NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Before we get started, if you are a member of the press or media, please disconnect at this time. This is a restricted line. Any unauthorized party in this meeting or any unauthorized use of the information communicated in this meeting is subject to prosecution to the fullest extent of the law. Any unauthorized person, including the media that is on the line at this time, please disconnect. Please note, today's call is being recorded.

Perfect. Good afternoon, everybody. I'm Aaron Rakers, the IT hardware and semiconductor analyst here at Wells Fargo. Extremely pleased to host a conference call -- Keynote Conference Call, our virtual meeting with NVIDIA, Manuvir Das, who is the Head of NVIDIA's Enterprise Computing Business. And we'll get into discussions here. Before we get there, I think Simona Jankowski from the IR team wanted to read a quick disclaimer on forward-looking statements.

Yes. Thank you, Aaron, for hosting us. Just as a quick reminder, this presentation contains forward-looking statements, and investors are advised to read our reports filed with the SEC for information related to risks and uncertainties facing our business.

Perfect. Good deal. So Manuvir, again, thank you for joining us. There's a lot to delve into here with NVIDIA. The company has been on fire in the data center business. So it's a great opportunity to have you speak to some of the key trends. So -- but maybe to level set, enterprise computing, it's broad-based. It's increasingly more and more diverse. I think you and I have talked about the BlueField product in the past. But maybe to level set the discussion, just a real quick overview of your responsibilities within the company, and then we'll go in to the questions from there.

Yes. Thank you so much for having me, Aaron. And on behalf of NVIDIA, it's a pleasure to be here. At NVIDIA, we operate as one NVIDIA, one business. There are no people other than the CEO who think of themselves as owners of a particular business, right? So for myself and my peers who report to Jen-Hsun, we do 3 different things for him. One is we're each responsible for some of the products. In my case, I'm responsible for DGX, which is our flagship server for AI, as well as EGX, which is the more mainstream hardware that is built by our OEM partners with GPUs in them. And then I'm responsible on the software side for our GPU accelerated data science software and as well as our new push now for enterprise-grade software for AI that we're making available to enterprise customers and all sort of the middle way, if you will, of AI. So that's one way of thinking about my role. The second way is we serve, of course, different customer segments and the customer segment that I focus on is really enterprise customers, and in particular, when they're doing their work outside of the public clouds, a lot of them work in the public cloud. And then finally, perhaps most importantly for this call, we all think about different areas of NVIDIA strategy. And the part that I work with Jen-Hsun on is our push to democratize AI across enterprise companies, which means really the move from the thousands of enterprise companies that are using AI today to the hundreds of thousands that are actually out there, right? So that's really my mission.

Yes. That's a great overview and exactly the topics that we want to hit on here.

So AI adoption, it's becoming clear that we move well beyond the hype cycle, right? It's an important driver of businesses. It's the underpinning of productivity growth in many instances across different verticals. I think NVIDIA recently mentioned, I think, 25 -- more than 25,000 companies is using NVIDIA AI for inferencing globally. How do you think about what's next, the next steps of enterprise adoption? And then we'll go into some of the product things that you're involved in that really looks to democratize that adoption rate.

Yes, that's a great question, Aaron. And I'll just clarify that the statement we made about 25,000, that was for using NVIDIA AI technology in general for both training and entrants, right? And some of them use it in the cloud, some of them use it in their own buildings and so on, right? I think, Aaron, when we think about the inflection point we're at right now, we are at an inflection point, okay? And I can describe that to you in 3 different ways. One way of thinking about it is, the success we've had with AI to date has been companies in particular verticals where there was a clear use case very related to the business they're in. For example, I'm an online shopping site, and I need to recommend what's the next thing for somebody to buy, right? I'm a health care company, and I'm doing research on drug discovery, right? So very specific to their use case. We've had a lot of success there, and we'll see that adoption continue. I think the inflection now is on horizontal use cases. So for example, conversational AI is a way to create chatbots for customer service. And it doesn't matter what industry your company is in as long as you're interacting with customers, you can benefit from having an AI-powered chatbot to communicate with your customers, right? So I think that's the first inflection, which is from moving from these very vertical use cases to more horizontal use cases, right? So that's one. I think the second inflection is, if you think of the AI journey for a company, training comes first, inference comes next. Because training is the process by which you hire data scientists and researchers and you say, okay, take on AI and show me that we can produce models that will help our business. And then when that work is done, then the company gets to the hardest step of, I'm willing to put this into production to use in my business, and that's when inference is done. So in fact, if you look across the enterprise base, you will find a lot of companies today who have done the training stage but are very early in doing the inference and actually putting this into production. And so I think in the next few years, we think you'll see a dramatic change there as this flood of companies gets to the stage of now actually using the AI in production by deploying it and doing inference. So I think that's the second way of thinking about the inflection. And then I think the third way of thinking about the inflection, which is perhaps a little further out. Today, when we talk about an enterprise company using AI, they need to understand what AI is. They need to learn how to do AI, how to do training, how to do inference. But going forward, I think you'll see more companies adopt AI in a transparent manner. And what I mean by that is I have some application I used today from some vendor, right? Maybe it's a document editing platform, right, like Office 365 Or it's an ERP system. And the vendor just says to me, hey, there's a new version of this thing, guess what? It's now infused with AI. And so it's better because of that. A great example is if you're editing in word today, you start typing a sentence, and it will suggest a completion for the sentence. And that is using AI, right? So in this case, the company that's adopting it does not have to learn AI technology at all, they're just deploying the next version of an existing product that has been infused with AI. So it's that vendor that has learned about AI and worked with NVIDIA to adopt AI rather than the end customer, right? And that's really true democratization, right? Because that makes it so easy for everybody to adopt, right? So I think 3 different ways you can think about the journey forward for enterprise AI.

And that latter point, just kind of dovetail a question off of that, does that infusion of AI at the application layer, does that necessarily say, hey, that application layer is going to increasingly need a parallel compute, accelerated compute underneath of it? Or is that, hey, this application infused that AI can run on general-purpose processors as effectively as a historical application. I was just kind of curious if that where that sits?

And the key phrase that you used there was as effectively because, yes, entrants can certainly run on general-purpose servers with CPUs. But it is increasingly the case that the more effective way, both in terms of cost and performance and capability, to run the inference is on GPUs, not just the training, but the inference. And this is why we believe that more and more servers, regular servers and data centers will have GPUs in them. Because more and more applications running on the servers will be infused with AI and they will benefit significantly from having a GPU in the server. It's no different than saying, can you do some graphics work with just the built-in graphics on a CPU versus then having a discrete GPU, perhaps, yes, in some cases, you can, but it's way more effective if you have a GPU, right? And it's no different here for the applications infused with AI. We already see now that they benefit greatly from having a GPU in the server. And this trend is only going to get more dramatic, Aaron, because the AI models that are being used within the application are becoming larger and larger and more complicated. And that's, again, where the GPU really makes a big difference.

Yes. So one of the questions I often come back to is that NVIDIA, I think, this last earnings call, Jen-Hsun had mentioned 10% of servers deployed globally today, incorporated GPU. I'm going to kind of tie 2 questions together. So the comment was also made that over time, a majority of servers could deploy a GPU. But when I talk to enterprises, ourselves being one of them, it doesn't seem like the traditional on-premise enterprise is going to have a big GPU attach rate. So what are we missing there, right? AI workloads running cloud? Do they run on-premise data center? I'm just -- how do you guys see that?

Yes, it's a great question and a great one to unpack. So number one, it is absolutely the case today that if you just look at how many servers are shipped, right, new servers that are shipped and going into enterprise data centers, servers with GPUs in them are in the single-digit percentages, right? That is the case today. And we expect that, that will change dramatically. The reason why it is the way it is today is because there has been a general [ mental ] model that you need a special kind of server to do AI, a server that you pack it to the gills with GPUs. You put 4 GPUs, 8 GPUs in the server, it's a bespoke thing. It's not a server that costs $10,000. It's a server that costs $100,000 or $200,000. And it's like a Ferrari. And when you buy one of those, you better be running AI on a 24/7 or you wasted your money, right? That's been the general model, and that's why the penetration has been what it is. The shift that we are working on now with our customers is take a general purpose server, a server you would put inside a VMware farm, a server you would rack and stack into your data center as a multipurpose server that you can run SAP on it, you can put a database on it, what have you. You take that $10,000, you put one GPU in it that costs a few thousand dollars. And now you've got a server that on the one hand, can run your traditional workloads. And when you're only using for that, there's no regrets because it's a modestly priced server. But on the other hand, it can also run your accelerated workloads like AI, right? So it's a multipurpose server. And when that begins to happen, we expect it will drive a significant shift in this percentage, right? Because it's the mainstream volume servers that will begin to ship with GPUs in them. And as Jen-Hsun said, we fully expect that in the fullness of time, this will become the standard. And so if you look enough years out, the majority of servers will have GPUs in them. All of them will have GPUs in them, but it is a journey, right? And so -- but I think this is a fundamental point that the difference between where we are at now and where we're headed is that today, the AI servers with GPUs in them are bespoke. They've put in a corner of the data center, they serve one purpose only, whereas the servers we're working with OEMs are now along with the software we've built, these are multipurpose, the new workhorse of your data center. The new server that you're going to rack and stack into every new footprint that you build so that your footprint can be future proof. And the reason is because more and more of the workloads that you run in your data center are going to require a GPU because they'll be infused with AI and they'll need that acceleration.

Yes. No, that's an interesting point, right, because I think myself and a lot of investors I talked to look at it. And to your point, right, there's 8 GPUs on average in terms of training. There's 4 maybe for inferencing. But what we're really talking about is just somewhat elasticity of bringing price points down to drive, coupled with the application ecosystem evolving with more AI infusion.

But That's absolutely right, Aaron. It's elasticity in the price point and it's elasticity in the usage point. This is why we did the work with VMware, right? You already have a farm of servers that are administered via VMware that are being shared out for different users and different workloads in your data center. And AI should just be thought of as another such workload. Your data scientists should just be thought of as another set of users who can come and access the same pool and your IT team should be able to just expand the pool naturally to these use cases, right? And that's why we did the work with VMware.

So you've driven a bunch of other questions that I could add on there, but I'm going to kind of try and keep with my script here a little bit. So how does that evolve? Or how does that entail what NVIDIA is doing as far as a full stack strategy company, right? You're responsible for DGX, EGX, as you mentioned, do you evolve down that path where in those environments where elasticity starts to drive more attach rate of GPUs? Do you want to be that full -- are you that full stack provider in that type of environment ecosystem?

Yes. Yes, absolutely. And I just want to be -- I'll be careful in answering your question because I also don't want to miscommunicate here, right?

Yes.

Really, the people who have done the best work in creating hyperscale, multi-tenant compute environments are the public clouds, right? And they've done a fantastic job of that, and they've really changed computing quite significantly. We have very deep partnerships with all of them. We work very closely with all of them. In that case, we are providing them hardware and software that they incorporate into their clouds, right, to provide a multi-tenant hyperscale environment. That said, the modern data center, regardless of whether it's an enterprise data center or a colo or a cloud, is going to look more and more like that. What you're trying to do is you have a shared environment of compute storage and networking, that is used for multiple applications. That's just how the industry has evolved, right? And NVIDIA is very committed to being a provider of that full stack, okay? And that means the hardware, the networking, the middleware software, the operating system of this infrastructure, if you will, as well as the AI software that you can run on top of that, right? We are absolutely working on all those fronts. We're a full-stack company. Now we are a platform company. And so what that means is we are happy to interact with people at any level. You can acquire just a hardware and build everything else on top. You can acquire our middleware software and do your own AI software on top. You can acquire our AI software and just go from there, right? And we're happy to work with you at any level. But we certainly have a viewpoint on what the full stack should look like. We are working on the full stack, and we make it available in different ways. For example, we announced a program this year where we're working with Equinix. And we are predeploying that full stack already for customers in Equinix data centers. So they can just choose to consume it there, right? Or they can learn from that and build their own. So -- and the reason we do this, Aaron, it's important because it is a full stack problem. Accelerated computing is not easy. You have to optimize the different parts of the stack together in order to get the best benefit. I'll give you a very small example of that, right? You think about AI and you produce these models and models can be large. And so you have to split the model up when you actually do the inference. Now you have to think about what are the hardware capabilities of each of my servers. So I can be smart about, in my software, how to split up the model. So you cannot do this in isolation, one without the other, right? And that's why we've done all this work together.

Yes, that's a great overview. And you touched on that a little bit with the AI LaunchPad product with Equinix. But the other thing that you've done is the -- I think the tagline was democratization of AI when you announced this product, which is the AI enterprise software suite. And could you help us appreciate what that is? How that's -- I believe it's a subscription license, how we think about that is democratizing AI and ultimately representing a pretty attractive recurrable software revenue stream for the company?

Yes. We're very excited about NVIDIA AI Enterprise. Of course, there's a revenue stream that we foresee that we are very excited about. But I think more than that, we're excited about what it means for enterprise companies, right? So I put it to you this way, right, that, what are the barriers to AI adoption today for enterprise companies, right? I think it starts from there, okay? The first barrier is that AI software is complex and difficult to create. There's a huge ecosystem today of start-ups to provide a piece here, a piece there, et cetera, but it's largely a DIY effort to cobble software together to do AI. The reason we created NVIDIA AI Enterprise was we said we, as NVIDIA, have actually built most of the software. We've spent years building the software. We've matured the software. So now we can package it all together and say to an enterprise company, just procure this piece of software, it has everything you need. It's enterprise grade. It has a support SLA. It's like Windows. It's like VMware vSphere. It's a [ bat ] form. It's the operating system of AI that you can go deploy and all your servers in your data center, right? So that was the first point of it to make it easier to adopt AI. The second point of it was, if you think of the dilemma or gap that you have with AI for enterprises today is there's 2 personas. There's the data scientist, right? The data scientist is doing the research, creating new models, Jupiter notebooks, all of that. Every night, they tweak the data a little bit. They get a better model, they like this model, right, more than the one they had yesterday. On the other hand, you've got IT. IT of the people who control the data centers. More importantly, IT are the people who are accountable for the software that is running in production that the business depends on, right? Where if something happens to that software tomorrow, whether it's compromised or not functioning, the Board wants to know what happened, right? My company's reputation has suffered, right? So you have these 2 distinct entities. And in the early adoption of AI, the data scientists have essentially worked around IT, right? They've either done their work in the cloud or they've built their own environments and gone forward. So you have to break this gap somehow to truly democratize AI. And so we built NVIDIA AI enterprise in this way where we said, if your enterprise IT and you are in the business of creating a farm of compute for your workloads and your users, then NVIDIA AI enterprise is the thing that you add to your farm so that your compute farm is ready to do AI. And then you can go to your data scientists and researchers and say, come here and do your AI work here on this compute farm because we'll be able to provide that deal. There will be a center of excellence where everybody can come and get their resources. And we'll have one place where we're working together so we can take the results of your work and put it into production, okay? And so that's the work we did with VMware, where basically we integrated NVIDIA AI Enterprise with VMware vSphere. So any IT admin who's familiar with VMware today. And they say, okay, I'll create templates of different kinds to spin of VMs in my VM farm, they can very easily create templates and they can stay to a data scientist, here's a Jupiter notebook template, come and use this one and now you're good to go, right? So that's the second reason. I'll give you a great example of this, Aaron. We've been -- NVIDIA AI Enterprise has been available generally for a couple of months now in general availability. Obviously, we've had our initial traction and successes. There's one where I will highlight to you without the name where, to me, it was the canonical example of the shift that we were expecting that we saw. There was an IT team. They read about NVIDIA AI Enterprise. They contacted us. They got on to launch pad. They had the experience of setting up an AI environment, okay? They were happy with that. They set it up. They went to their data science team, and they said, "Look, we set up an environment for you. Can you come try it out and tell us if it's good enough for you?" So they brought the data scientists, they used it, they spun our Jupiter notebooks, et cetera. They said, this is great. This is exactly what I want. And the IT team came back to us and did a procurement, right, for the software. So now they are in the loop. And they are not outside of the AI adoption, they're now actually part of it, right? And they are serving the AI to their data scientists. So that's kind of our expectation here.

Yes. Remind me just real briefly how it's licensed?

Yes. So we chose to license based on the way the VMware licenses vSphere because the product as we've done it today is based on top of VMware vSphere. So it's the same licensing model. Roughly speaking, it's about $2,000 per year per [ superior new socket ], okay, is the -- that's the model which for those who are familiar with VMware, will say that actually looks very much like the pricing for VMware vSphere. And it's also -- it's modeled the same way in terms of support period and SLA and all of that. So that the -- from the point of view of procurement, it feels just like a natural extension, right? That's a simple way of thinking about it. It's about $2,000 per CPU socket per year on a subscription basis. Yes.

That's perfect. A couple of other topics and a little bit of time we have left, I want to make sure I touch on is that. I've written a lot about it. I still feel like investors aren't fully either understanding or appreciative of another layer of acceleration. This idea of disaggregation, you and I have talked about BlueField in the past. So remind us again, how you see BlueField data processing units evolve? When we might see an inflection? And I think tied to that, Project Monterey with VMware, the importance of that as we think about things over the next couple of quarters a year or so?

Yes. And I'll put it to you this way, right? The modern data center, as we discussed, is multi-tenant, right? You have multiple people and workloads sharing the same equipment. And so the world -- the infrastructure is becoming zero trust, which means that you can't just put firewalls on the outside and let everybody partly inside. Every application, every user, every workload must be thought of as an adversary with respect to everything else that's running in the data center, right? And it turns out that in order to properly support a zero trust environment, you need the capability of a DPU. That's really the only feasible way, tractable way of implementing a zero trust environment. Because the amount of information going across the network that you have to process and analyze and manage efficiently to create the secure environment requires those capabilities of having actual processing on the networking interface, right? So we believe that regardless of whether you're a public cloud or a colo or enterprise data center, the trend is very clear, every server will have to have a DPU in it in order to be part of a secure zero trust environment, right? That's why we are so keen on the DPU. That's why you'll notice, Aaron, that when we talk about our DPU BlueField, we do not talk about BlueField-2 which is our current BlueField. We talk about BlueField-2 and BlueField-3 and BlueField-4 because there is a journey here, right? This is an inflection that the whole industry is going to go through. It's going to take multiple generations of the hardware and software for that journey to occur, right? And furthermore, there's going to be a lot of different players in this ecosystem, cybersecurity providers, virtualization platform providers, et cetera. And this is why we took the same approach with the DPU that we took with the GPU, which is we created an SDK, it's called DOCA. And the idea is that all of the software developers from these different areas have one way to program to the DPU to take advantage of it, right? Because otherwise, you'd have a piece of silicon that does one thing and you wouldn't really be able to support all the use cases, right? So I think what I was trying to convey there is why do we believe the DPU is so important, right? Because we think it will be in every server. And it has to be in every server because the industry is shifting to multi-tenant environments, right? It's going to be too expensive going forward to have siloed environments. Like this is my footprint that does one thing, and that's my footprint that does another thing, right? That's going to be too expensive. That's going to be too siloed, that's going to be too inflexible. And the cloud has shown that, right? This is the beautiful thing about the cloud that if you have one pool of infrastructure that can be used in multiple ways, that's the winning proposition, right? And so that's what...

I think to your point, the evolution of this disaggregation accelerated architecture, isolation evolution, if you will. The road map, as you mentioned, BlueField-2, 3, 4 eventually involves actually some convergence of GPU and DPU together. Is that -- should I think about that theoretically as that being maybe the inflection of really driving this elasticity where these things converge? Or do you think it maybe happens quicker than that?

Yes. I think it happens quicker than that. So that's not the thing. I'll come back to that. I think you talked about VMware Project Monterey, right?

Yes.

Again, that in itself is the way to think about the inflection. What is happening with VMware Project Monterey is if your server has a DPU inside it, then a lot of this work of enforcing zero trust and security, et cetera, is now put on to the DPU. So you can actually have these servers run in this multi-tenant fashion, whereas the CPU is freed up to actually run the workload, right? What's happened today is that more and more of this functionality has gone into the software and it has to run on the same CPU that the workloads need to run on, right? So we move it to the DPU, free up the CPU and more importantly, the DPU is built in such a way that it does that work much more efficiently than the CPU does, right? Otherwise, you're just shifting the problem. So we're very keen on moderate, Aaron, because I think that's what's going to drive the inflection. The next stage of it, as you said, what we are uniquely doing at NVIDIA with the addition of GPU and AI capability to the DPU just takes things to another level because now you can use AI to understand what is happening in the network, directly on the network, right? And so...

Okay. So it's not necessarily that convergence does away with the GPU as a discreet component?

No, no, not at all. Not at all.

Yes. Okay. That's perfect. That's helpful. The final -- in 2 minutes, and I know it's 2 minutes probably way too short to talk about this. But Omniverse Enterprise. I think any call that you've probably been on video recently is metaverse and how you plan to that. For your responsibilities, what does that mean in a brief 2-minute comment?

I think for me, as I mentioned, we're interested in providing enterprise-grade software to enterprise companies, right? And we have an offering with Omniverse. It's called Omniverse Enterprise. It's the -- in general availability, we've seen great adoption for this. Super excited because really, Omniverse is a way for people to collaborate when they're doing 3D design work, right? And the number of people who are looking to collaborate in this way is massive. There's 40 million people out there who need this kind of collaboration, right? And so we've created a Omniverse enterprise. It has a simple licensing model, again, based on subscription you can think of it in terms of thousands of dollars per such person per year, right? We have more detailed information, of course, in our documents, but you can build some simple math in your head of the opportunity here. We're also working on something called Omniversed Avatar for enterprise companies, which is it's not just about the creators and users, but it's about the work the company does where you have a digital form of it within Omniverse and that is its own journey and opportunity both financially and what it means for companies, right? So what is encouraging to us, Aaron, is this. We've worked for many years to produce Omniverse right? And what we found in the year that we've really now pushed it forward and brought together Omniverse Enterprise, the traction has been amazing. The appetite for this thing has been quite amazing. So it's another one of these things where I can sit and say to you, in theory, conceptually, the hypothesis is everybody is going to need it. It's going to change every enterprise company, but I can also say to you that in terms of the traction we've seen too far, right, it certainly seems to match that hypothesis, right? We see great appetite and interest in Omniverse Enterprise.

I can keep going on, but I think we're unfortunately out of time. Manuvir, I appreciate you helping us understand the story a little bit better. And thank you so much.

It was my pleasure. Again, on behalf of NVIDIA, thank you so much for giving us the opportunity to have this discussion.

Perfect. Have a great day.

Yes, you too.