NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Good afternoon, everyone. Welcome to Day 1 of Citi's 2021 Virtual Global Technology Conference. It's my pleasure to welcome Colette Kress, CFO from NVIDIA today. The format of our discussion is fireside chat. I'll kick it off with my questions, and then we'll do the questions from the audience. My introduction, I'm Atif Malik. I cover U.S. semiconductors and capital equipment stocks here at Citi. Welcome, Colette.

Thanks so much for having us. Appreciate it.

Colette, thanks for coming to the conference. Your keynote is always a big draw for our investors. And I always find new and exciting topics to talk about.

So let's start with the data center business first. Data center sales grew strongly, 16% sequentially in the July quarter. And you expect accelerating growth, with data center being the lion's share of the October quarter sequential growth. Can you talk about where we are in A100 adoption curve? And what is the approximate split of A100 versus V100 in the installed base?

Right. So first, starting off, our A100 is our Ampere architecture for the data center. Our A100 has been in market for approximately a year or so, and it has been widely adopted by the major hyperscales as well as the CSPs, adoption in terms of use or internal use but also for cloud instances. Our A100 is special. Our A100 comes with an HGX platform board, and it enables our hyperscales to not have to choose how you want to use the A100. It is capable of doing both, capable of doing both training as well as [ in consent ]. So it has been quite popular, easy in terms of adoption, in terms of general availability, almost in all cloud worldwide. But it also has encouraged enterprise adoption to be strong as well. Our NVIDIA-certified system with our OEMs has helped accelerate the mainstream adoption that we're seeing. We have hundreds of thousands of enterprises already looking at A100 and/or deploying A100. When you think about what percentage of our data center compute revenue is using A100, it is well over half. It is just popular that not only on our current architecture end market but we will be using other of our architectures as they will be continued to be deployed both for hyperscale clouds but also for enterprises. It's very common for people to add on to their current clusters with some of our existing architectures of the path as well. But A100 adoption, it's been driven by new AI use cases, new workloads, folks that are upgrading from V100 installed base, possibly some of them, but we still offer that V100 as well. Those new use cases are some of the most important areas of AI today in terms of natural language processing and/or recommendation engines. Recommendation engines fuel everything from the hyperscales business models to consumer Internet companies as well as enterprises looking ways to use that. So the A100 is doing quite well.

Good to know. And Colette, how is the inference momentum with the addition of A30 going?

Our inference platforms are doing quite well. Keep in mind, we have inference platforms that are specifically used for inferencing. We started this out years ago with the introduction of our T4, Turing 4, that was made specifically for inferencing. And now we have come out with the A30. The A30 is an upgrade to our T4 and is also specifically used for inferencing. Inferencing so far has outgrown the use of overall CPUs. Moving to GPUs and GPU's performance both helps them with latency, cost and overall scaling considerations. In the early days, several years ago, when we introduced the idea of using GPUs, we were starting with an overall market size of 0. We have now seen such great adoption of our platforms for inferencing. Within Q2, for example, it makes up a very meaningful part of our revenue in data center compute. Additionally, the revenue is up 2x year-over-year. So we have great momentum going forward, and we have lots of tools to assist those that want to use GPUs for inferencing. Whether that be Triton -- Triton adoption has occurred through many of our CSPs, this is an inference server that simplifies the deployment of AI models at scales into production. It powers open source inference serving software, and it also leads teams to deploy AI models from any framework that is out there. Our TensorRT, for example, has been downloaded 2.5 million times across more than 27,000 companies. So A30 provides 4x the inferencing performance of the T4. It also includes that MIG technology that you can also find on A100, and it supports a broad range of AI inferencing and mainstream enterprise compute workloads. So the customer interest is high, and it will continue to help power the AI adoption almost across all industries.

Great. And are both the hyperscale and enterprise verticals expected to grow in second half? Or is it a universal strength?

With our guidance that we provided for Q3, we indicated that not only the data center acceleration that we saw in Q2 would also occur as we moved into Q3. So we're expecting that acceleration. The demand is accelerating as we expect strength quarter-over-quarter from both our hyperscales and our vertical industries, including a ramp of enterprise that are using the specific NVIDIA-certified systems that our OEMs have developed. We expect strong growth overall in compute quarter-over-quarter, and we have great momentum, as we've discussed, in terms of the inferencing in those workloads that have outgrown the CPUs.

Great. And Colette, I was reading a blog somewhere saying that artificial intelligence economy will be nearly 4x larger than the mobile economy that drove the likes of Apple and iOS. You guys are the pioneers in accelerated computing and have employed deep learning in multiple new end markets. And one of the latest end markets, at least from investor interest this year, is software. And NVIDIA talked about a software-subscription model for SuperPOD hardware and Omniverse Enterprise at Computex in May earlier this year. How is the stand-alone software licensing adoption progressing? And when can we start layering some of the sales from software in the future?

So we've made a great effort to date in terms of allowing as many as possible to easily adopt AI into their infrastructures and work. We started with the hyperscale. So hyperscales are very well known to build up their infrastructures alone and often write some of their own software. The GPU is well positioned, with its CUDA development platform being able to work with all of the frameworks for many of those hyperscales to get started. But we're talking about a new wave, a new wave in terms of adoption that will occur with our enterprise and our vertical industries. Now what is different about them versus the overall hyperscales, it is very common for them to buy OEM systems or buy what we are seeing right now as NVIDIA-certified OEM systems to begin their work on AI. What they are used to doing is using enterprise software, such as VMware, as their underpinnings of system software to help them manage all of their data center systems. We are coming to market now with stand-alone software to assist them in their AI deployment. NVIDIA AI enterprise is one of the firsts. Dozens right now of our automotive, education, finance, health care, manufacturing and technology companies, for example, are looking and reviewing this software for their deployment. It's very common for enterprises to put their mission-critical software on a software program with overall support and work for maintenance as well. This is what we're doing here. So very like how they have sold overall VMware to them, this would be another part of that system software that helps them manage their entire fleet servers but also allow them to stitch together the work of AI directly to their applications that they would have as well. So that's one piece of it. Additionally, we can talk about Omniverse. Omniverse is another overall software opportunity that we have for us, an opportunity that more than 500 different companies are evaluating. What does Omniverse allow you to do? Omniverse really takes charge of what we're seeing as these virtual environments, connecting multiple and multiple virtual environments. We'll have our own domains about where we will be in a virtual environment. But the ability to move from environment to an environment is what the key things that Omniverse would enable you to do, the software that will be licensed to each one of the overall users as the users price in order to create those virtual environments. What are these types of companies that may be evaluating it? They're companies such as BMW, Volvo, Lockheed Martin. Why are those important for them to look at it? These are companies that have a significant amount of design, collaboration and large engineering forces that work together, both on the creative side of the overall design that instead of having to be in person, they can just go to the virtual space to complete their overall design. We're quite excited about Omniverse, been working on it for many years. You see it in terms of some of the important parts of the metaverse that is also being talked about at the infrastructure level. So here is an example of where we can both support the overall hardware infrastructure but also the software that creates these virtual environments. Now each of these opportunities can be multibillion dollars in nature, but it will probably take us 1 to 2 years for them to be more meaningful, and we'll be able to provide metrics along the way that shows our tracking to our progress long term.

Great. And just staying on Omniverse because I've been getting a lot of questions on this area. About 50,000 individual creators have downloaded Omniverse since it's beta in December of last year, and you have 2 million-plus CUDA users. So does your strong position in CUDA help you proliferate Omniverse? Because it is going to take a lot of collaboration across multiple areas to put out the metaverse. Does the CUDA position help you in your efforts in Omniverse?

Yes, a really good question. To think about some of the platforms, whether it be software and/or hardware we bring to market and how we think about it, we are rarely thinking at the full end use case of how software and/or the hardware may be used. What comes to mind is much of our developers that are on CUDA, developers that are on many of the AI frameworks continue to find new use cases all the time. We see the benefits of Omniverse for the 20 million designers and engineers that are out there. But we are assured that also our CUDA overall developers will even add on to that more, add on to it more to think about how to stitch together a full metaverse, how to provide it to more than just designer engineers, provide it to folks that are just looking from more collaborative folks. It could be folks like you and I in the future. I don't know. Maybe we'll do this conference in our virtual environment, and I'll have my digital twin there for you to overall see as well. So there's great opportunities by just providing the opportunity and the overall programmability for folks to use Omniverse, CUDA and many of our other products together to find new use cases.

Great. And let's talk about the Grace CPU. You guys surprised us a bit this year in terms of your announcement on Grace CPU in 2023. How is the team progressing on the CPU development? And what has been the data center customer response on the combined multiyear GPU, DPU and CPU road map so far?

We had talked about our road map that we wanted to focus on the modern data centers of the future. The modern data centers are racks of GPUs, racks of CPUs and racks of DPUs. Now the Grace overall CPU that we have announced, our ARM CPU, is a very specific CPU. It a CPU that enables that connectivity between the CPU and GPU from the onstart of data into the process that we can make sure that we are monitoring acceleration throughout the full process with the combination of the CPU and GPU together. So it is a niche product that is focused on AI workloads, focused on high-performance computing type of workloads and may be very key in terms of supercomputing as well. So we're in the initial stages, and we think in the next couple of years, this will come to market to be our first overall data center CPU.

Great. Let's switch to gaming. Gaming sales have more than doubled since April of last year. You guided sales up slightly sequentially in the October quarter. How sustainable is the current boom in gaming as the environment recovers from pandemic?

So our gaming demand is strong, but our growth is continuing to be gated by supply. Channel inventory in the market is still low, and we continue to work on providing more supply into the channel to improve those scenarios. The universe of gamers continues to expand. That universe expands because gamers are becoming more than just one-to-one play within the game. It moves to becoming a very important social place for them to connect with their friends. But now it is a full entertainment sport. Entertainment that meets all different types of users, people that are actually playing the games, professional gamers, first-time gamers, people that are training others to be gamers, people that are broadcasting. And then, of course, there's those that are just watching others game. The ones that are watching, for example, 500 million eSports of an audience. There's more than 700 million livestreamers, PC gamers on Steam, a very popular platform, is up more than 20% year-on-year. So during this last 1.5 years or longer, it's been an important time for people to find that entertainment, and that entertainment was gaming and a very popular sport that all took place on. But as we all know, once a gamer always a gamer. But it's important to look at some of the other things that have also driven a lot of the adoption towards gaming, a strong and powerful market of laptop gaming as well, laptop gaming that infuses some of the best performance, but also in a thin and light offering. So we're seeing folks not only build their own PCs with the desktop but also may have a second opportunity or their only opportunity, which is a laptop. So we have a record number of Ampere-based designs with the OEM, and we are well positioned heading into the holiday season with Ampere. Now right now, why is gaming popular as well is we are using RTX across all of our notebooks as well as our high end in terms of desktop. What this means as folks can realize some of the best games that are out there in real-time ray tracing that is just starting in its second generation with RTX. We've got a great upgrade cycle ahead of us still. Ampere does make inroads into our installed base, but only 80% of our installed base still has not upgraded to RTX. So we've got a great continued opportunity of Ampere in front of us. We're getting ready for the holiday season. And as we turn the corner to next year, we're still purchasing supply for the long term to fuel this market.

Great. Colette, just on the RTX adoption, I was surprised to hear that number, 80%. It seems like fairly early innings in terms of RTX adoption. And I know you guys don't talk about your next-generation road maps and products. But do you factor into where the adoption is of a current platform before you decide to put out your next-generation platform?

So that is correct. RTX is an important movement right now that are moving people to both upgrade and/or entering into gaming. And so 20% or less are upgraded either to Ampere or are able to work on RTX types of cards. So you're right that we have a big opportunity in front of us. It's common that we may change into a new architecture and not have reached a large percentage of our installed base on the most current architecture. As you know, we're always working every day on new architecture as well as architecture next, next as well. So we're going to make sure that whenever that time comes, we'll be ready to go, but it won't change anything. Our plans are our plans. If the architecture is ready, we're happy to bring that to market.

Good. And then on the supply chain, things have been tight across the board. Gaming availability has improved since May but remains tight. When can gamers get their hands on the RTX cards?

Yes. Channel inventory is still low. The overall prices in the market are a little bit higher than, what I would say, manufactured suggested retail prices are. So we know we have a ways to go in providing more of that inventory. We're starting to see prices decrease and therefore make it more affordable and available to our gamers, but we're still working on that. We work with all of our supply chain partners and ecosystems to improve that availability. It is a full piece of work across all of these different firms, meaning when we think about what we sell, it's very rare that we actually just sell chips. We are selling full platform. So we have to think about all of the components. But we also need to think about the components that our partners and ecosystems must also procure to finish up those systems. So right now, we're working across the industry to make sure that we are all working together to get that inventory to the market and to gamers. We do our best to serve the overall needs of the gaming demand not only for gamers but, again, for our overall graphics as it relates to our professional visualization as well. So this entire process includes a lot of different companies working on providing more supply.

Great. And then you've said the demand is going to exceed supply through the end of this year on gaming. How is the team managing supply? And where are the shortages most pronounced? Is it across wafers or substrates?

That is correct. We do expect that we will not be able to serve all of the demand, but we're working on it through that. Now the way that we are working on it is if you've watched in terms of our procurement process, you'll see us more and more with longer-term commitments. Longer-term commitments across the stack in terms of board chips for components, for all of the different necessities that we need, including contract manufacturers that help build out using the supply that we have. So we're working on all of these different pieces right now in terms of procuring for a longer period of time. Given the size of the company, we have quite a bit of optionality. We do have 2 high-end fabs that we are working with. We're probably one of the only large companies that have such a strong dual fab process. We couldn't be more pleased in terms of our partnerships with both of the companies as that partnership and the long term of working together has really helped us through this process.

Great. Another question on crypto mining. You guys have done a much better job this time around in terms of your strategy of offering low hash rate cards as well as CMP as a separate category. Ethereum is moving to a proof-of-stake model from proof of work, and some investors are concerned this might result in mining demand decreasing substantially. Given that bulk of ethereum mining is being done by GPUs, do you think there is a risk of perhaps oversupply in secondhand GPUs?

Yes. First, let me first start with addressing what we see as the effectiveness of the low hash rate cards. We established low hash rate cards as a strategy to serve gamers with our GeForce cards and then provide CMP cards to our miners as a whole. This, we believe, has been quite effective. The low hash rate cards were new this time around, and it really enabled us to market effectively GeForce for just the overall gamers. The proof of stake has been rumored for quite some time, and we have included in our forecast for Q3 that we would have a minimal amount of CMP revenue going forward. So resale activity could increase, but we're actually seeing several reasons why we believe it is not significant risk to our overall gaming business. For one, gaming demand is so strong and the channel inventory is low. It doesn't matter what day of the week it is, but if there is some rumor out there that supply is showing up, the gamers are out there and standing in line to overall get them. Additionally, gamers want RTX. Not just that they want a gaming card, they want a RTX gaming card, and it would be difficult for miners at this point to buy a significant amount of those Ampere GPUs as they're just not sold in large quantities, thus very unlikely that they have the ability to resell them into the market. So the CMP cards in quantity were provided to the professional miners, and it is not desirable for gamers, and we probably would not see those in terms of resell either. So given the availability and the mining performance considerations, it makes more sense for miners to purchase the CMPs and our gamers to continue to look for RTX as it becomes available.

Great. Just moving on to auto. Colette, when is the inflection coming in auto? And how would you differentiate your auto approach from other semiconductor competitors like Mobileye and Qualcomm?

So there are so many different microprocessors inside of cars. But I think it's important to note the difference in terms of our strategy versus the strategy of others. Our strategy is to build an end-to-end AV platform and AV service. Soup to nuts, from data center to an in-car computer, from data to training, to simulation to on-road driving. Many of our competitors are really focused on building fixed function point solutions, which is very different from our overall strategy. So our platform is open. We work with partners in many different ways or in any other different way that they are looking for. So working whether or not they want the full end-to-end platform that they want the software or whether or not they want full options in terms of chips or things for the data center. We are enabling a $10 trillion transportation industry that will likely all become autonomous in the future. Our architecture and workflow from the cloud to the car is essential. Now we've talked about a pipeline moving forward. A pipeline that takes us out to 2027 of the agreements that we've already signed with so many important car manufacturers out there. That pipeline is $8 billion by 2027. Now we expect an inflection point of that pipeline in calendar year 2023 as folks are still working in terms of the development of these platforms to go into the fleets. These fleets can be passenger cars, robotaxis and trucking. Now from time to time, we'll continue to update this pipeline as we are seeing many take interest of one of our more important deals that we signed with Daimler. Daimler is using our end-to-end platform and is also sharing the software revenue that they receive in terms of keeping the autonomous software updated in terms of those passenger cars. So as these new deals get added, we'll continue to update our overall pipeline going forward as well.

Great. [Operator Instructions] Colette, on the Arm deal, you guys have made progress with some of the fabless semiconductor players like Marvell, MediaTek, Broadcom endorsing you guys. And you also acknowledged on the last earnings call that the regulatory discussions are taking a bit longer than initially expected. An agreement with SoftBank allows until the end of 2022 to complete the acquisition. Investors ask why NVIDIA still needs to acquire Arm.

Well, we don't have to acquire Arm. We want to acquire Arm. Arm is a great company. Arm gives us an opportunity to address a much larger opportunity. It also gives us an opportunity to assist Arm on many different markets and workloads that they would like to get into. We want to see if we can expand Arm's IP licensing opportunities that allows us to offer NVIDIA's technology in large end markets, including PCs, including mobile, but also allow us to help R&D road map at Arm and turbocharge that investment. They would like to build out something similar to what they did with mobile in the data center and within PCs. We know both of those markets well, and we would love to help them build out the ecosystem and the work that's going to be necessary. We have the ability to reach more and more developers, developers on CUDA or otherwise, to the more than 15 million developers that are available at Arm as well. So we know that we can strengthen their work in terms of the data center and the PCs. Now we believe regulators will also see and understand the value of what NVIDIA can bring to Arm to both the ecosystem and then how the customers can benefit from more competition and availability to those customers. So we will be able to update as soon as we have more information, but we're still working through the regulatory process.

Great. And a question on the financial model, on the gross margins. How do you view your gross margins on a long-term basis given push towards higher supply chain continuity, even higher supply chain cost as the foundry partners talk about raising pricing and just structurally higher inventory?

NVIDIA, as a whole, is not unaccustomed to some of the things that are occurring because our focus is on innovation and continuing to build better and better platforms to meet the customers' needs. Better and better platforms generally means some of the best and high-end types of solutions out there. So costs have been continuing to rise from the beginning of time for us. But today, we're able to offset most of that with the overall mix that we're doing where we continue to take those platforms and build stronger and stronger systems. With our mix, we're able to offset a good amount of those cost increases. Mix and software are probably the largest drivers of our margins both today as well as in the long term in terms of what we expect our margins to do. So we'll continue to keep an eye on it. But right now, these costs are nothing new to us.

Great. I've received a question from an investor on auto. Auto pipeline is expected to grow to $8 billion in 2027. Please break down whether this is mainly software or hardware or something combined.

Our pipeline of $8 billion for our auto autonomous solutions out to 2027 is a combination. It's both the combination of hardware systems that are inside the car as well as the software opportunity, system software and/or DRIVE software that we are adding for both passenger cars, robotaxis and trucking. So it includes both hardware and software.

Okay. And then a question on OpEx. NVIDIA spends quite a bit on OpEx and rightfully because the size of the markets you're after are very big. What is the split on OpEx on the software hiring or the programmers and application engineers versus the hardware engineers?

Yes, a great question. So our focus in terms of investments, the investments that you see within our operating expenses, a vast majority of that investment is new employees, new employees that we have welcomed through this pandemic, and many of them still continuing to work from home. Some of our top areas of focus in terms of hiring is both hardware and software. I'd say it's a race in terms of which one is hiring more. We are hiring software engineers and hardware engineers for so much of our solutions. Our solutions, if it was just hardware, would not benefit if we did not have all of the software solutions that we need. Keep in mind, remember, we have a unified architecture, an architecture that goes across all of our products from a hardware perspective, but that is also true in our software. So our software engineers are working in terms of backwards compatibility but, of course, creating new capabilities for so many of our customers. From hyperscales, to enterprises, to gamers, to researchers, all of them are really focused on software that will enable them. So both are very important hiring opportunities for us.

Great. We're almost out of time. Colette, thank you for being a keynote at our conference. Thank you so much.

Thank you again for having us. Appreciate it.