video generation, which is a very specialized workload but an important workload within data centers. I just mentioned the potential of AI data processing processors because you need long-term memory and short-term memory. KV cache processing is very intensive. AI memory is a big issue You hope your AI has good memory, and simply handling the entire system's KV cache is a very complex task that may require specialized processors.

You can see that Nvidia's perspective is no longer just about GPUs. Our view is to examine the entire AI infrastructure and what these excellent companies need to handle their diverse and ever-changing workloads.

Look at transformers. The transformer architecture is undergoing significant changes. If it weren't for CUDA being easy to operate and iterate, how would they attempt a large number of experiments to decide which transformer version to use and what kind of attention algorithms? How to break it down? CUDA helps you accomplish all of this because it is highly programmable.

Now think about the way we do business: when all these ASIC companies or ASIC projects started three, four, or five years ago, I must tell you that the industry was very simple. It involved a GPU. But now it has become large and complex, and in two years it will become completely enormous. The scale will be very huge. So I think, it is difficult to enter a very large market as a newcomer. This is true even for those customers who may succeed with ASICs.

Host:

Investors tend to be binary creatures; they only want black-and-white answers of yes or no. But even if you get ASICs to work, isn't there an optimal balance? Because I think when purchasing the Nvidia platform, CPX will launch for pre-filling, video generation, and possibly decoding, etc.

Jensen Huang:

Yes, so there will be many different chips or components added to Nvidia's ecosystem of accelerated computing clusters as new workloads emerge. The people trying to tape out new chips now are not really predicting what will happen a year from now; they are just trying to get the chips to work.

In other words, Google is a major customer of GPUs. Google is a very special case, and we must show the respect it deserves. TPU has reached TPU7. This is also a challenge for them, and the work they are doing is very difficult.

Let me clarify, there are three categories of chips. The first category is architectural chips: X86 CPUs, ARM CPUs, Nvidia GPUs. They are architectural, with ecosystems built on top, and the architecture has rich IP and ecosystems, with technology that is very complex, built by owners like us.

The second category is ASICs. I once worked for the original company that invented the ASIC concept, LSI Logic. As you know, LSI Logic no longer exists. The reason is that ASICs are indeed great when the market size is not very large; it is easy for someone to help you package everything and manufacture on your behalf, charging a 50-60% gross margin. But when the ASIC market grows larger, a new practice called customer-owned tooling emerges. Who does this? Apple's smartphone chips. Apple's smartphone chip volume is very large, and they would never pay someone else a 50-60% gross margin to do ASICs. They use customer-owned tooling

So where will TPU go when it becomes a big business? Customers have their own tools, no doubt about it.

But ASICs have their place. Video transcoders will never be too big, and smart network cards will never be too big. When an ASIC company has 10, 12, or 15 ASIC projects, I’m not surprised, because there might be five smart network cards and four transcoders. Are they all AI chips? Certainly not. If someone builds an embedded processor as an ASIC for a specific recommendation system, that can certainly be done. But would you consider it as the foundational computing engine for AI, which is constantly evolving? You have low-latency workloads, high-throughput workloads, token generation for chat, reasoning workloads, and AI video generation workloads.

Host:

You are talking about the backbone of accelerated computing.

Jensen Huang:

That’s the whole point of Nvidia.

Host:

Simply put, it’s like the difference between playing chess and checkers. The fact is, companies starting to make ASICs today, whether it’s Tranium or some other accelerators, they are building a chip that is just a component of a larger machine.

You’ve built a very complex system, platform, factory, and now you’ve opened up a bit in some way. You mentioned CPX GPUs, and in some ways, you are breaking down workloads into the best hardware segments for that specific domain.

Jensen Huang:

We announced something called Dynamo, which orchestrates decomposed AI workloads, and we open-sourced it because the future AI factory is decomposed.

Host:

You launched MV Fusion and even told your competitors, including Intel, that you are involved in the way we are building this factory because no one else is crazy enough to try to build the whole factory, but if you have a good enough and compelling product that makes end users say, "We want to use this instead of ARM GPUs, or we want to use this instead of your inference accelerators," you can tap into it.

Jensen Huang:

We are excited to make connections. Fusion is indeed a fantastic idea, and we are pleased to collaborate with Intel on this. It requires leveraging Intel’s ecosystem, as most enterprises in the world still run on Intel platforms. We are merging Intel’s ecosystem with Nvidia’s AI ecosystem and accelerated computing. We are doing the same with ARM and several other companies we will collaborate with. This opens up opportunities for both sides and is a win-win for both. I will become their major customer, and they will give us access to larger market opportunities.

This closely relates to a point you raised that might shock some people. You said our competitors are building ASICs, and all their chips are already cheaper today, but they could even price them at zero. Our goal is that even if they price their chips at zero, you will still buy Nvidia systems because the total cost of operating that system—power, data centers, land, etc.—the intelligence output is still more cost-effective than buying the chips, even if the chips are free Because land, electricity, and infrastructure are already valued at $15 billion. We have analyzed this mathematical problem.

Host:

But please explain your calculations to us, because I think it is indeed difficult for those who do not often think about this issue to understand. Your chips are so expensive, how can it be that your chips are still a better choice when your competitors' chip prices are zero?

Jensen Huang:

There are two ways to think about it. One is from the revenue perspective. Everyone is constrained by power consumption, and let's assume you can get an additional 2 gigawatts of power. Then you want to convert that 2 gigawatts of power into revenue. If your performance or tokens per watt is twice that of others because you have done deep and extreme code design, my performance per unit of energy consumption is much higher, so my customers can generate twice the revenue from their data centers. Who wouldn't want double the revenue?

If someone offers them a 15% discount, the difference between our gross margin (about 75 percentage points) and others' gross margin (about 50 to 65 percentage points) is not enough to compensate for the 30 times performance difference between Blackwell and Hopper. Assume Hopper is an excellent chip and system, and assume others' ASICs are just like Hopper. Blackwell's performance is 30 times that.

So in that 1 gigawatt, you have to give up 30 times the revenue. The cost is too high. Therefore, even if they provide the chips for free, you only have 2 gigawatts to use, and the opportunity cost is extremely high. You will always choose the best performance per watt.

Host:

I heard from the CFO of a hyperscale cloud service provider that given the performance improvements brought by your chips, specifically regarding tokens per gigawatt, and with electricity becoming a limiting factor, they must upgrade to a new cycle. When you look ahead to Ruben, Ruben Ultra, and Fineman, will this trend continue?

Jensen Huang:

We are now building six to seven chips each year, each of which is part of a system. System software is everywhere and needs to be integrated and optimized across all these six to seven chips to achieve the 30 times performance improvement of Blackwell. Now imagine I am doing this every year, continuously. If you build an ASIC in this series of chips while we optimize across the entire system, it becomes a difficult problem to solve.

Host:

This brings me back to the question about competitive moats that I raised at the beginning. We have been focusing on this, and we are investors in the ecosystem, also investing in your competitors, from Google to Broadcom. But when I think about this issue from first principles, are you increasing or decreasing your competitive moat?

You are shifting to an annual rhythm, co-developing with the supply chain. The scale is much larger than anyone expected, which requires the scale of the balance sheet and development. The initiatives you are taking through acquisitions and organic growth, including Envy Fusion, CPX, and others we just discussed All these factors lead me to believe that, at least in terms of building factories or systems, your competitive moat is strengthening. This is at least surprising. But interestingly, your price-to-earnings ratio is much lower than that of most other companies. I think this is partly related to the law of large numbers. A $4.5 trillion company cannot grow much larger. But I asked you this question a year and a half ago, and today you sit here, if the market is expected to grow AI workloads by 10 times or 5 times, we know what capital expenditures are doing, etc. In your view, is there any conceivable scenario where your revenue in five years will not be 2 or 3 times higher than in 2025? Considering these advantages, what is the probability that revenue will not be much higher than today?

Jensen Huang:

I will answer this way. As I described, our opportunities are much greater than the consensus. I am here to say, I believe NVIDIA is likely to become the first $10 trillion company. I have been here long enough, just ten years ago, you should remember, people said there could never be a trillion-dollar company. Now we have ten.

Host:

The world is bigger today, going back to GDP and exponential growth.

Jensen Huang:

The world is bigger, and people misunderstand what we do. They remember us as a chip company, and we do make chips; we make the best chips in the world.

But NVIDIA is actually an AI infrastructure company. We are your AI infrastructure partner, and our collaboration with OpenAI is a perfect example. We are their AI infrastructure partner, and we work with people in many different ways. We do not require anyone to buy all products from us. We do not require them to purchase an entire rack. They can buy chips, components, our networking equipment. We have customers who only buy our CPUs, only buy our GPUs, and purchase other companies' CPUs and networking equipment. We can sell in any way you prefer. My only requirement is that you buy something from us.

Host:

We discussed Elon Musk with X.ai and the Colossus 2 project. As I mentioned earlier, it’s not just about better models; we also have to build. We need world-class builders. And I think the top builder in our country might be Elon Musk.

We talked about Colossus 1 and the work he did there, building a coherent cluster of hundreds of thousands of H100 and H200. Now he is developing Colossus 2, which may contain 500,000 GPUs, equivalent to millions of H100 in a coherent cluster. I wouldn’t be surprised if he reaches the gigawatt level faster than anyone else. What advantages are there in being a builder who can both construct software and models and understand the conditions required to build these clusters? Jensen Huang:

I want to say that these AI supercomputers are complex systems. The technology is complex, and due to financing issues, procurement is also complex. Obtaining land, electricity, and infrastructure to power them is complicated, as is building and launching all these systems. This may be the most complex system problem humanity has ever undertaken.

Elon has a huge advantage in that, in his mind, all these systems are working together, and all the interdependencies exist in one mind, including financing. So he is like a large GPT; he is a large supercomputer himself, the ultimate GPU. He has a significant advantage in this regard, and he has a strong sense of urgency and a genuine desire to build it. When will and skill come together, incredible things happen. This is quite unique.

Host:

Next, I want to talk about sovereign AI and the global AI race. Thirty years ago, you couldn't imagine you would be communicating with princes and kings in palaces, frequently visiting the White House. The president says you and NVIDIA are crucial to U.S. national security.

When I look at this situation, it's hard to appear in those places if governments don't see it as at least as existential as we treated nuclear technology in the 1940s. While we don't have a government-funded Manhattan Project today, it is funded by NVIDIA, OpenAI, Meta, and Google. We have companies today that are on par with nations. These companies are funding endeavors that presidents and kings believe are existential to their future economies and national security. Do you agree with this perspective?

Jensen Huang:

No one needs an atomic bomb, but everyone needs AI. That is a very big difference. AI is modern software. That is my starting point—from general computing to accelerated computing, from manually coding line by line to AI coding. This foundation cannot be forgotten; we have reinvented computing. No new species have emerged on Earth; we have simply reinvented computing. Everyone needs computing, and it needs to be democratized.

That is why all countries realize they must enter the world of AI because everyone needs to keep pace with computing. No one in the world would say, "I was using a computer yesterday, and tomorrow I'm ready to go back to sticks and fire." Everyone needs to transition to computing; it is just modernizing.

First, to participate in AI, you must encode your history, culture, and values into AI. Of course, AI is becoming increasingly intelligent, and even core AI can learn these things quite quickly. You don't have to start from scratch.

So I believe every country needs to have some sovereign capability. I suggest they all use OpenAI, Gemini, these open models, and Grok; I suggest they all use Anthropic. But they should also invest resources to learn how to build AI. The reason is that they need to learn how to build it, not just for language models but also to build AI for industrial models, manufacturing models, and national security models. There is a lot of intelligence that needs to be cultivated by themselves. So they should have sovereign capabilities, and every country should develop it

Is this what you hear around the world as well? They all realize this. They will all become customers of OpenAI, Anthropic, Grok, and Gemini, but they also really need to build their own infrastructure. This is the important idea that NVIDIA is working on — we are building infrastructure. Just as every country needs energy infrastructure and communication internet infrastructure, now every country needs AI infrastructure.

Let's start from other parts of the world. Our good friend David Sacks, the AI team is doing an excellent job. We are very fortunate to have David and Shriram in Washington, D.C. David's work at AISR and how wise it was for President Trump to place them in the White House.

At this critical moment, technology is complex. Shriram is, I believe, the only person in Washington, D.C. who understands CUDA, which is somewhat strange. But I love the fact that at this moment of technological complexity, policy complexity, and the significant impact on our nation's future, we have a clear-minded person who invests time in understanding technology and thoughtfully helps us navigate through challenges.

Technology is now as fundamental a trade opportunity as corn and steel were in the past. It is an important component of trade. Why wouldn't you want American technology to be desired by everyone so that it can be used for trade?

Trump has done a few things that are very good for getting everyone up to speed. The first thing is the re-industrialization of America, encouraging companies to build in the U.S., invest in factories, and retrain and upskill the tech workforce, which is extremely valuable for our country. We love craftsmanship, I love people who make things with their hands, and now we are going back to building things, building magnificent and incredible things, and I love that.

This will change America, without a doubt. We must recognize that the re-industrialization of America will fundamentally be transformative.

And then of course there is AI. It is the greatest equalizer. Think about how everyone can now have AI. This is the ultimate equalizer. We have eliminated the technology gap. Remember the last time someone had to learn to use a computer to gain economic or career benefits, they had to learn C++ or C, or at least Python. Now they just need to learn human language. If you don’t know how to program AI, you tell AI, "Hi, I don’t know how to program AI. How do I program AI?" AI will explain it to you or do it for you. It does it for you. It’s incredible, isn’t it? We have now eliminated the technology gap with technology.

This is something everyone must participate in. OpenAI has 800 million active users. My goodness, it really needs to reach 6 billion. It really needs to reach 8 billion soon. I think that’s the first point. Then the second point, the third point, I believe AI will change tasks.

What people confuse is that many tasks will be eliminated, but many tasks will actually be created. But it is likely that for many people, their jobs are effectively protected

For example, I have been using AI. You have been using AI. My analysts have been using AI. My engineers, each of them is continuously using AI. We are hiring more engineers, hiring more people, and recruiting comprehensively. The reason is that we have more ideas. We can now pursue more ideas. The reason is that our company has become more productive. Because we have become more productive, we have become wealthier. We have become wealthier, and we can hire more people to pursue these ideas.

The concept of AI causing large-scale job destruction starts from the premise that we have no more ideas. It starts from the premise that we have nothing to do. Everything we do in life today is the endpoint. If someone else does that task for me, I am left with only one task. Now I have to sit there waiting for something, waiting for retirement, sitting in my rocking chair. This idea makes no sense to me.

I believe intelligence is not a zero-sum game. The smarter the people around me, the more geniuses there are around me, surprisingly, the more ideas I have, the more problems I imagine we can solve, the more jobs we create, and the more employment opportunities we generate. I don’t know what the world will look like a million years from now; that will be left to my children. But in the coming decades, my feeling is that the economy will grow. Many new jobs will be created. Every job will be transformed. Some jobs will be lost. We won’t be riding horses on the street, and that will all be fine. Humans are notoriously skeptical and poor at understanding complex systems, and they are even worse at understanding exponential systems that accelerate with scale.

Host:

We have talked a lot about exponentials today. The great futurist Ray Kurzweil said that in the 21st century, we won’t have a hundred years of progress. We might have twenty thousand years of progress.

You mentioned before that we are lucky to live in this moment and contribute to this moment. I won’t ask you to look ahead 10, 20, or 30 years because I think that’s too challenging. But when we think about 2030, things like robots, 30 years is easier than 2030. Really? Yes. Okay, so I will give you permission to look ahead 30 years. When you think about this process, I like these shorter time frames because they must combine bits and atoms, which is the difficult part of building these things.

Everyone is saying this is about to happen, which is interesting but not entirely useful. But if we have twenty thousand years of progress, think about Ray’s statement, think about exponential growth, and all of our audience—whether you work in government, in startups, or run large companies—you need to think about the speed of accelerating change, the speed of accelerating growth, and how you will collaborate with AI in this new world.

Jensen Huang:

Many people have already said a lot of things, and they all make sense. I think one of the really cool things that will be solved in the next five years is the integration of AI with robotics. We will have AI walking around us. Everyone knows that we will all grow up with our own R2-D2 That R2-D2 will remember everything about us, guide us along the way, and become our partner. We already know this.

Everyone will have their own associated GPU in the cloud, corresponding to 8 billion people and 8 billion GPUs, which is a feasible outcome. Everyone will have a model tailored to themselves. The artificial intelligence in the cloud will also be reflected in various places—inside your car, in your own robot, it will be everywhere. I think such a future is very reasonable.

We will understand the infinite complexity of biology, understand biological systems and how to predict them, having a digital twin for everyone. We have our digital twin in healthcare, just like we have a digital twin when shopping on Amazon, so why not have one in healthcare? Of course, there will be. A system that can predict how we age, what diseases we might get, and anything that might happen soon—maybe next week or even tomorrow afternoon—and predict it in advance. Of course, we will have all of this. I think all of this is taken for granted.

I am often asked by the CEOs I collaborate with, what will happen with all of this? What should you do? This is common sense in rapidly developing things. If you have a train that is about to go faster and faster at an exponential rate, what you really need to do is get on board. Once you are on board, you will figure out everything else on the way. Trying to predict where that train will be and then shoot bullets at it, or predict where that train will be—it is accelerating exponentially every second—and then figuring out where to wait for it at an intersection is impossible. Just get on when it is moving relatively slowly, and then develop exponentially together.

Host:

Many people think this happened overnight. You have been working in this field for 35 years. I remember hearing Larry Page say around 2005 or 2006 that Google's ultimate state is when machines can predict questions before you ask them and provide answers without you having to search. I heard Bill Gates say in 2016 that when someone said everything has been done—we have the internet, cloud computing, mobile, social media, etc. He said, "We haven't even started." Someone asked, "Why do you say that?" He said, "We only truly start when machines go from being stupid calculators to beginning to think for themselves and think with us." That is the moment we are in.

Having leaders like you, like Sam and Elon, Satya, etc., is such an extraordinary advantage for this country. Seeing the collaboration between the venture capital systems we have—I am involved in it, able to provide venture capital for people.

This is indeed an extraordinary time. But I also think, one thing I am grateful for is that the leaders we have also understand their responsibilities—we are creating change at an accelerating pace. We know that while this is likely a good thing for the vast majority of people, there will be challenges along the way. We will address them as challenges arise, raise the baseline for everyone, and ensure this is a win, not just for some elites in Silicon Valley. Don't scare them, bring them along We will make it happen.

Jensen Huang:

Yes.

Host:

So, thank you.

Jensen Huang:

Absolutely correct