Dell Technologies Inc. (DELL) Earnings Call Transcript & Summary

Good day, and welcome to the Dell Technologies "Ask the Experts Anything" Call. Today's conference is being recorded. At this time, I would like to turn the conference over to Paul Franz. Please go ahead, sir.

Hello, everyone, and welcome to today's "Ask the Experts" Technology Q&A. We wanted to spend the next 45 minutes outside our typical financial cadence and focused purely on technology. The major technology trends we're following, our strategy and how we're innovating to support our customers. We plan for this to be the first of several conversations with our Dell leaders to stay tuned for future sessions. Today with me are Jeff Clarke, our Vice Chairman and Co-Chief Operating Officer; John Roese, our Global Chief Technology Officer; and Jeff Boudreau, President of our Infrastructure Solutions Group. I'll start by sharing our safe harbor statement. Dell Technologies statements related to future results and events are forward-looking statements based on the company's current expectations. Actual results and events could differ materially due to a number of risks and uncertainties, including those disclosed in our SEC filings. Dell Technologies assumes no obligation to update its forward-looking statements. Before we go to Q&A, I'll turn it over to Jeff to share a few thoughts.

Thank you, Paul. Happy to be here to talk about technology with you all. We can talk about a number of technology trends reshaping the landscape like data continues to explode and will define our world, multi-cloud by design, taking the right cloud for workloads, cost optimization, complexity reduction and operational ease. Zero Trust is the needed security shift that puts the good guys back in control. Compute and storage resources are moving closer to where the data is created and modernization, virtualization and containerization of the telecom stack. So I suspect most of our time will be spent talking about the most talked about technology today, generative AI and rightly so. Gen AI is an inflection. It's disruptive and it's game changing. What an exciting opportunity. Let's widen the aperture for a moment. There are several types of AI, including machine learning, deep learning, computer vision and generative AI, all are growing with new and exciting use cases like computer vision for manufacturing lines, machine learning for supply chain and chatbots that enable greater customer satisfaction. Today's $30 billion AI TAM is primarily driven by the first three. We view Gen AI as a new category of computing in the technology stack has very distinct characteristics. Gen AI doesn't replace anything, and it grows the IT TAM over time. It expands the use of computational machines to open up new ways of automation. The compute will be fed by massive unstructured and object storage infrastructure driven by more and more data. And the models created by Gen AI will need to be protected as some of the most valuable data ever created. Generative AI powered by LMMs, large language models, has dominated the AI narrative. These are amazing models generalized to answer any question. They are trained over very expansive data sets. The data set can be a one form like text or many forms text, pictures, audio source code, et cetera. The utility of these models with hundreds of billions of parameters is wide in general. They require massive scale, immense computational density to train them. From our customer discussions over the last 6 months, four generative AI use cases have consistently risen to the top of [CEO's] list. Customer operations, content creation and management, software development and sales. What we are seeing is customers wanting to use their data, processes and business context to train the model. As a result, we see generative AI having two distinct variants. The first being massive, multifunctional generalized models, GPT4 [bar], ChatGPT as examples and domain-specific like [ Galaga ], PubMed GPT or Dramatron or enterprise-specific models like stable diffusion, Bloomberg GPT and Code Gen. What's the difference you might ask? First, the core neural network-based language model is essentially the same. The difference lies in the number of parameters and the number of expert systems focused on unique functionality like common sense reasoning, translations, pattern recognition, reading comprehension and code completion. Massive multifunctional generalized models have many more parameters and expert systems. They are trained on a broad and large volumes of data. Domain and enterprise-specific Gen AI use smaller proprietary data sets with fewer parameters and expert systems. These models don't require a massive scale or cost. And in most cases, they can be trained up with a small cluster of servers with GPUs and perform inference on the edge with a single server. For example, an enterprise-specific model built with open source [ Falcon ] with 7 billion parameters, on your data can be trained with 4 Dell PowerEdge XC9680s. And once the model is trained, [inference] can be performed on a single Dell PowerEdge 760 XA. To use [Jensen's] characterization, we will have AI factories at the Edge, in the data center and in the cloud. In other words, everywhere, resulting in a range of Gen AI infrastructure solutions. Dell is uniquely positioned with our broad portfolio and services to win in these new categories of computing by helping customer size, characterize and build the Gen AI solution that meets their performance, cost and security requirements. We are very early in the cycle, not dissimilar than 24 years ago with server virtualization or as far back as 40 years ago with the introduction of the PC, disruptive, game-changing, more to learn, more to come. With that, we'll take your questions.

Thanks, Jeff. We'll ask each participant to ask one question to allow us to get to as many people as possible. Let's go with the first question.

[Operator Instructions] And our first question is going to come from Wamsi Mohan, please go ahead.

I guess the question I have is when you think about the technology differences up from Gen AI, it sounded like, Jeff, you're saying that TAM is all incremental. Can you talk about how you would characterize server configuration for Gen AI versus a typical industry standard server that you sell? Both in technology terms and in ASP terms. And I'd also be curious if you think about how the evolution of the interconnect is going to be, for now, seems like [ InfiniBand ] is fairly dominant. Do you have a view on whether that stays dominant or if it -- Ethernet starts to get incremental traction as well?

Thanks, Wamsi. Let me take a run at it, and John and Jeff can certainly fill any holes. I mean, first of all, we believe that TAM is expansive. This is game-changing and disruptive, and we just think it's an opportunity with what this enables. I called it in an inflection point in my talking point, This is really a fundamental change. From how we develop products to how work is done. And with that, we just believe this is incredibly -- an incredible opportunity for companies to reinvent themselves. When I think about specifically your question, what does the average configuration look like? There's actually a correlation of model size to how much memory you need. In your model [indiscernible] memory. And we've modeled that characterized to have a sizing capability to help our customers through this. But if you're running a $7 billion parameter model, 20 billion parameter model, you can almost double the amount of do twice as much memory to be able to run the model effectively. When you look at the ability to run the actual computational side and the configuration today with our 9680 with its H100s are clustered together for groupings of 8, allow you to extend the ASP of our products quite substantially. We think AI drives richer configuration. By the way, all the way down to the PC, doing more through PC drives a richer configuration. So we don't see any scenario that running these algorithms, training these algorithms, using your data to be able to build your models and then ultimately do conference out wherever you deploy it doesn't drive richer configurations and ultimately richer PCs. In terms of ASPs, they're significantly greater than the average server, just the content of memory and the content of the compute intensity drives that. Jeff, John, anything you would add?

No, Jeff. I think you're characterizing the TAM, it's just broad in general. I mean, I'm going to just add on the second half of that question, if you want, on interconnect. We'd like to talk about compute and we definitely like to talk about storage as components. Obviously, the network broadly in the AI world also gets disrupted to some degree. When you think about a training infrastructure, it's not a singular node. It's a cluster of nodes we're seeing in data centers, both InfiniBand and Ethernet. That's an ongoing -- two technologies competing with each other healthily, and we're seeing them both advance and we play both sides of that equation. And the reality is we will continue to see demand as computing performance goes up for a node, the I/O path to and from those computing nodes within the cluster will go up. Another dimension of the interconnect though, is the storage interconnect because it's not just the nodes in the neural net talking to each other, it's how fast you can feed the data into those environments. So this is an area where high-performance storage architectures that we happen to be quite good at are important, having gigabit links from the storage environment into the training cluster is as important as the train cluster itself being able to shuffle data between nodes. And then kind of a third dimension of interconnect is the IO path into the inferencing infrastructure, which that has profound impact because if we're talking about processing imagery or media, the IO path from the sensor, a camera or otherwise across a 5G network or across an SD-WAN or even within an enterprise environment, also it gets essentially higher utilization. So in general, storage, compute and networking all work in concert to build out these next-gen AI architectures and from our perspective, there isn't necessarily a particular winner around just Ethernet or InfiniBand, there's going to be competition there. On the storage side, [ Fiber Channel ] still plays a role along with InfiniBand and Ethernet connectivity. And again, what matters is how performant your storage systems are so they can actually fill those pipes. And then thirdly, broadly in the inferencing, it's going to require relatively high performance from any place where data is created through the place of inference which means that things like our Edge servers, if you didn't notice, are generally coming with a default higher performance [net], higher-performance Ethernet interfaces. And the reason for that is to be able to basically digest the stream of data that has to be inferenced at the Edge. So interconnect is a big deal. It's probably secondary to compute storage, but one of the three pillars.

And our next question will come from Simon Leopold from Raymond James.

Great. I guess I want to maybe get your perspective on how you expect the market to evolve and in that I'm imagining that much of the spending is biased towards hyperscalers building infrastructure and training AI clusters today. And that over time, we see essentially the uptake by enterprises, your customers to do more inferencing and to leverage their own data to utilize the systems being built by hyperscalers. And I guess I want to understand whether you think that's a rational way to look at the market or you believe the enterprises will build their own clusters. That's part one. And part two is really if we're right about enterprise adoption coming later, how do you envision that evolution or timeline of enterprise spending on AI implementation?

We'll see if I can unpack that Simon, there's a bit there. I'll take a swing out of here and see if I can answer your question specifically. I mean, for me, the big step back is this notion of inflection that I used earlier. And how game-changing this is and how it's really going to impact every sector and every type of function in every sector. And because of that, we're seeing CEOs of every size company have discussions of how to take advantage of Gen AI. And the reason for that, I think, are reasonably evident growth, it's still on the same page. The first is it's going to change the way our work is done. Two, it's going to change the way we build products. Three, it's going to change the way we service customers. And probably, four, I think it's obvious, but I think it's worth saying there's a productivity and efficiency element of this. And it's why I believe it's fundamentally additive to our industry. I've been at this -- I think that's [indiscernible] a long time, as you know, and how many times has the world dropped a 10%, 15%, 20% productivity improvement upon us. To my knowledge, that hasn't happened in my working lifetime. And that's what's here in front of us. And it's why it's the topic of every CEO and every Board. And because of that, companies are trying to understand how to deploy the technology, how to better understand it, how to use it on their business context, how to apply it to their business models? I don't think that can be done in the broad general foundational models. Is there a lot of great work to be done, is there a lot of applicability without question, we'll continue to see that. But when we want to start talking about your product development community, how you address your specific customers, how to use your marketing data to serve your customers better, how to use your telemetry data to serve and they build a better service model, we believe that type of information is proprietary, unique to the business. And as a result, we see enterprises building out that capability. Today, we have lots of pilots underway with medium-sized companies, large companies and enterprise scale companies. We see that continuing. What they're asking from us is help them size it, characterize it, get their data prepared and help them operationalize and build these AI factories that we refer to, that's the opportunity. How it's staging today, look, we're seeing a lot of work in the hyperscalers, the massive scale of the models that I described require that sort of scale. But that's not where we see the vast majority of the models being built and whether they're being deployed. John, Jeff, anything you would add to that?

So for me, I guess, also other things that we've talked about in the past [where] our customers are telling us, I would say there's concerns around data privacy and data security in regards to some barriers of everything being in a hyperscaler cloud over time and why they need to bring [ on-prem ] or in their environment. So Jeff talked about the different use cases and the attributes and the resources that you need to serve those things from LLMs, yes, along those at the hyperscaling [now] because they need massive scale. As you kind of lean into some of these domain-specific areas of the [inferencing], you need a lawless resources to go do that. And customers are looking at it because of some of the data privacy, the security. Jeff and I talked to you before, I did a lot of [indiscernible] around physics and latency and bandwidth matter and the need for real time and [ near time ] speed is all critical, which actually means into having infrastructure [ on-prem ] for the enterprise use cases. That's a big thing that we hear from our customers as well as we go forward.

Well, I think the other thing, John, you're a resident expert in this and Simon, maybe this helps build the case, we see the algorithms evolving with smaller data sets that are really honed for specific enterprises and specific, what we call domain and process-specific knowledge. And the fact that you're seeing decentralized AI occurring, which is there's a place where the model is trained, there's a place where the model is [indiscernible] and there's where inference is done. Back in the day, that was all the same system, but we're seeing decentralized architectures accelerated. John, if I misspoke, you should correct me.

No, no. I think it's important, we're getting in this moment in time where we see Gen AI as if it is the first time we've played with AI. And just to remind everybody, we were about to enter what many of us called wave 3 AI algorithms. The first-generation were basic machine learning algorithms, then we moved into the neural networks and deep neural network environments. And these -- what happened between Wave 1 and Wave 2 is things got bigger, meaning we saw this trend where you need a lot of compute capacity to build out a DNN and that was nice, but then the industry started to shift. We realized, hey, to make these commercially viable, we created technologies like transfer learning, which allowed you to take an existing model and only retrain one or two layers of it in the neural [indiscernible]. We did things like Federated Learning and that allowed us to distribute the learning process across all the way out to Edge nodes. Anyway, all that -- that was the narrative by the way, the day before the Gen AI world took off. It was basically a shift to rules engines, transfer learning, all kinds of things that would make the system more efficient and easier to deploy over a larger set of topologies, meaning small entities, et cetera, even out to a PC. Gen AI popped up and interesting enough, based on the nature of the first generation of it, things like GPT3, GPT4, ChatGPT Bar, et cetera, the problems that we're going after were generalized applications. They are very large scale. And quite frankly, everybody kind of assumed that AI for all eternity was only large things, even though the day before, it was trending towards distribution and smaller -- so our prediction is that now that we have a baseline of the large-scale Gen AI models, which are very good for public generalized services, now we want to take them to the enterprise, more specialized narroscope distributed environments. We'll just begin that journey again. We'll start to figure out ways to make them more efficient, which is code for reduce the burden on infrastructure, reduce the power consumption, reduce the amount of data necessary, but still make them valuable. And so we go through this periodically in the AI journey of jumping to a new order of magnitude and then figuring out how to optimize it so that we can run that thing not only in the privileged space of a couple of infrastructures, but anywhere the customer wants it. So that is likely going to be one of the big trends that we see as the market evolves, not just who adopts it but the composition of these systems inevitably will start to become more efficient, and that's code for being able to run it in more places and more diversity, which basically will catalyze the industry in our view.

I hope that helps, Simon.

And our next question is going to come from Erik Woodring from Morgan Stanley.

Jeff, I just maybe want to take a comment you made earlier and maybe expand on it. And that was you obviously -- a lot of people have been spending time on the ISG side about this Gen AI opportunity. I'd love if you could maybe tease out some of your comments on the PC side and the opportunity for CSG to benefit. Will AI at the Edge kind of catalyze PC refreshes? How do we think about the timing of that? What's the incremental kind of componentry you might need to add to your products to handle these AI workloads? Just maybe a couple of incremental details that would help us understand the opportunity for Gen AI specifically in PCs.

Sure. I'd be happy to, Erik. Look, any time that we can get a new technology that drives productivity into the best general purpose productivity device on the planet, it's -- we're better off. And when we look at what Microsoft's plans are with AI and future versions of Windows, it's doing just that. It's going to make the workforce more productive. And any time we've seen that with previous versions of Windows, it's driven a substantial refresh cycle. We think that's the opportunity here. And what's equally important about this refresh cycle, you're going to ask your PC to do more with some form of assistant, some form of, I'd say, do some language modeling or language processing, it's going to do some machine learning with the capabilities that we'll put on our PCs in the future. And that's going to drive a higher ASP. You're going to need a more capable PC to ask it to do more. Then that's good for business. We've seen ASPs increase over the past 3 years. The likelihood that continues, asking more of it is highly likely. And then you have the subcategory in PCs, one of my personal favorites workstations, engineers, developers, creators, designers, data scientists, working at the Edge using those high-performance PCs with GPUs in them to do more complex AI tests. You'll see next-generation PCs with NPUs in them, neural processors. And those are going to allow us to do some of the basic neural processing that John referenced earlier, that will be in every PC going forward. It's quite efficient way to do it, optimized for cost and power. We're pretty excited about the new PCs that we will be building on top of the embedded AI services and capabilities that we put into our service stack and our software stack already in our PCs today. Today, we do a lot of work around how to help customers optimize their performance by workloads, the telemetry systems that our service organization has tied into our PCs, we'll be able to extend that customer experience more broadly. So I know it was a long-winded answer. I hope that gave you some context of what we're thinking about, what we think the opportunity is and how we extend more AI to the Edge. John, anything I missed from a core technology point of view at PC?

The only thing I'd add is one of the challenges -- first of all, there's three different reasons why you need AI in a PC and you're probably -- you're familiar with them. I mean, one, obviously, is or do you need acceleration in a PC for AI. If you're developing models having a good high-performance workstation with GPU is pretty helpful. We know that with our precision [line], we're doing that for a long time. The second, though, is the proliferation of co-pilots, individually, each co-pilot you run probably doesn't need a massive accelerator. But if you think about the not-too-distant future, you're not running one co-pilot. You'll have a copilot doing transcription, a co-piloting translation, a copilot creating automated imagery, a copilot filling in the gaps and what you're talking about with contextual information. And as you add more and more "AI workloads" to the system, the idea of having a portion of your semiconductor allocated to do that in an optimal fashion makes a lot of sense. And that's why we can see this inevitable trend towards within the CPU or in other types of accelerators, more and more of that functionality, which kind of brings us to the third area, which is really the user experience, in general, it gets transformed as we do this. We expect just more immersive user experiences on the PC because there's more parties involved in it. These co-pilots actually manifest. They show up in better imagery. And so it's a pretty profound impact on the PC over the long term because it's the interface the customer has. And it's also the place where you localize many of these experiences around co-pilots. And the combination tells us we're going to need richer user experiences, spatial representation, just greater depth of field to be able to present the kind of data that we're at, and we're going to need both direct processing and co-pilot processing, if you will, as the number of kind of AI is working on your behalf around your increases. And we don't know exactly what they're all going to be, but we know there's going to be many of them, and we know that the platform they're likely going to run on will be a distributed architecture, which the PC is the kind of personal representation for the user. So it's a journey, but we don't see any -- we don't see any other path other than more and more processing on the PC, more of it dedicated to AI-type tasks and a richer user experience, which you can imagine all of those things are pretty good for us.

Thanks for the question, Erik. I hope that helps.

And our next question comes from [indiscernible] from Citi.

One of the questions that we get from clients is just as the spending is kind of being allocated towards AI, do you see that as temporarily perhaps taking funding away from a refresh, whether that's on the PCs or further dampening enterprise spending on servers and storage after a very strong calendar '22, if you may? Of course, we have the macro pressures, et cetera. So people are just trying to gauge if this temporary -- the spending on AI is sort of taking it away from some of the more spending that could have happened at least during the second half of calendar '23 on mainstream servers, storage and definitely as we look at the PC ahead of a Windows refresh.

I'd probably look at it through that lens, it's a little longer term. I don't think this is at the expense of one or the other. This is additive. Again, if I think about what we're hearing in our discussions with customers before use cases that I described and how CEOs are viewing this in the boardroom, this is really a discussion where generative AI can change the basis of competition. It changes the way we're going to develop products and serve customers that drives a significant productivity increase that I mentioned earlier, but to restate it because I think it's worth restating, how many times have we in this working generation had a 15%, 20% productivity improvement? I can't think of one. So this is additive. CEOs and boards are looking at this opportunity is not at the expense as that's the real productivity improvement that is out there, how can I not do this? And if I don't do this, I could be left behind, and if I'm the left behind, I may not catch up. This truly changes I think the basis of competition for many companies. It's going to disrupt cost structures. It's going to disrupt again how you serve your customer in a more intimate way if you can figure out how to get ahead of your competitors in any given sector, it's a huge advantage. So our experience with our customers and talking to CEOs and the market research we've done, suggests they're not thinking about, oh, I'm going to add AI and to not do this AI project. I'm going to actually extend PC lives by 6 months [indiscernible] I have to invest. This is that game changing. John uses the word inside our company as an inflection. This really changes the way to how technology is going to be used. And I think he's right. It certainly -- it's certainly in the discussions we're having really [our] leaders and I think how our company as a leader in our company, we're thinking about how do we take advantage of some of the coding assistance or just organizations looking at 20%, 30% productivity improvements based on the complexity of code. We think about how much work is actually language-based depending on whose research you look at anywhere from roughly 60% of all work is language-based. 60%-ish of that could be addressed with generative AI technologies. So when you think of it in that perspective, it's game changing. And as a result, I know leaders like myself are going, we need to invest. We have to stay competitive. This changes the way we're going to deliver and build products for our customers and serve them. And I know that wasn't quite a direct answer, but it's not about the TAM of this quarter and what's going to happen. It's that much of game changer. And we think of this in terms of this is an industrial revolution. This is the steam engine. This is the assembly line. This is the Internet. This is what the PC was 40 years ago and what it did to productivity. And it's all happening and happening in a much, much faster rate.

And our next question is going to come from David Vogt from UBS.

Great. This is helpful. I want to pull back and maybe spend some time on storage. I know you talked about richer configurations on compute. And ultimately, richer configurations at the edge of the network, particularly at the CSG side. But what -- when I think about generative AI and the other flavors of AI that you touched on earlier, what does this mean for storage demand, particularly are we going to see new sort of demand for whether it's more basic levels of storage, that software defined, that's a little bit cheaper to deploy? Are we going to see more all-flash systems deployed? How do we think about the data that's going to be developed and used for inference as whether it's hot data, cold data? Just trying to think about all the different permutations of how this could play out as the enterprise starts to spend more aggressively on storage. And then ultimately, can this, from your perspective, be delivered as a service from the storage side? And is that part of the thinking going forward from Dell's perspective?

Excellent question, David. We'll try to unpack it a little bit, I'm going to ask my resident storage expert to come in and help me in a minute. But if you step back and we look at what's happening, clearly, much of this data is unstructured. I mean it's going to come in, in a massive scale. So if we go back to one of the fundamental premises that's been driving the industry, the rate of data creation is not slowing, it's accelerating. It's absolutely the case. It's really accelerating outside of the walls of the data center, as we've talked about many times out on the Edge of the network and the form that it's taking is unstructured. And it's coming at massive scale. And the types of systems that we're thinking about have to be able to scale that way. They have to be able to respond. And to your point, while we use that [ position ] inside our company, the hot data at the Edge being handled in real time. Doesn't lend itself to take a long trip up to some cloud and back down. It moves itself being treated right there. The triage has to happen with the data is created. So the algorithm is going to be run. That's where you're going to see some of the micro tuning being done, where you'll see drift detection being done and where we'll be modifying the inference as a result. I mean, those things all play well to, I think, certainly our strategy, what Jeff and the team have been building in ISG and you think about old flash or we've done with [indiscernible]. You think about the scaled-out architecture that we've built with our unstructured products, we have a very broad and deep portfolio to meet the needs to where the data is going and the type of data that's being created. Jeff, I know you can make that sound a whole lot better than I did.