Nvidia Corp (NVDA) June 4, 2026 Conference Transcript

Good morning.

Welcome to day three of the Bank of America Global Technology Conference. I'm Vivek Arya, covers semiconductor, semi-cap equipment, and I'm really delighted and honored, and it's a real treat to have Colette Kress, Executive Vice President and CFO of NVIDIA to join us for a keynote session this morning, fresh off a number of announcements from GTC Taipei, right? So perhaps, Colette, if you could start with maybe giving us some sense of what NVIDIA announced, how it kind of fits your strategic direction, and then we can go and do a few other questions. But thank you so much for joining us today.

Thank you for having me. I'm going to give you one quick statement at the very beginning here. As a reminder, this discussion may contain forward-looking statements, and investors are advised to read our reports filed with the SEC for information related to our risk and uncertainties facing our business. Yes. GTC Taipei was a really, really great event. We do enjoy going to Taiwan, meeting with such an important part of our suppliers and helping our suppliers and that full ecosystem really understand the progress we're making and speaking to them directly on that. So there were great announcements that we did make. One of the top announcements is, let's not forget, Vera Rubin is on its way and definitely is in full production. As we had discussed part of our earnings, we indicated that, yes, it is planned for the second half. But more importantly, we talked about it. It's coming soon. It's ready for Q3. So Q3, we're looking forward to standing up Vera Rubin. As you can imagine, we're already in full production in order to make that plan within Q3. But that was one of the very first pieces of it. The other piece is some of the excitement that folks have talked about, which is referred to as Vera CPU. Vera CPU is a great opportunity for us to both continue to expand with an important area of agentic solutions. When you think about agentic solutions, that CPU is going to be an essential part, the director of directing that agentic work together. And we have the ability, again, to do a extreme co-design with what we enabled with a CPU. That extreme co-design is different than many other types of CPUs that are out there. It is on our own cores. It also has a tremendous improvement in terms of productivity, but also in terms of just its sheer performance is about a 2X versus any of the other x86 CPUs. We've added this to our portfolio not only within our full systems that we have for Vera Rubin or even what we have with our existing Blackwell, but this is also an opportunity to sell standalone, and we do believe that will be a big opportunity for us. This is also a time to talk about what we can do in terms of the PC. We've been working very feverishly with many different providers to help us on that, but this is putting together RTX Spark into the market. This is a great opportunity for AI types of PCs and very, very key in focusing those that will be doing a agentic type of work as well and using that PC with the performance. It's built together with MediaTek as well as our GPU, and that's a great opportunity for us. Lastly, this was a discussion to talk about really the diversification of both what we are building, but also the different types of customers and users that we see worldwide. We're in a unique position versus any other type of company. And you'll hear more of that, I think, in our discussion when we talk about not only just the hyperscalers, but a very important group of AI clouds and what they have built in terms of the market. We have so many different folks that we need to help with. That will be the enterprises, that will be the industries, that will be sovereigns. And we have that complete diversification in everything that we're selling. So those were some of the key highlights that we had.

So maybe let me pick up, Colette, on that point, because I think one of the very important disclosures you made as of the last earnings call was giving us this transparency between hyperscalers. And I was struck by two things. One is that both businesses, right, both your sales to kind of the large public hyperscalers and this other business to kind of the neocloud, sovereign, right, on-premise, they're about the same size, but the hyperscale business is growing faster. Like, I would have thought that the other business would be smaller and growing faster, but it's the same size and growing slower. So, if you could just walk us through what is the right way to interpret that disclosure?

So, let's step back and talk about our disclosure we've had. We've been providing, of course, the data center numbers that we've had. One of our focus was to provide you an understanding of compute and networking. But what was interesting is we're going on our now third generation of an extreme co-designed full data center scale system. We're taking that and breaking that down to show you compute and networking. But if you think about it, all of those full systems are always going to incorporate both our compute and networking together. And our attach rate of what we are seeing in terms of the networking is an enormous piece of that as well. We're probably more than 90% attach rate in terms of what we're doing on networking. So now let's step back. What did we want to show you with a new way to think through what we're doing? You get asked all the time in terms of what percent of our revenue was the hyperscalers. And that was, we had indicated for many years, many quarters, indicating that it's approximately 50%, plus or minus any single one quarter in terms of what we see. Now, we are growing underneath that. So, when you're saying it's 50%, that full growth coming from the hyperscale is there. They're a very important group in terms of what they have done in terms of standing up clouds. But what's interesting is the other 50%. And this is what we brought attention to, which is unique to NVIDIA in terms of our AI focus clouds. What we're looking at in terms of this group, in terms of the ACIE is what we are calling it, is a focus in terms of these types of AI clouds. These AI clouds are newly standing up, not a general purpose cloud that has been there, But focusing on either AI factories, but most importantly, focused on our reference architecture, because they need to know that full stack in order to put that into space. So what they've done is a fast-moving capability to serve enterprises, to serve the industries, and serve many of the different countries and regions that we see. We have AI clouds definitely here in terms of the United States. You see it building already in terms of Europe, and now also big areas in terms of Southeast Asia as well. These are important areas to serve this market. Folks are not going to move back into their own enterprises, data centers to build this. They need an AI cloud builder. This is where the token generations and pieces will be. So interestingly, that's the other 50%, and such a fast-growing, as Jensen actually communicated, likely the fastest growing piece, both their speed and the need to serve such a big market. You've got that ability now with this group. So we'll continue to be providing you that information broken out in that manner. There's a lot of interesting things happening in that.

Got it. And what I also found interesting is, despite this, you know, kind of daily debate, discussion noise about how much ASICs might be impacting your business, your business and hyperscalers, if I have my numbers right, was up like I think 115%, right? So it was definitely at a very strong place. Are you well represented at all hyperscalers? Is there some, you know, concentration at certain ones, you know, where people have custom chips? So how are hyperscalers kind of making this determination between where should they deploy their ASICs, where should they deploy NVIDIA GPUs?

Yeah, very good question to talk about the word diversity, because if you listen to each of the different groups, they're communicating the diversity. But let's step back and say, the king of diversity is probably NVIDIA and everything that we have done. Because each and every cloud that is there, whether that be a hyperscaler cloud, and or an AI cloud, we are also being a part of them. Also, when you think about the AI model makers, when you think about the foundation models. They are also 100% on our platform as well. So we have the diversity of both what the types of models there are. We have the types of clouds that it are. Meeting that together, we demand to be the biggest drawer of the diversity going together. The cloud operators, they have continued to use us not only for today, tomorrow, but in the future because what they are doing is building a significant amount of understanding of how we have designed at an extreme co-design. Most of them are actually continuing to sell what we have in the cloud for them, whether that be our Hopper architecture, whether that be our Blackwell architecture, and our future of Vera Rubin. What that means, it's sustainable to stand up, and we continue to watch It grow and grow in this area that they will be able to serve so many of the different markets because of what we have built in a full ability for them to not only do end-to-end types of solutions, but we can be ready for what we did at the very beginning of ChatGPT to what we are seeing now in terms of Agentec. And that diversity of all of our different customers is very key. Nothing that any other type of builder really has because we are supported and very helpful for them.

The next topic I wanted to touch on, Colette, was that, you know, there is this perception that as the market is moving from training to inference that, you know, dominated the training phase, that inference is going to be a lot more fragmented. Although when I look at these kind of growth rates, it tells me you're participating well in it. So how do you kind of frame the discussion that is it a fair pushback to say that inference, right? You know, like I saw this one headline which said, oh, the center of gravity is moving from the GPU to the CPU, right? So are you starting to notice a different kind of competition as the market moves towards Like, what percentage of your business today is in inference versus training, as an example?

Okay, so let's step back on the two pieces here. Basically, thinking about there's a training part of it and there's an inferencing part of it. The inferencing part is extremely important for many of these model makers or any of these new companies starting. Why? Because once you're moving to inferencing, now you've got revenue. As you've seen the growth right now in agentic types of solution, it is the only way to describe it is the growth is vertical. The percent is not the most important thing. It is vertically going up in terms of that. Once you reach a point that you are now at revenue, the most important next thing that you want to think about is how do I make sure I can get a profit? And how can I make sure that I am using the best of breed in order to serve that? Therefore, for those tokens, I want as many tokens as possible in the shortest amount of period of time, and I want the most efficient use of that type of compute. And that's where we come in. Because when you think about our systems, they've all been designed, thinking about these models, thinking about the full software stack that is also necessary for it, but being the most productive as well as the most lowest cost. So this inferencing is not any easier. It's actually harder if it is just a static ASIC, because we have been continuing to design over and over again. We still look at this as helping them focus on what type of use they want to use for the inferencing because it's not that simple, but it's one case as well. As you're working right now that says low latency as we go forward, there is a piece of agentic that says I'm going to go back to the director of the CPU and ask that CPU to assist in this piece of it. But again, that hard work is still going to be necessary on our full end-to-end systems and with that important part of the GPU as well. So what percentage is of our inferencing? Almost every one of our systems, both is using for training, but is also easy to move over to also do inferencing. With Grace Blackwell, it was the first time that we'd also seen inferencing first for some of the systems that are built. Most of those AI model makers that you see right now and the models that they've put out there, Grace Blackwell is very, very key for much of the inferencing that is happening.

Got it. One other thing that I've noticed is, you know, with the last few years, we saw a lot of deployments, and I think it was harder for investors to kind of connect the dots to ROI and monetization. But this year, we have actually seen much better. Is that a fair representation of what's going on? What are you hearing from customers? Are they noticing real ROI now? And what has kind of created this switch?

That's correct. The ROI has absolutely represented a significant amount with the cloud providers. The clouds were already making a great return as they continued to sell on the cloud. Now, when you are also supporting those foundational models and what their work is doing in a Gentic, you are seeing not only enterprises focus, but you have the entrance of consumers as well. Folks working on a Gentic types of work, even at a personal level, is also very key. That is enabling the revenue, which is therefore enabling more and more profit that allows them again to serve more. Most of them right now are in a position that says the compute is tight. There is still a shortage of supply of compute stood up that they want more and more that they are trying to get. And we're working with so many of them to help them as this agentic plan that we knew would be a great part of the next stage is actually almost a greatness that is causing, again, more and more compute needs. And we're going to work very hard to serve many of that.

You know, we have seen just insatiable demand, right, from these public LLM, you know, companies, if I can call them that. Now, the skeptic will say that, you know, they are trying to get their hands on compute wherever they can find. Does it not mean that hardware, to some extent, is being commoditized because of that? Because if they don't care whether they're using a five-year-old GPU or the latest generation GPU, they're just trying to get compute, whatever. So does that speak to commoditization or does that just, you know, say that there is just a...

It's actually the opposite. It's actually given that full strength of agentic, it is forcing them to say, I've got to find the best of breed to do that. I agree that they're working on many different solutions to try and put that together. But you really have to look at where would the lion's share of that come from? We tend to continue right now growing our capabilities, growing our ability with more and more supply. And we are serving a vast majority of what is needed right now in this agentic role. And I think that's going to continue. The perception that, hey, any chip will work, we're not a chip. It does take a full end-to-end data system to do that because you really have to think about the work that it needs to be done in agentic starts at the onset of the information landing in that data center all the way through to the end. No simple one chip would be able to solve that. We think about all the seven different chips that we'll even have with the Vera Rubin fully designed for these types of solutions that we can answer all different types of questions all within one full system. And that's why we are continued to be looked at. But we have certainly engaged with a lot of different companies, matchmaking with them, where how can we help them both obtain the LAM PowerShell? How do we help them in terms of standing up the compute as fast as possible for what they need to do?

Got it. You know, given, you know, you mentioned LAM PowerShell. Those are very, very well, I think, recognized constraints. Do you think they are constraints to a point where your existing customers might say, you know what, it might make more sense to just upgrade my existing GPU infrastructure? Like, has that upgrade been a factor so far? And when does it start becoming a factor?

Yeah, it's an interesting debate that says you have data centers that may have already incorporated the hopper architecture. What is interesting right now is the Hopper architecture in the cloud is actually earning more money than the very first day that they had it, because that has been something that has continued to improve, whether it be Hopper, whether it be Blackwell, and you're going to see the same thing with Vera Rubin. We continue to improve it with software to align with what currently is happening in the market to provide that with inside those systems. So that is actually helpful for them. A, the depreciable life may be a certain amount of time. The useful life is very long. If you think about it, why they want to keep it up and running, the time down to then build another data center versus this is actually quite useful for us is much easier for them right now. So particularly in this time, the brand new data centers are also important because you want your best on that. If you spent hours, years trying to find the exact land power shell, given that it can be a very tight area right now to find, you're going to want the best. You're not going to sit there and say, I'm going to use something and give it a try. They know that NVIDIA's architecture has continued to ramp for this, and that's why many of them are on that plan for all of these different data centers.

So right now, to your point, the focus is obviously Greenfield. It's a lot of the newer products. And then existing, it's try and get as much, right? So upgrade is not yet a factor. But do you think there is a point at which that tradeoff makes more sense that you can generate so many more tokens, right, and generate so much more revenue that that offsets whatever depreciation benefits you can get from an older generation product?

At some point, it will get to be there, but it's an interesting time right now, given the tightness. Those new data centers that are being built are absolutely built with the most sustainability, but also thinking about the efficiency of power, and most of them are liquid cooled. When you're looking back in terms of a hopper, you'd have to change that data center. It will happen. They will move that. But right now, moving ahead, getting ready for Vera Rubin is what the key thing they're doing.

And I'm glad you mentioned software, because I find that sometimes that is sort of underappreciated about the full stack aspect of the platform, right? And the reason one can extract a lot more value, even from existing hardware, right, is because of that constant innovation in software. So maybe talk to us about what is the role of software? What is the role of having all these domain-optimized libraries and developers? Because sometimes I find that people just kind of treat it with a throwaway line of, oh, the CUDA mode is broken. And I imagine it's a lot more than just one operating system.

If you discuss with any of the foundation model makers or any of those creating these types of solutions. They are all focused, yes, as the underlying CUDA, but those key things is the domains and the libraries of software that we have. With you think through any type of AI cloud that is standing up. One, they're working on, can I design a data center? Can I stand this up? They need a full end-to-end architecture. They don't have thousands and thousands of software engineers that for over the last 20-25 years have designed software that is both backwards compatible and forwards compatible as much as they are leveraging this and could up. So our work in terms of those libraries for each of the industries, each of the enterprises is going to be fundamental particularly to the AI clouds and the work there. They can get a full stack. They don't have to rethink redesign it but it's more important to that even within the cloud and you have the foundational models in the cloud they too are also very keen onto that stack of software that we are working together with them to continue you know to enhance we have our own models with inside those models are leveraging our software and building upon our software because we do also understand how models are working.

Absolutely. And none of that is something that ASICs can do.

Absolutely not. It doesn't have the ability. It's a fixed. It was determined at the point that they went and designed it. The time that they went to go tape it out, it's done. It doesn't have the ability to change over that. But we have that because we have a full platform to revise and any part of that from here and going in the future.

I wanted to touch on two more topics in the last few minutes. One is on constraints and then the CPU side. So on the constraint side, you know, given the strong growth that you're seeing, what are you doing to, are there constraints? Like what is the constraint? How are you dealing with it? And, you know, can the fact that, you know, you have prepared for it, is that a competitive mode by itself also in this industry?

So it's been interesting to watch as we knew the growth and we have been continuing to talk about the growth going forward. Remember, we think about 2025 through 2027, if we think about Blackwell and Rubin together, that's one trillion dollars. We knew that. We knew that that opportunity would be in front of us and we knew it would probably exceed that. What that means is you have to be thinking about your suppliers, your full ecosystem, because I would say it's not just the supplier piece of it. You also have to think about the land power shell as essentially a supply. It is a principal piece of what we need to do to complete. So our work is longstanding. Our supply is not just ordering what is there. Remember, often we are in deep design with many of our different suppliers at the very onset of what we are putting into our full systems. That design product is working with them way ahead of where we are now in market. You would say probably more than three years ago, those are where some of those discussions started. And then that is a continuous, ongoing focus in terms of the supply we need. We have some significant amount of purchase commitments. It's a number that I don't think any significant company would look at that and say it's a small number. We're essentially at about $124 billion of commitments.

The entire logic semiconductor industry was that a few years ago.

It's an enormous number, but keep in mind, there's an ongoing every day. You never know what day you are fully where you need to be from the supply. Demand arises the next day, and we're in that continued discussion with our suppliers. No one's immune to the supply constraints. We all have our continued work trying to help them with it, but given our size and given our longstanding work with them, we can continue to think long-term with them. It's important to understand it's not necessarily how they divvy up supply. the question is how do they stand up the capacity the manufacturing lines to continue what they're doing and that's where we often come in how can we assist how can we help you understand what is going to be necessary their first question is how many manufacturing facilities do i need to support nvidia so it's an important area i can make the same type of statement for land power shell, how do we help them stand that up? How do we align that with people that are interested and didn't have enough compute, but now need more land power shell, we can also be continuing to work that ecosystem as well.

Right. So despite the inflation we see in everything around us, you think NVIDIA can kind of sustain the kind of margin levels despite all this because of the planning work and pre-borges commitments?

Nothing as we see moving forward changes to where we are today. Even when we think about putting Vera Rubin in market, we have done so much work and is very well appreciated with many of our customers, both on the manufacturing side and the cycle time. How long will it take me to get this product into the first day of revenue from what they're putting in together, and they've actually really seen a true difference. We're on our third generation, and that is moving very fast, but also very efficient, not just focused on the system, but the full manufacturing of what it takes to get that up and running fast. So at this time, I don't see anything different going forward. Does this continue for us? Yes, and we'll find out more later, but that's where we stand.

Understood. Before I ask about the CPU, one other thing I kind of remember from what was said at GTC Type A was the ability to improve monetization, right, per gigawatt. I think there was a kind of a roadmap given of, you know, from 40 billion towards 60 to 80 billion. What is driving that? Because those are pretty big numbers, right, and pretty serious. I mean, that's like your content expansion with every new generation.

Every new generation takes more and more focus on every last piece of that system and getting it accelerated. Not focusing on a very single one chip. One of the comments we've made in terms of air movement, it's seven different chips. When you think about the importance of just the switching capability, the connectivity of moving, how fast can we move things across? How can we make sure all different types of asks of the compute are being held and accounted for in terms of that? So that's where the CPU now becomes an important part in terms of how we put that in there. It also thinks of all of the different switches. It thinks about in terms of the adapters and pieces of that. You've got a blue field opportunity included in this. And then also what we can do just from a sheer GPU and the overall development of that. So you bet we're going to continue to see not just because we're focusing on the chip is because you focus on that full data center and that capability.

Makes sense. And then just the final topic in the last 90 seconds, which is on the CPU side. So what drove your decision to invest in the CPU? It's not the biggest part of the content. What has made it more important? And why invest in ARM, right? Why not just, I mean, you were using x86 and many of the head nodes before. Why not just continue to stay with that?

Yeah, the design that we've done is very well aligned to agentic type of solutions. That verification of the agentic work that needs to be done. So the hard work sits with that GPU and now that CPU is really right in line for what that's going to be necessary. However, it doesn't mean that you just want to build something off the shelf because you still want that connectivity of the software. How does all that work together? It is more performant. Remember, it's a 2x more performant, which is an important piece. But it is also our own cores in terms of what we put together. If you don't mind, I'd love to take the last 30 seconds here to remind folks In terms of where we've really gotten to in a very good place in terms of returning capital to our shareholders.

Ah, yes, of course.

I know. It's a really great opportunity. You know that we've had to focus in terms of on our suppliers, our ecosystem to build so much of what we've done. But given now where we stand with a significant amount of cash flow generated from these full systems, The amount that we can return to shareholders, 50% or more, absolutely, is a key focus of ours. Not just for today, not just for tomorrow, but for the long term, we'll continue to focus on this. But we also added the dividend. Our dividend has been with us, yes, for 13 years.

Since you joined us.

So that is correct, as far as when it started. And it's now time to also serve a very important market that also has the ability to, from a capital return, choose in terms of our dividend as well. A great opportunity for a dollar a share per year. And that is, again, an area that we will continue to focus on for growth, not just our share repurchases, but also our dividend. So I just want to make sure that all were clear.

Why limited to 50 percent?

Why not 75 percent? we're working on it and we'll let you know in terms of there but how about we'll go through this year and a dollar each maybe next year when we when you're here for the keynote maybe by that we'll come back next year we'll talk about it sounds good thank you thank you so much really appreciate your time