NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hey, good afternoon, everyone. Thank you for joining us again on PSS Friday series here. We have a pretty exciting lineup here. We've got Ian, who had some -- you guys see him there. And I can tell you, there's no better person to talk about AI and data center compute than Ian right now in this world. And I'm going to turn it -- with that, I want to turn it over to Simona. She's got a couple of statements she wants to make, and we'll go get started.

Thanks very much, Harsh. Before we kick off the Q&A with Ian, I just wanted to remind everyone that we may make forward-looking statements as part of today's conversations. So investors are advised to look at the reports that we file with the Securities and Exchange Commission for any of the risks and uncertainties that relate to our business. Back over to you, guys.

Thank you, Simona. So Ian, there -- I meant that very seriously. There's not a better person to talk about the trends in the data center business. You've been doing this for a while. You guys are at the top of the game when it comes to compute. Could you maybe talk about what you see as the primary role when you talk to your customers in data center today? And then perhaps what your customers are indicating or where that role is going to the next 5, 10, 15 years out?

Yes. I mean one of the main drivers right now in the conversation is now how to digitize or cloudify or accelerate different industries, whether it be in automotive, in retail, in social media, in all things that we do, engaging with the Internet or engaging again in retail. The biggest honest driver for that was the productization and commercialization and enablement of AI for those use cases. And it's very challenging to -- AI is a scary word. It was pioneered by the hyperscalers originally because of the fundamental research that had happened in order to activate it, and obviously those took management first. The thing that's trending now is the shift toward the commercialization of AI into the broader enterprise [indiscernible], whether it be deciding how products are bought or recommended to you. We've all experienced improved ways of doing conversational AI, how we're talking to the cloud and our devices. And that's also extending more broadly into the broader area of data science. Now with the era of now going from big data to big AI and big data science, we now have tools to actually understand our data much more intimately. And that's driving a whole new era of computing in our data centers, either in the cloud or on-prem. And that's the opportunity we're seeing and driving. NVIDIA, while we do make the accelerators for AI and data science, we're now expanding in the conversations about how to accelerate the entire data center. What does the data center of the future look like and needs to be built today to meet those incredibly intense computing demand to achieve those business insights. So that's why you see NVIDIA as a company moving to becoming more of a data center company for both AI and data science and innovating, not just at processor level or at just the core software level but of the data center as a whole.

Okay. So that's pretty interesting. That gears towards some of the recent moves that you guys have made in terms of bringing in some critical pieces. So functionality is rising each year, now more than ever with COVID. COVID just sort of put probably a pause on every aspect of the data center, I suppose. What functionality rising -- what kind of stresses does that touch? So if I had a -- if a company had a data center that's a year old, and all of a sudden, you have a COVID economy, what kind of stresses does that put on a data center that's even just slightly old? And then how are you guys helping your customers overcome those stresses?

Well, first off, the -- it's driving an accelerated path toward more and more AI applications. And one of the challenges about AI, it's amazing technology, is that people are discovering new ways of doing things and more intelligent networks, more intelligent AIs to give a better experience or to recommend better products when you click on things, and of course, increasing ad revenue and user experiences, which drives the economy of the cloud. The challenge is that AI is moving really quickly. If you look at state of the art just a few -- 2 years ago, 3 years ago, you have neural networks like ResNet-50, which allowed you to understand what's inside of an image. The newer neural networks, much have gone beyond just understanding images, to understanding speech and language. And those things are much more higher level understanding that they need to do. And the neural networks, the result are much more intelligent and much larger. It's the true human intelligence, to be able to understand the human language. Bugs, dogs, animals can recognize pictures. That's a very, actually simple brain function. Understanding language is a more intense one. As a result, those neural networks from like just a few years ago, are literally 3,000x more computationally intense. If you compare like a ResNet-50 neural network, what it takes to train a network like that, to training some of the modern natural NLP, natural language processing models like Megatron, which was recently published, with 3,000x. So that's what's driving a lot of the rapid innovation and rapid deployment of the latest possible technology to enable the data scientists to apply those networks to their particular problem, to their particular domain. Whether it be understanding what's published on a web page and doing summarization, or if it's an automated call center for refilling a prescription or asking for -- or understanding a financial call and understand the text of dialogue and it's happening, and being able to flag it across all of the different information. That's a natural language processing, and it's a much more intensive activity computationally, and that's what's driving a lot of innovation today. The -- both on the training side, we have to develop these networks. There are now tools where people can take existing networks like Megatron, like BERT and do transfer learning. So apply that free training network to their domain. It's getting much easier to apply AI technology, which is exciting. And so as a result, proliferating the need for AI data centers everywhere. There's a flip side of that, which is obviously the deployment side. You train and then you turn around and you deploy the AI. The neural networks are getting big enough that they are requiring to be accelerated. In the past, the early networks can still run on the legacy CPU-based data centers or CPU-based edge devices or points of presence. Today, with the new neural natural language models, they just can't deliver the real-time experience and run the neural network in real time. So we're seeing a big growth in accelerating at the edge, moving more of the -- with our edge accelerators, both at the edge of the data center, the points of presence and even in the embedded and telco use cases, where we want to put the acceleration as close as possible and by the lowest possible latency and the best possible experience for the end user.

So that's fascinating. Sounds like a lot more is still coming. It sounds like we've only gotten into the tip of what can happen with accelerated computing and AI. I want to talk about something you mentioned earlier, bringing and accelerating, sort of the entire data center. So you guys recently acquired Mellanox, which was the bigger one. I think there was a software company acquired as well. So we know that Mellanox is computing at the surface -- sorry, connectivity at the surface. But Jensen was giving, I think, a keynote for one of the events out of his kitchen and he talked about -- so I don't know if that's fascinating to you, he talked about maybe offloading some things to perhaps parts and pieces of the Mellanox product set. Could you maybe expand on this thought process? Is this what you're referring to as sort of hyper driving or sort of putting the turbo in the entire data center?

Well, the basic unit of compute is moving from a server to the entire data center. Now people today, when they think about building an AI data center, they don't think about just building servers, we'll make sure they have GPUs in them to get great acceleration. The AI today is moving from a single server, single -- with multiple GPUs in it, to training across the entire data center, where you train a neural network, allot for language processing, for example, or for video understanding to understand video content. You typically train across many, many, many nodes. Upwards of some of the largest neural networks we've trained have been up to over 2,000 nodes across the entire data center for a modern neuro network. The reason for that is, it's time to train. We can't -- these data sciences' work is very important. And these new neural networks, we need to complete in a certain amount of time. Otherwise, it's just not productive to their work. So today's data centers have to be multi node and multi mode capable of doing AI training. And that's one of the important abilities that Mellanox, with InfiniBand, brings to the market. We can now optimize the entire computing stack from the server design, from the GPU and its software stack, to the server design, to the interconnect to build a complete AI data center. And by having Mellanox part of NVIDIA, we can now accelerate that road map even faster. There's also a flip side to that, which is, of course, the edge in the deployment side. If I can take that neural net and deploy at the edge, my requirements are different. I need to make sure that I have the best hospital network latency. I'm often -- and I care a lot about security. I want to make sure that I have -- my models are encrypted, my data that I'm transferring is encrypted, but yet operating at full line speed. And the DPU, the data processing unit, or SmartNIC that Mellanox provides, can do all that line encryption. They can do a lot of the computing in the network where necessary, both in the training steps and the deployment steps, deployment needs, and provide the best possible encryption security for delivering -- for operating on this data, which is the end-user customer data, which obviously, it's critically important. And for the people who have developed the technology, they know -- they can trust that their models, their IP are safe and secure because they never leave the accelerator itself. These are different ways where we can accelerate and move the ball forward. Both for building an AI data center that isn't just a composition of a bunch of servers, but truly operates as one data center GPU holistically for training the network. And then all the way down to the edge, where we want to provide the best possible service and security for streaming the AI in production.

Wow. Fascinating, really fascinating. So it seems like a pretty critical -- it seems like there's just -- it's not 1 plus 1 equals to 2 plus some financial equation. It sounds like there's some pretty good technological advantage that you guys will have. In the minds of common folks -- and this is a question that I got a lot, believe it or not, like what's the difference between hyperscale today versus -- a data center for hyperscale versus an enterprise data center? And the perception is, well, the data centers, the guys like Google and Amazon are really savvy, really sophisticated, the ones by bigger companies like Walmart, et cetera, are pretty good, but not quite the same. Is this actually the case? And are you seeing companies like larger big retail organizations and others trying to push their data centers to be more like, more cutting-edge and therefore investing a lot of money into their data centers?

I mean, we're obviously seeing data center growth in both vectors, both in hyperscalers for their internal use cases, hyperscalers for their cloud services, infrastructure as a service, as well as the enterprise on-prem. And that's always an equation that people have to make for themselves, whether or not they need to rent or buy. And that's a TCO calculation. NVIDIA, we serve both markets and both -- we see growth in both areas because of the strong demand for accelerated computing in general. There is a difference between hyperscalers and enterprises. The hyperscalers obviously have a deep bench of technical talent. They invented many of the core technologies around AI that we all benefit from today, though their engagement -- our engagement model then is a little bit different. We do provide the basic capabilities and foundations for accelerating their workloads, and they can design the next generation now themselves. We're bringing it to market in a variety of ways, how we talk to our phones or how we see recommendations off their websites for products we want to buy or use or links we should put on. On the enterprise side, obviously, I don't have that as necessarily deep of a bench to develop foundational technologies. That's where we can -- NVIDIA can help, too. So we have obviously immense experience in ourselves of building AI technologies for our own use cases, which we use internally in developing our own products. The work we're doing for our self-driving car, we've built up a very strong ability to develop those, our own AIs for our own self driving vehicles, as well as the work we're doing in robotics, health care, smart cities. So what we've done is open those software stacks to the rest of the world, in fact, made them freely available. So they can -- and we have our Metropolis stack for Smart City, Clara for health care and Jarvis for conversational AI, Merlin for recommender systems. These are higher level stacks, SDKs, in some cases, complete solutions, where enterprises can take them freely to their particular use case, their particular modality, apply some transfer learning, maybe retrain that neural network to recognize the words that are relevant to them, like prescriptions you're calling in, or financial data, but they're starting from a known good and highly intelligent neural network and software stack that's ready to deploy, making it easy for enterprises to take advantage of AI and deploy it is our goal. And we've done that through our -- by providing prebuilt containers, pretrained models that are starting point for it. So while we are enabling AI across all of those channels and vectors, the engagement model for it is a little bit different. But we're the one AI company that's working with every AI company. And by doing that, we learn across all those industries, decide how we can help move the ball forward and then make those technologies freely available to help accelerate the whole AI and data science acceleration.

Fascinating. It sounds like you guys are able to cut down the time that a lot of these organizations would have just starting from scratch and doing the training. They can take a lot of what you provided, that moment's kind of not perfectly ready-made but close, and then get to it. I wanted to ask about edge computing. You brought it up earlier. So we hear that edge computing is going to end up being one of the big things down the line. It's going to have some level of intelligence. But then every time you get into a complex application, the edge device wants to talk back to the data center. So how do you guys see that evolution at NVIDIA? Do you think the compute will reside at the edge? Or do you think for any kind of semi complex application, you are pulled in into the data center again for analysis and trends, whatever else you want to do.

Well, as devices become more intelligent, they're going to do something at the edge. In some cases, there's really 3 things that matter. One is latency, how long does it take to ingest the data, make an intelligent decision and taking action as a result. And that can -- in some cases, doing the -- connecting backups to the cloud or back up to the data center, which may be hundreds of miles or states away, it just introduces too much latency that they can't get a real-time experience. The example might be online gaming. We need to have low ping times in order to have a good gaming experience. The same is true for AI. If I'm talking to my device and it takes hundreds of milliseconds or even 0.5 second to a second of just getting through the different layers of the Internet to connect back to the cloud, just to ask a question on my phone, that's a bad user experience and people aren't going to enjoy it. So latency is critically important. The second is cost. Certainly, if we can push more of the computation into the device and provide the right level of acceleration at the device level, we can save money by doing it where it's needed and not necessarily having to move into the servers and dealing with all the backhaul, including the networking cost for that much bandwidth. It would -- just might be prohibitive. The third, obviously, is data sovereignty. There are certain use cases where it's just -- it's too difficult to ensure data sovereignty, to go all the way back in the cloud. It's much more logical and simpler just to do the AI in the device. We see this a lot, obviously, in health care. Doing AI in the medical -- at the CT scanner or the MRI machine, and there's work being done actually in COVID right now, to understand, to process images. We never have to send that patient data into the cloud, where, of course, there's much more hit up considerations, and it's just simply do it in the device. So as a result, those 3 things are driving more -- and of course, the overall interest in applying AI technologies and the AI technologies are getting bigger and to deliver all those things, we have to do them cost effectively and do them with acceleration to move the latency requirements. So that's really one of the big drivers around our new EGX platform. We're building accelerators, which have both Mellanox networking and NVIDIA GPUs in a single accelerator. And working with an entire ecosystem of edge system partners, OEMs and elsewhere, to help bring edge acceleration right to where the data is being ingested and where the intelligence needs to happen. Those devices will always be in connection to the cloud as the AI makes decisions and sometimes make mistakes or gets corrected. We always want to make sure that we provide the information back to the data center so that we can build the next AI and build that virtuous cycle. And do that, of course, safely and appropriately.

Fascinating, fascinating stuff. And so I'm going to go into a topic that I know is very near and dear to you guys, conversational AI. Every time Jensen talks, he gets really excited when it comes to that topic. So what are some of the new and exciting things that are going on at NVIDIA to the extent that you can talk about it publicly?

Yes. Conversational AI is -- it's an incredibly rapidly growing field. In fact, there isn't an AI company out there that's isn't trying to beat the other guy and getting the better BERT or the better NLP model, NVIDIA as well, and we welcome all of it, obviously. The conversational AI is a great demonstration of where we are in artificial intelligence. Before, like I said, understanding what's in the picture, the simplest -- that is a basic intelligence test, but one that is fairly common actually in biology. And we'll -- every insect to dogs and cats, they all can see images and recognize what's in the picture. Understanding language, being able to take a web page and summarize it. Being able to ask a question of AI and get an answer back in a human and understandable and productive way. It's the real Turing test, and it requires a level of intelligence that obviously is far above and beyond. So -- and it's just revolutionizing the way we interact with computers and our kitchen counters and our cars. And it is -- so as a result, this is a very active area of investment, both in the research community and as well driving in the software and engineering side as well. For NVIDIA, we're trying to help bring this technology to all the world's enterprises. We invested in a platform called Jarvis, which is an end-to-end conversational AI stack that provides some of the basic foundational components of speech and conversational AI, from automatic speech recognition, ASR; to natural language processing, NLP; as well as text to speech, taking on a string of text and making a human-sounding voice out of it. One of the reasons why the voices we're hearing from computers, from our phones, our kitchen counters sounds so real, is because they're entirely generated from AI-based models. Some of those models are the most complex models, the most intensive models to compute and require some of our largest GPUs to run, even in real time. But the results are uncanny and quite convincing to get the right breath notes and tones. Beyond just the foundation of ASR, NLP and TTS, you also need to do higher order technologies, like having proper Q&A engines, having proper dialogue managers. So you can understand a thread of conversation, which is going to be unique to each domain, whether it be ordering prescriptions or asking for directions or looking for a restaurant. Dialogue management, it's another area where we can help the community move forward. And we actually provide early dialogue managers, working with partners who provide those commercially. The third is things like speaker diarization, understanding who's talking at one time. We've all been on Zoom chats where everyone is talking at the same time, and the text-to-speech engine gets pretty confused. Having good AI can help us there, too. You can understand that I'm talking versus you talking versus someone else talking, potentially in a room where multiple people are attending and be able to isolate it. This is also important in a car scenario. When are you actually talking to the car versus talking to the person next to you? And doing sensor fusion, where you're combining camera-based technology, looking at where is this person looking at and doing gaze detection, and combining it with a conversational AI engine to recognize when I should be paying attention and when is this a separate conversation. As you can tell, this is an area where computers truly become more human, and as a result, radically accelerate how quickly AI gets deployed and really, in use cases. That makes it a very exciting technology to work on.

That's amazing. That is fascinating. So I wanted to ask about a business question. NVIDIA wins a lot. We can see that in the numbers. When you guys win, what do you think is the reason or the combination of reasons that allow NVIDIA to win so very often and particularly in compute situation with the data center?

Yes. I mean it's important to recognize that AI, and more broadly accelerated computing, it's not just a GPU question. It's not just who has the better chip. You can build a chip that's fast. There's a lot of transistors in 7-nanometer. What makes accelerated computing successful is the entire stack and the ability for the ecosystem to get their applications, whether it be AI, data science like Spark or HPC applications like for your home or Amber, to be accelerated. And when we first started this journey, we've produced an incredibly fast processor that could accelerate all those workloads but actually had no users, had 0 applications. Because in order for them to be successful, they had to be accelerated at the software level, at the application level. And there are many ways to do that. We can do that at the lowest level by programming the GPU directly. You can do it at the SDK level. We have many core libraries that people can use. Or you can opt to stack further to domain-specific SDKs or toolkits like Jarvis I just talked about. Or all the way to the application level or the ISV or the application liability itself has, have already been accelerated. So in order to be successful in accelerated computing, you need to have invested in the entire stack. We need to work with the entire ecosystem. It's one of the reasons why NVIDIA has more software engineers than hardware engineers, is because what makes us successful is our 20 years of investment in the software platform, software ecosystem for accelerating computing wherever we can make a difference. And that's the most important part of the equation.