NetApp, Inc. (NTAP) Earnings Call Transcript & Summary

All right. Looks like folks are still trickling in, but maybe in interest of time, let's get started. Good afternoon, good evening, everyone, depending on where you are. I'm really delighted to host this webinar with the team at NetApp, and really the focus of the webinar is how is NetApp poised to benefit from the broader enterprise AI deployment and how they are differentiated versus what seems to be an ever-evolving competitive landscape. A couple of housekeeping items before I introduce the speakers. We'll aim to keep this webinar to around 50 minutes. We'll have the executives from NetApp kick this off. Actually, we'll start this with Kris giving some safe harbor disclosures. We'll get into some Q&A from there. But I would encourage that the intent of this is to be interactive. So if folks in the group, and it's quite a large group here, have any questions [Operator Instructions]. So with that, we'll be delighted to have with us Kris Newton, who is the VP of Investor Relations at NetApp. In addition, we have Russell Fishman, who is the Senior Director of AI Solutions at the company. Russell is responsible for solutions, product management, including NetApp's AI solutions portfolio. And he's been at this function for the last 4, 4.5 years, but he's really been at NetApp for the last decade plus. We also have Hoseb with us, who goes by, I guess, only his first name, who is the Senior Director of Global AI, Data Analytics Sales and Go-to-Market with responsibilities and oversight of NetApp's AI sales and go-to-market strategy. Hoseb has been in this function for the last 5 years across multiple geographies, starting up in Dubai and in the U.S. currently. And so maybe those -- introductions. I'll pause, Kris, I'll let you go through your safe harbor disclosures before we get into the questions.

All right. Thanks, Amit, and welcome, everyone. Today's discussion may include forward-looking statements regarding NetApp's future performance, which are subject to risk and uncertainty. Actual results may differ materially from statements made today for a variety of reasons, as described in our most recent 10-K and 10-Q filed with the SEC and available on our website at netapp.com. We disclaim any obligation to update information in any forward-looking statement for any reason. That said, I will remind everyone that this is much more focused on technology and the use cases and the market evolution as we see it for AI rather than a financial update. Back to you, Amit.

Perfect. thank you, Kris. I guess, Russell and Hoseb, maybe just to start off with right, spend a few minutes just talking about the history of where NetApp has played when it comes to AI. And what has the team been working on over the past, call it, 5, 6 years around this AI development. And really, the intent is just to understand kind of the history of the company were -- on the AI side and then where do you see this going a year forward?

Sure, Amit. Let me take a step on that. So we all know that this whole AI wave has been around for multiple years, right? So in 2016, NVIDIA obviously announced the DGX-1 before that, there were other people doing a ton of different analytics type of workloads. But really, the AI wave started with NVIDIA announcing the DGX-1. And obviously, after that, NetApp saw an opportunity here, especially in those areas where -- we call it, the deep learning at that time and predictive AI. We came up with the reference architecture with our ONTAP storage with DGX-1. We called it on ONTAP AI in 2018. And since then, we started adding a lot of customers in different verticals, the strongest being the health care, public sector, et cetera. And that's where our mission and journey with AI started. That's where I started taking over the mission, moved to U.S. to build the team and also our product and engineering teams built that references with NVIDIA and kept going on, right, till we saw the wave of obviously, gen AI with ChatGPT and OpenAI. And we continue with the innovation. We -- our reference architectures kept growing with NVIDIA. We added the SuperPOD 3 years ago, then obviously, the A100, the H100, et cetera, with the NeMo Retriever right now. So we kept innovating in the past 6 years, adding customers, as I said, in different sectors. Backing then when we were doing deep learning, customers were leveraging our technology to build the models, even sometimes used to call machine learning. So one of our first customers was a large health care provider in United Kingdom, who basically deployed ONTAP within DGX-2 from NVIDIA at that time to build models to feed the CT scan images into those models and predict whether a tumor is benign or malicious, et cetera, et cetera, right? So until today, where we have, for example, recently closed a large SuperPOD opportunity in one of the leading genomics and health care companies, where they will be building models that will be trained on the world's largest human patient data, right? So this has been the progress that we had in the past 6 years, whether on the sale side and the product side and the engineering side.

I'll just add to that. I think one of the big changes that we've seen is this inflection point in the last 12 to 18 months where AI has moved from being solely the domain of companies that are looking to innovate and get ahead to something that is seen as a necessity to stay in touch with their competition in their industries. And to that end, one big change that NetApp has made in our fiscal year, this current fiscal year is that we've moved from purely a specialist sales motion to one that is far broader that touches every single one of our sales force. And I think that's just a reflection of the interest level amongst our customer base, amongst all customers to want to engage with NetApp around AI.

Perfect. I think it's a great point, which is -- just assume you suddenly get more wide -- wider in terms of deployment beyond just the 4, 5, 6 hyperscalers. I think one of the questions I get asked and I see a lot of folks -- you folks get asked a fair bit, which is what is the "AI opportunity" for NetApp going forward, right? How do you kind of put dimensions around that would be really helpful. And maybe as you answer that, I would love to understand that. In 3 buckets, if you may, which is one kind of an incumbent installed base, how does that kind of stack up? Secondly, the next new infrastructure build and then third, there's a vast amount of data that actually needs to be prepped for enterprises for the AI use cases, right? So maybe just talk about what is the AI opportunity. How do you size that up for NetApp as you go forward?

Yes. I would love to first start by characterizing the workload a little bit on the AI side. So you've got definitely the training workload, which is basically building those models where data gets fit into the GPUs and then it trains the model. And obviously, you've got the inferencing piece, which -- basically putting those models into production and then -- now inferring on those models. And then obviously, on top of that, you've got techniques like fine-tuning and RAG. So once we segment those, we will better understand the opportunity for us. When it comes to model training, let's start with that. Model training is actually something that we have been engaged, as I mentioned, since the early days of the DGX inception. Also in addition to that, GPU-based other OEM-based servers, right? We have customers today who have deployed servers from different vendors with GPUs in them, but they rely on NetApp technology to feed those GPUs with their data and training those models on. So I can tell you that NetApp has been very instrumental in that space. The opportunity for us over there, whether a customer is building a separate cluster to train those models, or they are leveraging their currently existing data that sits on ONTAP technologies or NetApp technologies, whether it's in the cloud or on-premises, we have built all those integration points to be able for those customers to feed that data into the GPU clusters. So I want to repeat that in the training, we've been really, really helping a lot of enterprise customers. We have customers all the way from 2 DGXs, for example, into more than 80 DGXs in one of the largest model training clusters that we have today in one of the financial institutions. So in model training, we see a good opportunity for us. We see that opportunity growing. But more than that, if you come to the RAG and the fine-tuning that is where we also see a tremendous, tremendous opportunity for NetApp, which actually sparked a couple of months ago, as Russell mentioned. For those folks on the call, RAG is basically a technology, which is called Retrieval-Augmented Generation is where you don't need to retrain those models using huge clusters, but you can deploy those pretrained models but make it relevant to your businesses, make it updated because traditionally, those models will be trained on a year-old data, right? So that model gets into production. And then if you use technologies like RAG, and that becomes very important now, you would be asking where is the storage in this RAG technology. The storage comes actually -- the RAG technology requires a preexisting data to be converted into numbers, into numerics that the models would understand better. The industry analysts, a couple of them already out there saying that, that process will actually generate almost from 1x or 100% more data to almost 10x of the data because those embeddings would generate -- would need more numerics to be stored on. So that is the opportunity. Now that is an opportunity a lot for our installed base, and this is where NetApp is really well positioned in the market, being in the market for more than 20 years, storing that [indiscernible] for our customers, enterprise customers for almost 20 years gives us the ability now to transform those data that is sitting on NetApp installed base into this embedding. So that is a huge opportunity for us that we anticipate in the next couple of years, especially in the RAG technology. And obviously, the last piece is the inferencing, right? The inferencing is something that we anticipate, for example, these RAG models when they start generating new types of documents, PDFs, videos, audios, et cetera, one might be saying, but those would have been generated anyways without the RAG models or without the gen AI. But I would say the pace of those data being generated is going to get accelerated. Because if you were able to do -- to produce a movie a year, now you would be probably able to do 2 movies a year, 3 movies a year, same with documents, right? So those type of new documents that would come out of inferencing, we anticipate that to be also a good business opportunity for us. Now put a wrap around this is the cloud and the Hybrid Cloud story we have. Because we believe that AI is a true hybrid workload so that means it is an opportunity for us both on-premises as well as in the cloud as well as in a hybrid model as well.

Yes, I'll just add to that. So a couple of things. So if you look at how industry analysts look at the amount of unstructured file data that sits on NetApp, it's north of 30%. So we have an outsized share of that market, and that is the field that's driving the current gen AI wave, particularly. So we're extremely well positioned when it comes to gen AI. But more broadly, my suggestion to the folks on this call is to follow the data. I think that what you're seeing is you're seeing new entrants coming into this space who don't have any of the data, who inherently have to start by taking data from somewhere else, creating a new copy of that data and managing it. That's not just expensive, it's complicated to manage. It even causes problems with things like the AI Act recently launched and put in force by the European Union, where the ability to know where your data is and how it's being managed becomes not just a case of best practice, but one that is subject to regulatory and compliance concerns as well. So I think that NetApp is extraordinarily well placed from that perspective. So incumbent installed base for sure, what we see is certainly in the training environment, these are all new infrastructure builds. So all of that business has been in that space. Obviously, when it comes to generative AI, we're talking about taking -- augmenting existing environments, so we see that as well. And in terms of just the amount of data needs to be prepped, we'll talk about that, I think, as we go through this call and we'll get into a little bit more detail, but it's significant. And as Hoseb said, it's also growing, not just in terms of the amount of data that's growing, but the way that data needs to be prepped and worked through actually generating more data in and of itself to run through techniques like RAG.

That's great and yes, I do want to spend a little bit of time just on RAG and inferencing, what that means to be focused as it goes forward. But when you think about some of these kind of capacity requirements, if you may, I'd love to kind of understand when do you expect to start seeing benefits from the growing data capacity requirements that some of these AI models have?

Well, it's happening now is the answer to the question. We're firmly in this gen AI wave. But I'll take a step back and just give you some sort of industry numbers, if you will, on sort of what this data growth looks like. So Hoseb mentioned this technique called RAG or Retrieval-Augmented Generation, which is really the leading technique used to augment foundational models with context for a particular company's data environment. And as part of that process, that data is -- go through a process that generates these things called vector embeddings and these embeddings get stored in a vector database. That vector database then gets enacted to the foundational model at the point of prompting. So when a user asks a question, it is able to use the vector database to connect to underlying data sources. I kind of think about it like a librarian with a bunch of books behind the librarian, the librarian themselves doesn't know content of the books, but knows how to go find that information and how to read those books and get the right information to answer the question. In terms of how much data is generated, and the answer is what I think the industry as a whole is continuing to refine its answer in this space. But what we have seen as an industry is somewhere between 4 to 10x of typical office file environments. So that means if you have, for example, a terabyte of office file data, that would be Word, Excel, PowerPoint, PDFs, et cetera, et cetera, the sort of content that typically gets used in, for example, an enterprise knowledge management use case, we would be seeing somewhere between 4 and 10 terabytes of additional data generated as part of that RAG process and that data doesn't disappear, it stays there. It actually continues to get updated as the source data changes or expands. Interestingly, this data also with current techniques doesn't seem to be compressible in any way. So that actually translates pretty clearly into an increase in raw capacity requirement. But I've given you one example. There are other examples. If we look at another common example, which is using code bases for Retrieval-Augmented Generation, we're seeing 200x data increases. So it really depends on the type of data. And I think as the RAG market gets more mature, I think the industry will continue to evolve standard metrics around this. But clearly, there's a significant opportunity there. In terms of when, we're seeing a huge amount of growth and interest in AI, one of the reasons why I mentioned before that we've really engaged all of our -- NetApp's general sales force, for example, and engaging with customers around this is because the appetite is there. Gen AI is probably, I would argue, one of the most accessible forms of AI. It, in many cases, moves directly to a value phase from a customer ROI perspective. And so that increases the level of interest from customers who may not have had the sophistication to go build models from scratch. So we really think that, that wave is happening now and it's something that we expect NetApp and other folks in the industry to gain benefit from in the next 12 to 18 months.

Got it. And as you said, just think about all this, how should we think about the timeline of how data is stored through the AI life cycle in its entirety, right? Like what happens to the new data that's created? You touched a little bit on how you can't compress it and everything else. But I guess will there be different performance requirements depending on where you are in the life cycle be that training, tuning or RAG or inferencing, for example?

Yes. So there's a lot to unpack there. I would say that one thing that we have seen as an industry is that because data through a typical life cycle for data in AI has been highly siloed if folks don't use solutions like the ones that NetApp provides. But if you allow each of the different personas that typically are involved in making AI real for a company to do their own thing, you could end up with as many as 6 redundant copies of the same piece of data through that life cycle. So there is -- it's really interesting in terms of an optimization opportunity that obviously, NetApp is very well placed to take advantage of. And I think what's kind of embedded in what you just asked, which is around the performance requirements. The reality is there is a significant variation in the performance requirements, in fact the workloads -- we kind of talk about AI as a singular workload, in reality it's lots of different workloads with lots of different use cases. Each one of those use cases with their own performance characteristics. So what we're seeing is customers are going to be best served by a truly flexible solution that can deal with data at all stages of that pipeline. So whether it be at the data unification phase or the prep phase or the training phase or the model management phase or the inferencing phase or the feedback loop phase. Each one of those is really a different set of performance requirements, is a different type of characteristics. And one of the things that they're naturally focused on is taking the amazing ONTAP leading storage operating system and applying in lots of different ways through that process. And I'll just say one other thing that's kind of interesting here. I think again, this is one of the things we'll probably get into a little bit more as we go through this is that if we think about all the workloads NetApp has seen over the last 20 or 30 years, I would say that AI is probably the most hybrid workload we've ever seen. And actually, if you hear other folks in the industry talk about the AI workload solely through the lens of cloud or on-premises, they're doing that because of their portfolio. But if you actually go and talk to customers and when I say customers, I don't just mean lines of business, but data scientists and data engineers. What you'll hear is that, that flexibility and the fluidity that they need to move and manage data, both on-prem, cloud through service providers and other places, bringing data to GPUs -- and GPUs and AI to the data. They are going to be very well served by the strategy that NetApp is pursuing in this space.

Got it. I guess, this would be a good time maybe for you folks to talk a little bit about what do you see the competitive landscape as right now? And maybe just talk about -- there are a lot of companies that are, I guess, for a lack of a better word, AI-washing the narrative. I'd love to understand, who do you see really in these customer deals that you give your customers? Who else are they looking at? Looking for RFPs, for example, and maybe also contrast how has this landscape changed from a historical perspective to you folks?

Listen, I'll give you some facts and I'll let the audience decide who is AI-washing or not. But I think it is very -- it is very obvious that we need to clarify what is being identified as an AI customer today, right? Especially in the storage world. If we want to easily call someone just using data to do some dashboards and say they're doing AI, we can call that out day and night. But I want to really specify when we say our work in the past 6 years, and you heard me say a lot NVIDIA and DGX-1 and GPUs because we truly believe that those are the -- really the use cases that there will be no argument around that, whether the customer is doing an AI or not because we all know GPUs and the purpose that customers are utilizing them. So the competitive landscape, I mean, I would say, has changed over the past 6 years. And it changed dramatically. In the past, in 2018, I mentioned there were probably 2 publicly traded storage vendors, they were certified with NVIDIA. So you could see us compete with another vendor in that specific deal. And it was -- the market wasn't hot as it is today. So you could basically identify those customers, handful ones. As we grew in the market, obviously, COVID hit, right? And our priorities have shifted. If you look at NetApp's history, we were probably the only ones from a storage standpoint who kept believing that this market is going to grow. And we kept investing in this space. We kept our teams together, including myself and our specialist team because we knew this market is going to grow. And in the time of 2019 until 2021, '22, we were probably taking customers left and right. We didn't see much competition coming in. Our priorities have shifted in other vendors. But after 2022 when SuperPOD started becoming more and more important in this space, obviously, a push from NVIDIA came on that front as well. We started seeing HPC type of storage vendors come into the play in our deals, right? So we started seeing, I would say, a healthy competition on that front from vendors, also on not public, also start-ups get into the play. But from an enterprise level storage player, we don't see that much of competition in our deals. Now you have to be very careful here. Again, very specifically talking about DGX opportunities, GPU opportunities, we also compete with other server vendors who also tend to provide storage, right? So that's a natural competition over there, too, where a customer might buy a GPU from a server vendor and the storage. But I -- as I mentioned in the past, we also were able to attach our storage to customers who bought GPUs from service -- from different types of vendors who also provide storage. So we've also seen some competition over there. But if you look at NetApp, I think we are the ones that provide that end-to-end, as Russell mentioned, portfolio that could cater customers' requirements, whether they are building a 2-DGX cluster for training, whether they are using cloud like technologies like SageMaker and others, and their data is in the cloud that they would love to feed it to those GPUs for training or other techniques like RAG. Or if they are building clusters like SuperPOD, then NetApp has a solution for that.

And I'll just add, the one thing that separates us out, I think, from our competition from -- in a competitive landscape is that we aren't competing with the hyperscalers. I think if you go and look at the other folks in this space, especially the ones that like to talk about themselves as AI specialists -- by the way, if you ever hear that, NetApp is a specialist in AI. It's just that we don't only specialize in AI. The reality is that we don't compete with hyperscale as we partner with them. And that is based on a legacy of building out these 1P, first-party services, cloud storage services at each of the 3 major hyperscalers and seeing integration between those services and the native hyperscaler solutions, some of which actually just happened to be announced today. So that puts us in a very different position. And it also means that, again, back to that idea of fluidity, the flexibility that we offer the actual AI practitioners in managing AI workflows in a hybrid manner. We're not trying to force customers into one modality or another. We can support whatever they want, I think so as well as supporting best performance requirements, we're also able to support however a customer wants to construct their AI data pipeline.

Got it. I have a question -- a couple of questions from folks on the call here. One of them is just -- it's on the competitive landscape. Is there a different set of customers or different set of competition you see in the cloud side versus hybrid? And really the question of the -- intended around vast data systems and where do you see them play versus not? Maybe we'll stop to that and there's one more after that.

Yes. I think where we see our value is in the enterprise data level management capabilities. We see a lot of customers, especially on the enterprise side, the large pharmaceuticals, the large financial institutions where they really care about their data being secured, they really care about their data being managed efficiently where they have been relying on NetApp technology for many, many years. and not only price performance density, right? That is also another area that we really need to be very careful about is, hey, there will be customers who are really looking at price density performance. And then obviously, also being certified with NVIDIA, which means we have met the requirements to serve those workloads. But when our value comes in, as Russell mentioned, is this workload being hybrid. I was talking to one of the financial institutes last week who happens to have data sitting in the 3 hyperscalers as well as on-premises. And they wanted to create this foundation for their data to be sent to those GPUs, wherever they are, they really like the strategy that we presented in terms of creating the data fabric. Us being natively in the 3 hyperscalers literally differentiates us from any other storage vendor out there in the market. If you want to be only on price performance density, you can just put 10 competitors next to each other and they all can compete on the price and the performance. But our value, again, I repeat comes from enterprise-level support, enterprise-level security, enterprise-level data management, our history in the market for almost 25-plus years. And also our native integration with the hyperscaler consoles in the major 3 hyperscalers.

And I'll just add to that. Performance is a requirement that is table stakes. If you look at the way that NVIDIA has gone about talking about performance when it comes to their relationship with storage vendors, they have spent a lot of time and effort creating validation and certification environments that essentially level the playing field from a performance perspective, you either make -- meet the requirements or you don't, if you meet the requirements then let's move on. And certainly, NetApp meets the requirements across all of these solutions like SuperPOD and BasePOD, et cetera, et cetera. I think the other thing I would add is you don't get fold between the difference between having a cloud offer and having a native 1P cloud offer. The way that we get treated as an OEM essentially to the hyperscalers in the delivery of our first-partied ONTAP-powered cloud offers is completely different to the way that they treat the marketplace offers that some of our competition rely on. Certainly, when it comes to integration with their AI services, the hyperscalers are incented to do that work between their services, which remember, these NetApp-powered services are actually native services in each of the hyperscalers. Again, we just made some announcements specifically today around that with AWS at the AWS Summit in New York City, demonstrating our commitments jointly with hyperscalers to integrate our technologies directly into their first-party AI services.

Yes. And maybe one more thing. One of our messaging to our installed base, specifically and also our customers is like you really don't need a siloed architecture to run this. And if you look at NetApp, right, in the past years, every different technology came, we try to integrate the solution with our existing tools, with our existing ONTAP, with our existing operating system, which means customers do really not need to buy some new technologies in their data centers that will require them to hire and train new individuals to -- especially for our installed base, which is really, really good compared to others, right? So this is one of our differentiators is not -- we're not trying to create a separate architecture in a silo to run this workload. And as I said, if we really meet the certificate, then we met the performance requirements already set by NVIDIA, which are very high standards, right? So there is no reason why would someone just deploy a different architecture to run the workload that could run on the architecture that they've been using for many years.

And you were just talking about NVIDIA, I was hoping you could -- if you just expand on your relationship with NVIDIA. And maybe you can frame it around what enabled NetApp to get SuperPOD certified earlier than some of your public peers, at least their comparable products? And maybe related to that, you can just talk about what do you think differentiates NetApp SuperPOD versus offering some -- your competition right now?

Yes, I'll take that. We have a very strong relationship with NVIDIA. We're not claiming to be the only one, but we have a very strong relationship. It started back in, as I said, in 2017, '18, and we continue building up on that relationship. As a matter of fact, we were very proud to be the only storage vendor being called out on Jensen's keynote at the GTC, the recent event from NVIDIA. Your question about SuperPOD. It's very important, and thank you for asking that because that's SuperPOD certification, us having the SuperPOD certification and the BasePOD certification. And for the audience on the call, BasePOD certification today, in the past as well, relied on Ethernet as well as an NFS file system, whereas the SuperPOD certification, in the past again, relied on the parallel file system and an InfiniBand requirement. So they were very distinct, 2 different architectures. And then in 2020 and onwards, we went through almost 6 to 8 months of testing with NVIDIA because of all the requirements they had to be certified, and we were 1 of the 3 vendors actually to be initially certified for SuperPOD. The reason we had that is because our portfolio was able to do that. Our portfolio was rich enough and had partnerships well enough to be able to pass NVIDIA SuperPOD certification early on so that we can provide our customers the ability for them to choose the architecture and the solution that fits their requirements. So when I walk into a customer, I'll be asking them, "What is your workload? What are you trying to do?" And if they say, "We want to go SuperPOD. That's the architecture that we were recommended by NVIDIA," fine, we've got that for you. If they say, "No, we don't have InfiniBand in our data centers. We're going to rely on BasePODs with Ethernet and all that," we have that as well with our BasePOD certification. So that has been a very strong proposition for us. And I can clearly say that we have customers who are leveraging both technologies. We also have a customer who's leveraging both technologies in one time, meaning they all have a SuperPOD for their large language model, training type of workloads. And we also have them using BasePOD to utilize technologies like containers and RAG and NeMo Retriever and the different things from NVIDIA. So we are very strongly positioned in that space, and that is all because of our rich portfolio that can manage those different type of requirements from the customers.

Yes. I'd just add that this is a situation that might surprise the investors to hear actually makes it much easier for what you might describe as a legacy supplier, someone who's been in the industry for a while because the market is very fluid. It's changing rapidly. The technologies that are being leveraged are changing. And it is relatively easy for us at NetApp with our very broad portfolio of capabilities and rich technological integrations that we've built over time to quickly adapt to a changing market. Some of the start-ups, the privately held companies are having to build everything from scratch every time anything changes. And they have virtually an inherent interest in keeping the status quo, which is kind of ironic in a space like AI, where the market is moving so quickly.