Exactly.
Good morning.
Good morning, everybody. Welcome to the Goldman Sachs Communacopia + Technology Conference. My name is James Schneider from Goldman Sachs, and we are really thrilled to have NVIDIA Chief Executive Officer Jensen Huang with us today. Welcome, Jensen. Thanks for being here.
Thank you. Great to be here. Thank you.
Now, Jensen, last year, your prediction of $3 trillion-$4 trillion in AI infrastructure spending by 2030, I think, raised a lot of eyebrows in the investor community. But it seems like we are actually really rapidly progressing toward that figure right now. What technical advancements or market developments with respect to AI-
I just think we should just take a pause and acknowledge that I was right.
We're doing that.
Just take a pause. Take your time.
What do people miss or underappreciate, and sort of where do you stand with respect to that outlook now?
The question is, why did I know that? The question is, why did I know that? It is actually fairly simple. The last industrial revolution made it possible for us to power everything, and we distributed power everywhere. Then, of course, the internet made it possible so that we can find anything. That is the big idea. You plug into the wall back in the old days, Ethernet jack or a modem, and now Wi-Fi.
You plug it in, and you can find anything. It is not a small thing. It is a very big deal. Because you could find anything, and because everybody wants to find everything, we had to make it part of the infrastructure. Because we can distribute power anywhere, we want to power everything, we distributed power everywhere. It became part of infrastructure.
Now we can find anything, find everything, but obviously that is not what people want. What we all want is to know everything. We want to ask anything, know everything. That is the layer of computing that we are building now that makes it possible so that whatever you want to ask and whatever you want to know, you can. That is the big idea with artificial intelligence, and a computer is in the middle of doing that.
The first instrument was called a dynamo, and the next one, obviously, called the computer, and now we have these AI factories, and they produce numbers just like the internet produces numbers. At some level is that simplistic. Now, the question is very simple. How did the computer industry go from that to this, and what is the implication? There are two compounding problems, two compounding challenges.
The first one is that in the last generation of computers, it was used by humans, and because you are looking things up, and you bought the computers as a tool, as a terminal, and so that you could find anything, find everything.
Now, in this new world, as long as you prescribe to the idea that everybody wants to ask anything, find and know everything, if that is a world that you prescribe to and that you believe in and that it could somehow be integrated into almost everything we do, then the question is, how do you build that layer of computing around the world? That computer, as it turns out, is a generative computer. It is not a retrieval computer, and I think I have explained this to you guys in the past. The last 60 years, everything was pre-recorded, and we put it on storage somewhere.
When you touch something on the phone, it goes and retrieves that piece of information, and in fact, it probably retrieved somewhere between three to 10 pieces of information. Based on who you are, the cookies that are associated with you, and your previous track record and your previous preferences, a recommender system would recommend one of those pieces of information to you, but it was all pre-recorded.
That model of computing is called retrieval-based. If your question—if you want to ask anything and know everything, you want to actually know it, you do not want to find it, you want to know it, then you have to generate the answer. You cannot reasonably retrieve that. The reason for that is because in the old days, you have this idea called a recommender system, and it is based on your preferences.
Well, if you want to ask anything and know everything, then that preference has to be replaced by something else, and it is called context. The query is called a prompt, and the surrounding environment includes the context and also your preferences. The combination of all of that, you have to generate the answer.
Basically what that says is that a new layer of computers that is continuously generating answers based on all the queries that are coming at it has to get built. That is no longer a bunch of storage. It is a bunch of computers. What does that computer look like, and what is the algorithm it runs? It is almost like everybody has our own recommender engine. Think of it that way. Instead of having one giant RecSys engine for all of Meta, you now have a recommender engine literally for everybody.
These recommendation engines are really complex because the AI models are large. It has to be smart. It compresses a lot of the world's information. Now the second problem goes like this. It is the end of Moore's law. You guys, the first time I said it some 15 years ago, there was a gasp, like I said something that hurt somebody's feelings.
It just says transistors do not scale anymore, and it is not a big deal. If the transistors do not scale 2x per year or get half as big every other year, then the question is what is going to happen to the future as we are trying to do this other thing? It is called generative AI or artificial intelligence as you might like to think about it.
In this new world, where it is the end of Moore's law and we need a lot more transistors, then several things has to happen. The first thing that has to happen is you still want to get 100X and 1,000X improvement every few years, and if you want to do that, then you cannot just live inside the chip. You have to co-design, which is the reason why NVIDIA became a co-design company, extreme co-design company. You guys hear that all the time from us.
The second thing that you have to do is if you want twice as many transistors, you got to build twice as many chips. Which is the reason why NVIDIA invented NVLink. You see, first thing that we did was we used CoWoS, and that made it possible for us to fuse multiple chips together.
Then that was not even satisfying, so we took a whole bunch of chips and we connected them together into NVLink, which is the big breakthrough today. If you do not have NVLink, if you cannot have really excellent scale-up, and scale-up technology is really hard. Scale-out is hard. Scale-up is incredibly hard. If you do not have that, you are dead in the water because Moore's law is your enemy, and these models are getting larger and larger and larger.
Then the third thing that happens, the compounded result of that is something that I predicted a while back, which is the semiconductor industry is going to be X times that larger. The reason for that is very simple. Demand of the semiconductor industry has been growing for some time, and yet Moore's law was a depreciating deflationary technology. That was happening at the same time.
If you don't have the benefit of deflationary technology, if demand accelerates even further than that because instead of one billion people using tools and we sleep, now we have hundreds of billions of agents and have to think. It's not like a chatbot. You don't hit it one time. You got to think iteratively. You compound all of this together.
The semiconductor industry is going to just keep getting larger and larger, which is what we're seeing now. These two fundamental ideas, that we have a new layer of computing with a new application and the end of Moore's law. Meanwhile, people are expecting these AI models to be smarter and smarter because they don't like wrong answers. Then the compounded result of that should result in a very large industry. Anyway, that's the end of my talk.
Excellent. I think one thing that investors have consistently questioned is the ROI of this technology. It's hard to look at a P&L today and say, "Oh, here's the gross margin," or whatever, but coding has clearly been a killer app for the industry. I think it's fair to say we're not going back to the old way on coding for sure. As you look ahead, what other applications or tasks do you think could really move the needle on both adoption and ROI for the industry?
Everything's coding. When you buy a cookbook, you see the recipe, that's just an American word for coding. When you ask somebody how to do something, that's coding. When you codify a business process, that's coding. Everything's coding. If you want to do something, if you want to discover the best way of doing something, and then after that, repeat it over and over again without anybody deviating from that process or that methodology or that best-known process, however you guys. Right?
There's a lot of words about it. It's all coding. We code in a lot of different ways. We code to make omelets. We code to repeatedly close our books and manage our supply chain. We code in order to communicate with each other in a consistent way. The connecting fabric, for example, the supply chain, is one giant large piece of code.
Not one code, but trillions of pieces of code. Everything that we do is in fact coding. It's a sensible thing. It's also the easiest thing for AI to learn, and the reason for that is because there's a right answer. The best answer is hard to find, but the right answer is not hard to find. Coding is an important part of it. A derivative of coding is, of course, bug finding, and a derivative of that, which is a very large market, is called cybersecurity.
The reason why there's so much conversation today about cybersecurity is because the industry is getting ready to launch some products. What better way to create demand than to create a problem. You know? Who doesn't want their market to be hysterical about their product and line up around the corner for it?
There are responsible ways of doing it, there are less attractive ways of doing it, but there is a lot of demand creation about cybersecurity today because new products are about to be launched. If you can code well, you must be able to debug well. Red teaming is finding a bug. Blue teaming is patching a bug. It is not a complicated concept. The fact of the matter is cybersecurity will likely be the next major use case of AI, and it is going to run continuously.
It will be a great new business opportunity for the labs. It will be a fabulous opportunity for CrowdStrike, and we have a big partnership with them, using open models to create red teaming and blue teaming and have the asymmetric advantage of swarms of Nemotron models that are running continuously.
We have a partnership with Cisco I think today, Palantir, Cisco, NVIDIA, they are building. Cisco is going to offer an entire AI factory platform from NVIDIA. Palantir is built on top of that. That is going to be taken out to all the countries and companies around the world to help them build either proprietary AI models based on Nemotron or cybersecurity red team, blue team models based on Nemotron.
Anyways, this is just, it is the time when it is the next click of AI coming out to the next market. Then, of course, there will be future use cases. But at the core, code is very important. I think that one of the things, Jim, that I was going to mention is if you look at. We spoke about the industry and all, and technology and all, but I am here to sell some NVIDIA stock.
I do not want this meeting, the agenda, to be unclear. We are the world's first and only growth value stock. I think we are in every. People are trying to figure out which one we are. It is like, we are both. You can be both at the same time. Why is it that we are both at the same time? Let me do some forensics on this for you. NVIDIA is incredibly misunderstood.
As large as we are, we are insanely misunderstood. The reason for that is this. We invented the GPU. You cannot uninvent the GPU. When you are NVIDIA, you cannot uninvent NVIDIA. Everybody knows because they have known me for 30 years. Most of the AI researchers in the world, most of the tech CEOs, they grew up on products I built. That is how old I am.
When an AI researcher comes up and says, "I used GeForce RTX 2060 when I was eight years old," that is not a compliment. They are just counting the years. My point is, we have always been a GPU company. It is not who we are today, but unfortunately, that is where we started. Does it make sense? We are not unproud of it, but most people think NVIDIA builds a chip. You need airplanes to ship what we build.
Each one of our system, each chip, if you will, one GPU is 2 tons. One GPU is now, it is not $399, not $399. It is $8.5 million. That is one GPU, all connected with NVLink, two million parts, right? 250,000 kilowatts. That is a GPU. We ship thousands of them. I was just seeing the reports this morning. Grace, Blackwell, NVLink, 72 racks, month-over-month increase 27%.
You don't have month-to-month increase of 27%. That compounds. That's called a high growth value stock. Number one, we went from Hopper, which is about $18,000. Now, Hopper is the name of an architecture, not a chip. Blackwell's the name of an architecture, not a chip. These architectures are expanding in their scope.
Hopper was about $18,000 or so per GPU system, Blackwell went to about $25,000, and Vera Rubin is about $40,000. The reason for that is because we're offering more and more and more of the overall AI factory. We see the AI factory in our head. We're trying to use extreme co-design to overcome the challenges of Moore's law.
Between algorithms and software and system and interconnect and new technologies that we invent along the way, we create a generation every single time that's many times faster or more productive in token generation capability than the generation before. The first thing that we do is we're increasing our SAM of the world's CapEx. Okay?
So mission number one. Not only are we growing, we're also capturing more at the same time. The second thing is, it is incredible, but there are more model companies today than there was a year ago, than two years ago. A lot more. There are a lot more frontier model makers today than there was a couple of years ago. There's a whole bunch who are starting up right now with a whole bunch of great ideas. Okay?
My point is, we're the only company in the world that actually runs every model. We didn't use to run Gemini. We run Gemini today. Of course, Grok is doing fantastic. Can't wait to try it. Of course, Meta's Muse. These are all new. These are all net new, right? Not to mention, we didn't use to run Anthropic for a lot of different reasons. We didn't have the money to invest in them as a young company.
Now we have more money and we're happy to invest in them and help them. Anyhow, our share of Anthropic is growing very quickly, and of course, we're delighted to see Anthropic and OpenAI and all of these labs growing. Okay? The second growth that most people don't see is we're the only company that benefits from frontier closed-open models. We run everything.
There's no company in the world who is indexed to open models, which is growing incredibly fast, except for us. You can't find that number anywhere. OpenRouter is a good place to go look at these things, but you could see the growth. The world needs both closed models and open models, and we address them all. Okay?
The third thing is the AI overall market is growing incredibly fast. What we see are just the CSPs. We just see the cloud service providers. Remember, they're enterprises. Some of the great names of enterprises building AI factories are, of course, Jane Street and Hudson River. Just about every quantitative trading company in the world is shifting into this new model of doing predictions.
And of course, drug discovery with Eli Lilly and Merck and Bristol Myers Squibb and many others have now created basically their robotics lab. It is basically lab in the loop, AI in the loop for wet labs. You need a supercomputer to do that. You have enterprises, but also the regional clouds. One of the biggest challenges, and I have been talking about this, and you want to make sure you understand the strategy.
Upstream, the supply chain is very challenging, and the reason for that is because obviously we are growing super fast. Packaging is a challenge. DRAM is a challenge. LPDDR DRAM is a challenge. Connectors are challenging. Everything is challenging. Voltage regulators are challenging. Everything is challenging. Wafers are obviously challenging. All kinds of different challenges upstream.
But remember, the supply chain goes all the way to the end downstream until somebody stands up a computer and turns on the service. We have been thinking about the supply chain upstream and downstream, and one of the advantages that we have, because our go-to-market, remember, we run everything. This is the power of general purpose versus specialization. Do you remember two years ago, three years ago, everybody used to say specialization is better than general purpose because it is faster?
Well, if I can make general purpose faster than specialization, then specialization is all good. The reason for that, excuse me, generalization is all good. The reason for that is because generalization gives you fungibility, durability, versatility, rentability, and very importantly today, because of capital constraints, everybody's balance sheets, investability. Finally, we have a computer. The NVIDIA Compute is finally a computer that could be asset backed.
We can use it to secure loans, and that is a very powerful capability, and frankly, the only computing stack in the world that allows you to be able to say that, because nobody could step back in on it. Our market includes CSPs and of course, the regional clouds, the neoclouds. The power of the neoclouds is this.
They secure land power and shell for us that the CSPs have already exhausted. Just remember this. This is a very big deal. Many countries and many counties and regions want to secure the power and land for their own companies. We have neoclouds around the world from Nscale, and of course, you got CoreWeave and Nebius, and they are doing fantastically.
Some new ones that are on the verge of going public or are in the process of filing, whether it is Nscale or Lambda or Firmus. You are going to see a whole new crop of really, really exciting neoclouds with hundreds of billions of dollars backlogged together. This is a capability that allows us to secure land, power, and shell downstream and be diversified in our way of going to market.
These three ideas, the whole AI market we serve completely, all of the models we serve completely, and of course, the world's data center CapEx, we addressed a lot more of it, which is the reason why we are growing so fast.
Yeah. Can I follow up on a couple of those points?
Yeah.
One on supply constraints. You talked about a couple of weeks ago in your earnings call, high confidence in delivering 70% revenue growth next year, unconstrained demand growth of over 100%. You talked about upstream and downstream. Are you more concerned about the upstream component shortages, or are you more concerned about the downstream, which you may have a little bit less direct control over in terms of land, power, shell, and data center availability?
Well, the downstream is where NVIDIA's advantage is incredible. Upstream, our advantage is because of our scale. We have the largest supply chain in the world. I have been working with the supply chain and these partners now for coming up on three decades. When I make a prediction, they come true. They like it when they are people who are right.
Because they have to put a lot of money at play and because NVIDIA has a track record of actually being helpful and truthful and good at predicting these things. Not to mention, we can create our own market. Do not forget, NVIDIA's platform is the only one that you can take all the way to the end market by yourself. Remember, we go to market through the CSPs, but we go to market through the OEMs. Look at Dell. They are doing incredibly.
Almost all 100% NVIDIA. I think it probably is 100% NVIDIA. Supermicro is doing well. Lenovo is doing well. HP is doing well. Cisco is coming in. We have 100% of the world's enterprise go-to-market can take NVIDIA to market because we are a full stack AI factory platform.
You come in, you can install the whole thing, add our software to it, and be basically up in a couple of weeks and get to work. Because these things are so expensive, the economics of it is so high that if you do not make it productive as fast as possible, the anxiety is really quite incredible. If you look at our go-to market, we have diversity of channels. We have many ways to get into market.
Because of our neoclouds and because of sovereign AI capabilities, all driven by the fact that we are full stack, we have a rich ecosystem, we can reach all of these different marketplaces. The downstream part of it is a huge advantage for us. There is no question land, power, and shell is a problem. When I hear these people talk about these big numbers, the question is, and we are tracking every single gigawatt of land, power, shell around the world, literally everything on the planet.
Just think about all my partners. How many neoclouds is reporting back to us? How many OEMs are reporting back to us? How many clouds are reporting back to us? How many AI native companies are reporting back to us? We are working with everybody. We know where everything is. It is pretty amazing.
We have secured a lot of it. How can you otherwise be upstream? Obviously, TSMC would never tell you. They would never say something like that they can somehow project how many gigawatts sold. How would they know? Because it is far away from downstream. But we are all the way upstream and downstream, which is the reason why NVIDIA's position is so good.
You feel good about closing that gap?
Yeah. We could grow 70% year-over-year, and we are confident about that. We have a year to go work on improving that. We are going to come to work every day and make our living and get more supply.
Yeah. Excellent.
More fried chicken.
More Denny's.
Yeah. More Denny's, more dinners. Whatever it takes.
Excellent. Your hyperscaler and large enterprise customers clearly can finance their spending independently.
Oh, did you guys see what we announced this morning with Australia? To give you an example, in Australia, there's a whole bunch of energy. We're working with all the companies and data center companies in the region. We stood up 2 GW for 2027. Just put in perspective, 2 GW for 2027. It's not a lot, but it's $80 billion.
Yeah. It's a lot.
You guys are so hard to impress these days. It's like two companies, but that's okay. I don't even think that was it. Was it a tweet or a press release? Was it a blog? Anyways, it's a very big deal because I love working. Australia is a region that have excess energy, as you know. But they need several things. They need technology.
This is where NVIDIA's full platform AI factory comes together. They need ecosystem off-takers. $400 billion of VC funding went into AI natives in the last six months. $400 billion. AI natives, the definition of an AI native is somebody who spends two-thirds of their raised money on compute. There is no such thing as a low CapEx software company anymore.
Every technology company is going to be a high CapEx company in the future, or high OpEx or high CapEx, but basically a high compute because there are no non-AI companies. We are working with just about every AI startup in the world. We have the whole platform. We have off-take. They bring the regional land, power, and shell, and we work with them to stand that up.
Lastly, capital. Sometimes we invest in them. One of the transitions, and this is the big idea that is happening right now, we are working with the financial industry and many of you, and really appreciate the work together. We are moving NVIDIA Compute from technology to an investable asset. This transition, and I think when we make this transition, it will happen here fairly quickly.
I think this is going to be a huge needle mover for people to recognize that our computing systems have long-term value, has durable value. At the moment, none of it is being valued. This is a huge untapped opportunity for us, and there is just so much evidence that the compute is productive. We sell it for, let us say, one- gigawatt data center is probably like $60 billion.
Let us say it is over six years, $10 billion a year. Let us just do some simple math. That $10 billion a year currently is being rented out for 50, as you guys know. The revenues per gigawatt right now is about $50 billion. That is the economics of—that is how productive our technology is. You can still rent Voltas. Volta is 10 years old. Obviously, Mike just talked about renting Ampere's.
All the Ampere's in the world are all rented out. Durability. Fungibility. NVIDIA runs every model. Every single lab can use us. Because our capacity is so large, it is like TSMC. One of the values of TSMC, of course, is technology, but one of its most important values is capacity. I can build my company on top of TSMC.
You can build your startup on top of NVIDIA, no question about it. We are not going to be your problem. You do not have to engineer us into existence. We are here for you as an AI platform. The way I see TSMC is the way that AI natives and AI companies see us. We are a foundational platform of the AI ecosystem, foundational platform of the AI industry.
To your point, you are providing, in some cases, guarantees or backstops for those AI labs or neoclouds to go secure land, power, and shell for their operations. You have done that a number of different ways over time, including the $500 billion platform you talked about recently. You said on your earnings call you do not believe that circular finance—
Hopefully, most of that $500 billion is going to be asset-backed.
Yeah.
That's the big idea.
Yeah.
That's the big idea.
So maybe help, for people who are skeptical, explain maybe why you believe it's not circular.
Well, it's not circular because we put a little bit of money in, and a lot of money comes back.
Yeah.
Is that finance talk? I look at the spreadsheet, we put in one and 100 comes back in. Is that circular?
Yeah.
If that is, let's do more of that. Not to mention the companies that we're investing in, we see their pipeline because we brought the pipeline to them. People are starting to recognize that wherever I invest, it's not a bad place to invest because I'm an informed investor. I'm not taking any risks. We're not smart like you guys. I need a sure thing.
So when I see they're-- We bring the AI platform to them. They have to secure the land, power, and shell. We help them with financing, but very small part of it, but the most important thing is we see their offtake because that financing doesn't come together without the offtake, and that offtake is $100 billion. It's lined-up contracts. It's total contracted value. It's real stuff. We know where the demand's coming from and the quality of the demand.
So we see the bigger picture, which is the reason why it gives us confidence to do it. Also, I think it's really smart strategy. I think it's necessary for us to help the industry create this network of neoclouds, which are going to be distribution channels for NVIDIA's architecture. It is also a place where all these AI natives that are being founded could land.
Remember, AI is not going to be just global. AI is going to be regional. The reason for that is because there's so many regional intelligences, and that's by definition, you're going to see AI becoming much, much more regionalized. So we want to have hubs of NVIDIA partners all over the world. So that's how. The circular part of it doesn't make much sense to me because the numbers shows it doesn't.
Yeah.
Yeah.
The returns.
The returns are too great.
Yeah.
It's too good.
Maybe, we only got a couple of minutes, but I would love to hit on physical AI before we close. That's an area where you've really seeded the market across everything from automotive to industrial robots to humanoid robots, et cetera. How quickly do you think that market is going to take off? What is going to be the killer app in physical AI, and how big an opportunity you think this could be for the company by, let's say, 2030?
The first killer app for physical AI is just self-driving cars. There was a time, and it made no sense to believe that, there was a time when people thought the way to build a self-driving car is just to drive billions and trillions of miles. Obviously, there would be a lot of casualty in between, but you wouldn't want that. What you are really looking for is a thinking car, a car that can reason.
We made some groundbreaking work in this area. It is called Alpamayo. It is the world's first reasoning and thinking car. It could see an environment it has never seen before, and it could reason about how to break it down, just like agentic systems are reasoning. It reasons, it breaks it down into things that it understands and can think about quite easily, just like us.
The number of miles that you need is actually quite few. Post-training is necessary, and I think physical AI, for example, has arrived for self-driving cars. I think in the next couple to three years, you are going to see really great progress. Obviously, Waymo, obviously Tesla, NVIDIA's Mercedes partnership. There is a whole bunch of others that are being teed up.
Our partnership with Uber. Lots and lots of partnerships being teed up. That is the first application. Derivative applications of that are things called AMR and warehouse delivery vehicles and inside logistics centers, and we announced a big partnership with Amazon. They have the largest fleet of inside warehouse navigation systems. You are going to see a whole bunch of that kind of stuff, whole bunch of applications like that. Okay, grocery delivery vehicles and all kinds of things.
The second part is manipulation systems. Manipulation systems are making really great progress, and today's manipulation systems are all pre-programmed, just like the old computers that we talked about was everything was pre-recorded. The current manipulation systems are pre-recorded, which limits the market size to just the biggest car companies.
If you want manipulation systems to be useful for the mid and medium-sized manufacturing companies, which in Germany is called Mittelstand. Here in the U.S., they just call them supply chain partners. In Japan, there is their version of Mittelstand. Basically, the industrial supply chain are a couple of hundred million dollar companies. If you look at ASML supply chain, it is a whole bunch of Mittelstands. Okay?
A bunch of technology companies, each one of them is probably a couple of hundred million dollars, and none of it can be robotic because none of it is big enough. We need smart robots, that reasoning systems, and so that capability is probably a couple of years away. After that, physical AI will permeate into everything from telecommunications, for example.
6G is basically physical AI. Instead of using cameras that understands the physical world is using a different spectrum of electromagnetics, radio waves, but it is basically sensing the world. We have a partnership with Nokia, and we built a platform that is based on NVIDIA's CUDA and AI, and it is called AI-RAN. The system, I think it is going to be phenomenally successful. It is basically a distributed edge data center. All of this is, I think, probably about five years into cooking.
You will just see more and more of it. In the meantime, NVIDIA is already successful in physical AI because before you deploy the model, you have to train the model. People know that Elon and I worked together in the Tesla supercomputers that he called Dojo, and that has a whole bunch of GPUs inside. That is for training these self-driving cars. There is a whole bunch of examples like that.
Yeah. I think I have virtually got to wrap up there. Thanks so much, Jensen, for being with us. We appreciate it.
All right, guys. Thank you.