Coming to you from the Windy City. Welcome to Let's Talk Shop, a podcast about all things cloud and enterprise tech. Listen to insights and guest interviews with IT thought leaders and professionals. Now here's your host, Elias Nasser. Welcome back, everyone, to another episode of Let's Talk Shop. Today we're going to talk about, well, you guessed it, AI. It's kind of the only conversation in town these days. And for that matter, we've got NetApp to discuss AI and I've got a ton of questions on, well, howis I changed the storage landscape? What are they doing? How is it different than what they were doing with analytics and big data just a few years back? And to answer that, we've got a friend of the show. Russell, welcome back to the show. How have you been? Good. Thanks for having me. Uh, I appreciate the time today. Exciting topic. Excited.to get stuck in. Is there any other topic you're talking about with customers today? I think you think I was the only thing that's going on in the world. Uh, yeah. Look, AI's everywhere. Everyone. It's the one conversation I can walk into any customer and have a conversation. They want to talk about it. They want to learn more. Um. And listen, um, I'm excited to have the conversation with you today. Amazing. So before we get started, for those that don't know you, maybe what do you do for NetApp, exactly? Yeah. Thank you. Yeah. I'm, uh. I'm a senior director of product management. I look after the NetApp solutions business. Um, been doing that for a number of years, but.most importantly, within that, uh, I, I've been looking after for the last four and a half years. And so solutions is all about connecting customer outcomes with product. So it's really about using products in the, in the environment that actually delivers an outcome. So we're really customer focused. I spend all my time talking with customers um, and making AI real for them. Right. Rather than just putting products out there that do AI I type things, but actually sort of focused on outcomes and delivery so that it's really a great place to be. It feels like it'ssuper satisfying to help customers actually get something working and delivered. Um, and yeah, excited toconnect on it today. So Netapp's been touting intelligent data infrastructure for a while now. It's again, everycompany that I'm talking to is kind of repositioning their message around AI. But when you're talking at it from again, from a NetApp perspective, how has intelligent data infrastructure made an impact on the data pipeline, whether it's ingestion, whether it's processing, the whole data pipeline? Howhas it made a difference? Yeah, it's a good question. I if I look at the market in general and I look at customers and also, you know, frankly the way that people like Nvidia like to talk about it, you'd think I was just training. Honestly. I mean, because that's,all that people like to talk about. And I don't blame Nvidia for doing that. They'vekind of. They sold a lot of chips because of that. So they've sold a hell of a lot of chips. Right. Andit's all about you know they talk about performance and etceteraAndit's always great to hear, you know, industry visionaries like Jensen talk about where AI is going and what it really means. But, um, you know, the reality, though is training is actually only a part of the inference. Is everything now right? Well, inference is where the value is, right. I mean, you know, you build something, but if you don't use it, where's the value? Right. But before you even get to training, how much time do you have to spend ingetting your data ready for AI or AI data ready, so to speak? Um, and I'll tell you,know, one of the things that we've learned as we've gone down this journey with,our customers, even us internally, is that although it might not be obvious at the start, data is probably the biggest challenge that customers have and organizations have as they go down the path of AI. The ones that haven't started don't think it's a problem. And I you know, as long as I get the GPUs kind of GPUs something success. That's all they think, right? Um, but actually the ones that have tried it realize really quickly that it's a data problem. So thatsort of keeps us super excited about, you know, what can we do to help customers bridge that gap between, hey, I've got a bunch of data that I know I can use for AI and actually delivering value. So is it if we're talking about inference is now kind of where the conversation is moving towards or where it's going to be, or where the value is going to be to towards what you said. Doesn't that shift the conversation towards, well, I want all the GPUs and thento well, now I want all the CPUs and thenwhere does NetApp fit in? How does you how do you fit in. Yeah. So that's that. Yeah. So that that's a super good question. Firstly I'll say yeah. You know I know everyone likes talking about GPUs. We're going to see a you know as things move forward we're seeing a variety of different accelerated compute infrastructures and architectures, but the thing that's common throughout it is data. Right. So if you think about you're building engines, um, but you need fuel. Uh, gas, gasoline to drive those engines. So, uh, with a gasoline. Right. So it starts with, you know, knowing about where your data is and how you bring it together in a way that makes the data useful. Uh, preparing the data. So getting the data in a form that you could actually use for something like training. And I'm thinking about sort of, uh, full end to end. Uh, traditional AI life cycle. Then we're thinking about, you know, actually getting that data, um, you know, in a very highly performant fashion into a training environment that's obviously popping out some sort of model. That model then needs to be tested and ultimately deployed, potentially at scale. Um, there's data that needs to be driven into those inferencing environments. And then frankly, you're,learning as you inference what works and what doesn't work. And that's a continuous feedback loop that has you go back into retraining or what have you. So in that way, we sort of think about it as a, you know, this,virtuous cycle of data. And one of the things that's really interesting about it, and I know your listeners may,or may not realize this, is that firstly, AI is not one workload. I mean, there is no one AI workload. There are many,applications of AI and even techniques like retrieval. Augmented generation isn't really a workload, it's just a technique that you can apply to lots of different use cases. Sofirstly there's a plurality of these differentworkloads. And they all have different requirements right. But even within a specific workload if you think about that pipeline you know it's again it's not one thing. It's multiple stages. Each of those have very specific requirements on what you need to manage the data, the performance you need, theability to protect the data to provide things like data governance, etc., etc. all that sort of stuff means that you need a fairly flexible all approach to data. It's not about one size fits all. It's really about how you apply a sort of consistent approach to data management in lots of different ways through that process. I don't think I'll tell you. I don't think that's something that organizations have really fully appreciated yet. But as we move past this kind of getting AI, working into getting AI efficient and effective, this is going to be I think this is the next big thing, honestly. Okay. So if we were to drill down a little bit, what specific features, for example, of NetApp would be potentially beneficial for performance, for scalability, yadaAnd can those features be used across multiple environments like on premises, multiple different cloud providers? What specific features can you share thathelp with scalability and performance? Yeah, typically for AI workloads by the way. So again, the whole topic overshadowing topic. Yeah. Look, I mean, um, you know, at first I'd say that it is definitely about the right performance for that right workload in the right stage of the pipeline. So as I was mentioning, you know, it,isn't a singular workload. It isn't even a singular, um, sort of process. There are lots of elements to this, uh, to most of these deployments. And so actually it has a lot to do with finding the right performance, which gets you the right sort of efficient and effective price point as well. So you can do the whole thing really efficiently. It's, you know, you could absolutely buy a Lamborghini to go down to the drugstore and pick up a prescription. You probably, you know, that's an expensive way to get to the drugstore. It will work. It's a nice way to get to the drugstore and listen. It's one of those things. If I had the Lamborghini, I'd probably think a bit differently about this. But the point is,that, um, you know, there it is all about having the right tool for the right job, but doing it in a way that's really consistent. So,absolutely, you need to meet those really high end performance requirements. And you know, folks like Nvidia have done a really good job through their reference architectures like Superpod setting a bar, saying, hey, listen, this is the required performance you need to get the most out of our GPUs. And unlike all the vendors, we absolutely meet those requirements, right? You have to go through a pretty rigorous certification process with Nvidia where they they're testing you at scale and they have a very clear view. You either make you either meet the performance requirements or you don't. There's no overachieving because at the end of the day, they are very clear about what is required to get the most out of their GPUs. So we absolutely do all of those things. Now, the question is for the rest of the data pipeline beyondtraining. So we saw the data ingest, data prep, um, inferencing, um, and retraining and fine tuning. Uh, you know, do you have the right performance for that. And so one of the interesting things about NetApp is, you know, we built this,storage OS over, you know, a long time ONTAP, um, and it's fairly famous in the industry because it's been around, you know, a long time, but it that is. Yeah, butit's interesting. It has. But at the same time, it is incredibly agile and modern in that it has adapted to all of these new workloads every single time a new workload, a new requirement has come out, you know, at ONTAP has adapted and grown in terms of its capabilities and feature set to achieve that, but it's also available through lots of different form factors. And I suppose this is probably the most key thing, right? So I mentioned, hey, listen, you know, there's lots of different requirements. Each one of those bits of that data lifecycle are kind of different. Um, so how do we deal with that. So we deal with that by giving you a sort of very consistent data plane and a singular control plane. That means that as your data moves throughout this, uh, this process, you may be doing it on different form factors of ONTAP. Right? Everything from not being funny, everything from spinning disk todifferent types of flash, critically Cloud, which I'll get onto in a second. Um, you know, but you're doing it in a consistent way, but you're matching each time you're matching the performance and feature requirements to the specific specifics of what's needed, without making it massively complicated, I think that's always the risk. You think, well, hang on a second. This sounds really complicated, Russell. Yeah, I can see that. Butif but if it's consistent. Yeah. If we're using the same OS and the same control plane, you know we minimize that complexity. Absolutely. Right. But do any specific features of ONTAP come to mind that help with scalability and performance? Yeah.Soyou know, so thinking about what I was describing. So let's talk about features like our synchronization technologies like SnapMirror. Right. And SnapMirror is going to help you replicate data in a super efficient way between points often with extremely low latencies. Right. We do. That's a more of an asynchronous replication technology. We have synchronous replication technologies too, right. Like MetroCluster. And then if we go beyond that, we have things like FlexCache and FlexCache allows us to take our, you know, uh, the,namespace, make it instantly available at a, at a remote site and then, and then slowly make that data available as needed. So performance for me, by the way, you know, I think it's important to note this, right. We performance at a pure level comes down to things like, you know, the protocol you use to access the files, um, and enhancement technologies like what Nvidia calls the GPU direct storage or GDS. Right. There's all that sort of stuff. And but that's what you'll hear most folks talk about. Do we meet the performance? I tend to minimize those conversations by saying yes. And Nvidia said wedo. Right. So you kind of have to just trust that it's happening now. I'll tell you one thing. That's actually the one thing that's very interesting is that we have a very distinct NetApp has a very distinct point of view on how to deliver extreme performance, uh, for storage. Right. And traditionally, you probably know this. They you know, most companies have focused on parallel file systems historically, right? There's a whole bunch of them out there. Um, and some,very well known, some open source ones, like,lustre would be kind of a famous open source one, of course. Right. Um, and then there are folks that have tried to do sort of a more standards based approach, but have ended up with the bane of every single data engineer that I've ever spoken to, which are custom clients. Right? So essentially, hey, yes, you get accelerated access for training, for example. But to do that, you need to install a kernel drive, a kernel mode driver into your Linux distributionYou probably need to put some kernel patches in there toget it working really optimally. It's not upstream, it's not part of the mainstream Linux distro. So you're adding the stuff in and it creates complexity and frankly instability. Again this is one of those nightmare. You go and talk again. You go to talk to data scientists and data engineers. I hear this every time. This is a real problem. So Netapp's got a very unique position that we have gone down the path of completely standards. So NFS right? NFS is not new. People have heard about it before. They're probably thinking, wow, I've heard the promise of NFS over the years. It's never really delivered. The performance that is all changed, right? We actually delivered our, um, our, uh, super pod, uh, our new super pod certification, uh, based on NFS. Right? Um, so we got that extreme performance based on NFS. We've been optimizing NFS. We've been basically contributing upstream to enhancements to NFS. The long and short of it is you can now take a standard Linux distrowith NFS. You can connect that up to NetApp, and you're going to get that extreme performance without the need for custom clients. Okay. So Netapp's been in this game for a very long time. Um, as far back as I can think of, I've known NetApp and NetApp was there way before I came to be. Right. So yeah, before I, we were talking about analytics and big data and high performance compute and all of these other things. I feel like at least from the way I followed the progression of the industry, all of the things you're describing, we already had for analytics and big data and high performance compute. So how does where does the AI magic come in? Like what do you have now? Because again, you've been tackling these types of workloads for a long time. And maybe some you know, some viewers didn't know that. But Netapp's been at this game for a while. But when you sprinkle an AI, how does it give you something you didn't have before? Yeah, I mean, AI is absolutely a, you know, is kind of a revolution, right? In terms of the impact it has oncustomers, these organizations looking to,kind of optimize to create new revenue opportunities, etc.,So yeah, I don't want to underplay the importance of AI. And you're absolutely right. There'sa whole bunch of stuff that NetApp has been doing for a number of years that has suddenly become super useful in AI. So I'm going to start there, and then I'm going to. That's a very good point. So I'm sorry to interrupt, but I think that's a critical point. And I'm glad you said that because hopefully our viewers, our watchers understand this. But some of these capabilities were already there. Maybe they weren't as useful for everyone, maybe a niche, um, community of users or customers was using them back then. And now it's with the with all of this AI thing going on, a larger amount of users and customers and enterprises are using them, as opposed to just the data scientists. The niche use cases that you have, you had before. Would that be fair? Yeah, I think it is. I think that thereal issue has always been that data manageability, features that have traditionally been locked up behind storage admins are extremely useful in these in a lot of these areas, but they're not very accessible. I mean, that's always the problem, right? And certainly the speed at which AI has been moving and sort of the approach that folks like when I talk, when I go out and talk to data scientists, data engineering organizations, a couple of things that I hear. Firstly, theydon't really care about infrastructure. Sure. It's always been the case. Who are you? Why do I want to talk to you? And that's and that by the way, that you know and that's been doing this for about seven and a half years, right. And, you know, when we started, it was definitely, you know, we focused on things like infrastructure simplification and that sort of stuff. But as we got further down the on the road, what we realized is,that again, those features you were talking about, they're actually very useful in data science workflows, for example. Right. So I mean, let's just take a very simple one, like as a data scientist, I am gathering data and I'm trying to generate a data set that has been prepped in a way that enables a training job to be successful. Okay. Very simple. As I do that, I'm typically iterating that data set a large number of times. Right. So it's not once or twice I might be iterating it 1020 times. Um, as I do that, I might need to go back, right. So I might need to say, oh, that didn't work. What I just tried didn't work. That labeling job didn't work. I'm going to go back and make a change. And once I've actually done that, I'm normally not working next to the accelerated compute. So I then need to move that to accelerate compute. All those things sound very storage oriented. I just described snapshots and clones. Sure. Right. I just described, uh, synchronization technologies and data movement technologies, things that are very,sort of traditional NetApp value. So one of the first things we did was, hey, how do we take that value and make it accessible to these data scientists, right. Data scientists and data engineers, they aren't infrastructure people. They don't want to use infrastructure tools. They certainly don't want to be raising tickets each time they want to do something like this. So,the whole idea has been, hey, how do I sort of package those things up and make them useful to these people? I'll give you an example of how we've done that. Right. So, um, MLOps tools, right? Um, there's a whole bunch of them out there, some very famous, obviously open source tools like MLflow. Uh, we've actually, uh, in the last few months, we announced an integration with, uh, one of the leading, uh, enterprise MLOps companies called Domino Data Labs. Okay. With Domino Data Labs, uh, we integrated at a essentially an API level into our ONTAP operating system. Now, ONTAP operating system doesn't mean on prem, doesn't mean cloud. It means everything actually. Right. So this does. Yeah. This is not a and we'll get on to cloud in a minute. But you know uh it means everywhere. And uh, what it means is as a data scientist sitting in my MLOps tool, kind of my Microsoft Office for Data Science, I now have the ability to do all the things I mentioned in a super accelerated and efficient way, right? So from their perspective, it's quicker, right? Um, I don't have to raise a ticket with the storage team to do it. I'm not running out of spaces. I'm doing it because it's super storage efficient as well. Right? I don't have to get up and walk to the water cooler because I'm working with a 200 gig or, you know, two terabyte or even bigger data set. And every time I make a change, it takes, you know, 20 minutes, half an hour, two hours to,modify that data or to, uh, move it to somewhere where I need to,have it, which isn't where I am locally. Um, as you go through all of those things, you know, suddenly these features become super useful. Now, do they know they're using them? No. But do they get a better experience when they're using them? Absolutely. So wetend to go talk to like a AI center of excellence. This is kind of the conversation we'll have with them. Hey, you can make yourusers lives better by I integrating these capabilities in and we're doing that with other. If you think about the sorts of, uh, data, uh, sort of data workflow components and the sorts of workspaces that folks will use. Jupyter Labs, Jupyter notebooks, Jupyter Labs would be another good example of this. Um, you know, how can you make that super accessible through, you know, Python or what have you? We've been doing all that for years, so that that's one way that we're making our features sort of super accessible, uh, to folks. Yeah. So you're essentially creating an abstraction layer that simplifies all of these things. And I think that's a great that's a great thing. Um, now everything's got pros and cons. So my next question is a little difficult. I can't even have a great answer for it. But it's a question that a lot of, a lot of it's a lot of people's minds, which is the idea of locking. Now, I've always told the customers that I've advised that you can never eliminate lock in, but you could reduce lock in to a large degree. And when I look at what NetApp has done, especially with your partnerships and your integration partners, whether it's Nvidia or Intel or AMD or the hyperscalers or, you know, Equinix, whoever. Right. So Netapp's done a great job of diversifying, being open, you know, using ONTAP, like you said earlier, everywhere you want to integrate with S3. You're able to do that. You're able to abstract, able to extend some of these features. You want to do it on prem. You could you want to do it another cloud provider. You could. Now two things I want to ask you and two challenges that I want to ask you. Historically, even before this whole thing came about, one of the challenges for Lockin was just the sheer size of the data. You know, if I've got, you know, petabytes on NetApp, I'm,kind of locked into NetApp just because of how difficult it is toleave. I've got petabytes of storage, the same thing with Amazon S3 andothers. But in this case there'stwo aspects to it. Right. So there's the sheer size of the data that's creating a sort of lock in. But there's also well, now you've kind of integrated everywhere. So I'm not locked in necessarily with Amazon and I'm not locked in with, you know, any other provider. But I'm sort of locked in with NetApp, am I correct or am I missing something? I think NetApp has got an absolute commitment to open standards that, you know, permeates everything that we do, right? So that means open standards in terms of data access. And we just mentioned NFS before. That's an example of how we're trying to avoid locking in our customers. We also have a, you know, sort of a history of contributing to the open source community and creating standards. A really good example of this that you, your listeners may not know is the container storage interface for Kubernetes. So obviously, the standard for persistent storage for Kubernetes actually was a NetApp invention and contribution to the open source community. So that gives you a super easy way to plug in, frankly, other vendors toyour Kubernetes environment. So we recognize there was a better way of doing it, but we contributed it. Um, and it's now the standard in the industry. So, um, I think that we, you know, we obviously think we're the best of it. Um, and so, you know, I, we don't consider that locking. We consider that just choosing the best vendor. That being said, um, look, um, when you walk into our customers, you know, yes, there are customers that are all in on that app. There are customers that use us alongside other storage technologies to, um, and especially in the cloud as well. So if you think about our experience in the cloud, we have first party essentially OEM services in the cloud, which means that we've OEM ONTAP to each of the three hyperscale or the three main hyperscalers. Right. So that would be Amazon, Azure and Google and those the services you can buy that have the word NetApp or ONTAP in them are not NetApp services. To be clear, wedon't sell them. We don't uh, we don't market them. Uh, we don't support them. Actually, that's all done by the hyperscalers directly with the customers. But they do use the same ONTAP technology. Um, and, you know, even within the cloud providers, you know, people are using a plurality of,different storage technologies to,achieve their goals. So, listen, weknow we're in an ecosystem. Ecosystems are really important to us. Um, we still think we're the best. So, you know, if you choose all NetApp, that's not necessarily a bad thing. Um, buthey, listen, I will tell you,know, but maybe, like, a nice kind of follow on to this might be just a,conversation around, you know, how we're helping customers get their AI data ready. Right? Because the basis of your question, I think, is reasonable in that NetApp has such a significant proportion of the world's unstructured file data sitting on us. Right. Um, and, you know, we have a huge amount of it sitting on us. And, uh, you walk into customers. A lot of organizations certainly walk into fortune 500. So many of them are going to have, you know, NetApp, at least at a minimum, looking after their unstructured file data. So,we kind of, you know, as AI has become bigger and certainly as unstructured data has become more important in AI, specifically with the rise of large language models and techniques like retrieval, augmented generation, we've really sort of felt like we had a responsibility tohelp our customers get their data AI ready. Right. And so that's really been, you know, one of the sort of guiding principles for us in the last year, we announced, uh, back at our user conference last year at INSIGHT, uh, we announced that, um, that we were extending our Ontap operating system to do a whole bunch of new stuff that was designed to help customers get more value out of their data. And, um, that's a really interesting sort of way in which we see storage evolving from what are primarily retrieval based systems, which is where kind of storage has been up to now, to something that is more,like a queryable interface. Maybe that's the best way of describing it, right? So how would you get a data AI ready? Is it a matter of helping them with, you know, more metadata, like what would specifically NetApp be able to do to get the data kind of AI ready? Yes, it's a good question. Metadata is definitely part of it, right? So metadata in of itself, you know, it doesn't mean anything. Right. Unless we actually say what types of metadata. No,you obviously that's what you were saying. But I'm saying for the audience, you know, metadata generically is just a way to store additional data alongside your existing data. So the question is what are we doing with metadata. So if you already have your data, let's just start with the sort of do you have your data. Now if you have your data already in NetApp and we've you know, that data has, you know, has been created and has gone through its life cycle on that app. There is an incredible amount of information that NetApp can derive from that data, right? So let's think about just some basic stuff. So access controls and policies are obvious right. But data classification you know do you know what that data is. Right. Andthat's a feature actually we already have uh, that we,announced we're going to bring into this new sort of, um, this new these new enhancements, extensions of our existing ONTAP operating system, um, data lineage. Right. So, you know, we are we understand the data lineage because we understand where that data started, where it's ended up, how it's been, you know, modified and changed over that period of time. Um, it, you know, thatum, we're not the only place that you're going to understand data lineage inside your organization. I think we're a key sort of contributor, if you will, to thestory of data lineage, rather than just being a kind of a basic file space. Um, and then, um, the last one I would, I would call out would be data quality, right? So is there some way to derive some measure of data quality by having the file sitting on there? And the answer is yes. I mean, all of those things I just mentioned are types of metadata that we intend on making available to our customers now. Um, why is metadata so important? Those additional bits of information can help, uh, techniques, let's say retrieval, augmented generation. It can provide a way for rack systems to weight the inputs. Right. So if you're able to weight the inputs based on, let's say, a measure of data quality, it's likely that your retrieval augmented generation system is going to come up with a better answer. Now, this might not be obvious to your listeners, right? Because a lot of folks have gone down the path of like, rag chatbots and have started off with the idea that the data that they're going to use for it is, um, essentially hand-picked. It's curated. Right. So I guess I'm going to feed this rag to 100 documents that someone has picked out very specifically for my environment that are the gold standard. And it all sounds great. And in a POC that works fantastically, right. The problem is, once you get to the real world, yeah, no one does that right. No one is doing that in the real world. No one. Um, the reality is,you end up consuming all the data in your environment, right? And if you consume all the data in your environment, how can you tell one piece of information that's good or better, you know, higher quality than another piece of information that might, you know, essentially be conflicting with that first piece of information, right? Um, that's where we can step in with these types of metadata. That's a really good example. Another good example would be policies, right. So let's say you want to manage the way that your data is moved and used. Maybe you're subject to legal or regulatory compliance concerns. Maybe you're in an industry where there are specific requirements, let's say financial services, life sciences. Maybe you're in a region like European Union where the AI act right, and you are required to restrict the ability for certain types of information to move in certain ways with metadata at the storage layer. You can actually enforce those policies in a way that you couldn't if it was further up the stack. Right. Um, you know that that's the sort of stuff that we can do. Yeah. So I mean, that'spart of, you know, data, you know, you you're pretty much describing data governance to a large degree. Um, what about traceability and security, which I'm sure the system has, especially for, you know, high secure environments or where, you know, you're in countries or regions that require, um, you know, adherence to certain, uh, rules and regulations. Um, canyou do that? Can you apply that only to the on premises environment, or can you have these this data governance go across your multi-cloud providers, even though every one of those hyperscalers is offering your service as a tier one type of service where they're offering it. Yeah. ONTAP still kind of be that umbrella across all of them. Absolutely. And actually, I suppose there's actually a really interesting statistic thatfolks may,or may not be aware of. Right. So in a traditional AI data life cycle, if it's if people are just left to their own devices, um, a single piece of data may be copied as many as six times. That's from what we've seen in the real world, right? I don't think that's a particularly, uh, controversial number. I think, you know, others willrecognize that number. And by the way, when that piece of data is copied six times, it's not copied six times to the same system. It's copied six times to lots of different systems. So you end up very quickly. I think the problem statement is you end up very quickly with, um, if folks remember going to like, you know, fairgrounds. And there was a, there was a game people could play called Whac a mole where this little thing, you know, you had the mallet and you, you'd whack these,little plastic things on the head as they pop out. That's how I feel that, you know, when CISOs and CDOs look at this new world that we live in of AI and they're thinking, how do I keep this under control? And the reality is,when you have that many systems, you know, the idea of replicating your governance or security policies to every single one of those systems, but just the management overhead or the ability to be have visibility to what is going on across that environment is impossible almost. Yeah, it'simpossible. I think that people are pretty much got away with it up to now because, you know, no one's really noticed. Now we're seeing and. They didn't really have a choice, like, okay. I mean, fixing it was very difficult anyway. Very difficult. So,that's changing, right? So firstly, you know, as I, as I kind of mentioned, you know, I think NAB's got a view that we can do this with one OS, one sort of data plane, if you will, one, you know, compliance and security environment. But to your point, that critically extends the cloud as well. Right? So that's a pretty unique thing. Right. And you know, I'll tell you that we you know, I believe NetApp believes I think the industry is coming around to the idea that AI is intrinsically a hybrid workload. Right. Sowhat the hell do I mean by that? What I mean is,that in the long run, we believe that most customers are going to be best served by extending their AI workflows on prem to cloud. Now, that doesn't mean it's one or the other. And by the way, there'll be customers that only go all in on prem, all in, on in cloud. We think most people will be some combination of two. There's a whole bunch of reasons we could spend a whole hour conversation just talking about why that is, but let's just leave it at that. There are very significant investments that the hyperscalers are making, not just in accelerated compute, but tooling that makes it sometimes better to use those things than try to do it yourself. That's basically what it comes down to, right? And so we want to make it as easy fororganizations to,take advantage of that as is humanly possible. Right. And as you say, doing that in a way that your security and governance and data manageability posture remains intact as that data moves seamlessly on prem to cloud that your again, your CTO and CISO can put their finger on the data wherever it is in that process. They know where it is. They know who can access it. They know it's controlled that.is one of I would argue that is one of the most significant challenges. That is going to be the difference between where we are today with AI, which is small number of sophisticated organizations doing great things with AI to everyone. I mean, the masses, every organization, even those that haven't, you know, don't have the sophistication but just want the outcome, being able to take advantage of AI. And I think that's the crux of why I always advise customers that come to me specifically with how do I eliminate lock in vendor locking in particular? My answer's always been, and it's been controversial. Lock in is good. And I just want to come back to this because and I'm going to put myanalyst hat on here for a second. For those that are watching and viewing us folks without lock in, without a lock in with NetApp, you can't achieve this. It's impossible to do this without some layer of lock in. And I'll also tell our viewers a lot of them I talk to and advise. You can't use the best database in the world without being locked in. So if you're using an Oracle or a SQL, you're locked into these, right? So lock in is good, and lock in is a topic that only came about when the cloud came about for a variety of marketing reasons, but I don't think customers should be afraid of locking lock in is what's going to get you the features that Russell's been describing, specifically the advantages that you can take off from a governance from a security perspective. So I just wanted to reinforce that, in my opinion, folks that are trying to reduce lock in, you always want to reduce it, but lock in is good, and that's what's going to get you some of the features that you're seeking. Shifting topics on you just a little bit. So some of the benefits that, um, you know, AI is bringing in a lot of insights, a lot of capabilities, a lot of actionable knowledge for us and some of the challenges that we've had over the years. So some of the challenges we've had for the last 50 years is, you know, budgets are always tight. You're always asked toreduce cost. And, you know, for the longest time, these were all manual exercises for at least my career. I've not used an automated system of any sort, even analytics, to tell me how do I reduce cost? It's always been, well, let's take a look at, well, how much am I paying for storage and how much am I paying, you know, for the different tiers of storage and maybe some life cycle managementwithin it. But what can NetApp offer today based on all of the things that you're talking about from an AI perspective that maybe can contribute to cost reduction? Yeah, I mean, yeah, cost reduction is one of those things that never goes out of fashion, right? Ifthe economy is doing well, people want to save money. If the economy is doing poorly. They want to save money. Yeah,Solook, I mean, there's a few different things here. And so remember NetApp is baking AI into our products as well. So I'm actually gonna let me just start and give you an idea of how AI is helping us transform customer experiences with NetApp. Right. Um, and help them reduce costs. Right. So, um,onereally great example of this is our data infrastructure insights, uh, product. Right. Data Infrastructure Insights is essentially a management optimization tool. And it's not just storage. It does the whole environment. And what it's able to do is it's able to look at everything you've got. And that means, you know, compute, network storage, uh, through the lenses of different workloads. One good example, and probably quite topical right now would be VMware. Right. And looking at the whole thing and say, hey, how can you optimize it. And we brought AI in to help with things like workload placement. Right. How do you optimally place workloads to keep your, uh, core count low and to and how the role that centralized storage can play in helping to optimize yourVMware. Right. So that's one example, right. Um, another good example would be our Active IQ platform. Okay. Um, Active IQ, you know, has actually been a platform that has had for a long time. Uh, we've brought more and more AI capabilities into it. Essentially, we're able to do things like, uh, predictive failure analysis, right? Uh, I know the first time customers organizations buy NetApp, and they might receive a drive in the mail. And they're like, why did I receive a drive in the mail? I didn't request it. But of course, what's happened here is they're Active IQ platform has determined that you might have a failure.It hasn't happened yet. Right. And we're going to change that out before you even knew it happened. And there's a whole when we used to do it used to when we started doing that,sort of predictive failure analysis, it was kind of more rule based. Right? We would look at metrics. Those metrics would be sent back to headquarters. We'd whenthey opt in. And of course, and based on that information, we would determine when something was about to fail. But of course with that, I we're actually able to realize that things are about to fail without being so strict about a specific, uh, one specific metric, uh, or indicator that we're looking at. So that's another kind of example of how AI is improving our customer experience. And I think a third one. Right. Really interesting one anti-ransomware protection. So, uh, you know, ransomware, obviously the scourge, something that affects almost you know, every company is thinking about how to protect against ransomware. You know, I never throw shade, but, you know, one of our competitors, maybe this week announced that they had been subject to a ransomware attack. It's a shame they weren't using our technology. Maybe we could have helped them with that. Yeah. Great day. Good job. Come on. You had. To do it I agree, I agree. I couldn't. Resist, I couldn't resist. Um, yeah. But no. SoNetApp now it's been building this thing called Anti-ransomware.detection is actually an AI based system that sits locally. And why does it sit locally? It sits locally because we want to be the first, you know, one of the sort of well, actually not first line actually in an enterprise security context, you could argue last line. But the point is,that we have an extremely close affinity with the underlying data and,our,ability to detect ransomware and,critically, uh, make sure there's no data loss, even when we detect it is so strong. We actually alpha, uh, you know, an anti ransomware guarantee now, NetApp quite, uh, you know, I would say, uh, we're very, um, held we hold ourselves back from we're very conservative in terms of offering these sorts of guarantees. But, um, in this case, we have done that's howmuch faith and trust we have in it. It's a continuously evolving AI model that essentially sits locally. And that's another way that we're using AI to improve customer experience. Right. Um, so those are kind of. Ransomware is a great point because historically and we've had a lot of those, right? You know, you change the passwords or you delete the data or whatever the case is. So havingAI with Anti-ransomware support, you know, can alleviate a lot of that. That's a great point. Yeah. And nowthat'skind of you know, using, you know, NetApp. And then, you know, how does AI help our customers? Now, let's say you're trying to build AI, right? You're building AI and you're using that to build AI. How does NetApp help you save money there? Right. So there's a few different ways. Some things that we've announced and we'll be delivering shortly, some that we already today. So I already talked about the idea of finding optimal way to deliver different parts of the data pipeline. That is a key value prop that we have. 100%. Right. Um, but there are other things that we do as well. I'll give you one good example. Um, in retrieval augmented generation systems, those that have gone down that path have realized pretty quickly that, um, vector databases can get quite big. Right? You may have heard it described as vector bloats. I mean, there's all sorts of names for it. But, you know, what we see is anything from 4 to 20 x, the original data set size, um, is what you end up when you generate those vector embeddings and store them in a vector database. It's not surprising. Vector databases are absolutely set up for speed at the point of retrieval they're not set up forstorage efficiency. But if your data set. Remember we talked earlier POCs. They tend to have these very curated data sets right. So you know whenpeople do POCs for Rag, they say, well, I don't really care about that because, you know, my data set is, you know, 50GB. I mean, okay, so 20 x I mean, it's okay. Yeah, we're talking about a terabyte, but it's still, you know, it's still manageable when you get to the real world, people are indexing the entire, you know, they're bringing in the entirety of their data set. Yeah. And when you do that, you're talking about quite a lot of data, maybe multiple terabytes or more. When you start multiplying that by up to 20 x. So suddenly it becomes a lot of data. And that's a lot of capacity that you have to bring in. So one of the things that we've been really focused on is how to make, um, these vector embeddings of vector databases more aligned with the way that we store. And the reason we do that is so that we can do it more efficiently. So rather than having 20 x vector float, you know, maybe we get it down to 1 or 2 x, right. Sowe talked about this at our user conference last year. Um, and it's something that, you know, our customers will get access to uh, this year. Yeah. And that's a good example of, you know, I can see more data, more space. How can we make it more efficient? So you've talked about, you know, rag at length, you've talked about, you know, obviously there's content generation in there. You'vekind of taken a subtle dig at some of the competitors. But maybe everything is everything you're describing. What sets NetApp apart or are others? Do others have some of these technologies, like what sets NetApp apart from its competitors? Yeah, that's a really interesting question, because of course, you know, when and we've seen this, if you think listen, it depends how long the folks that listen to this have been in the industry. Right. But for you and I being in the industry long enough to know that this is a story that's played out multiple times and you workload comes out, the industry believes that it is so unique and so different that it needs a fundamental change to the way we do data storage. A bunch of new entrants come in, they build something new and, well, what's happened traditionally. Let'sthink back to previous times. This has happened, right? Um, VMware, right. When VMware came out, there was a whole bunch of very specialized VMware, uh, storage solutions that came out. And,you know, I don't think any of them survived long term. And there's a real good reason for that. I mean, listen, I don't want to poothe relevance of innovation in storage. And it helps us all stay on our toes. But ultimately, what it comes down to is, can you do this in a consolidated fashion without having to create new data silos? That that's basically what happens now in the era of AI. Hi. Um, there are some absolutely new things that are happening, and some of those things are relating to, you know, how much of these AI stacks and data workflows do you want to bring into the storage? Right. Andlisten, you could go way up the stack with NetApp. I guess the way of saying it is NetApp is very um, is it's not conservative, but iscautious because what we see are these AI stacks evolving extremely rapidly. Right. So rapidly that when I go and talk, for example, to a data scientist or a center of excellence at one of our customers, what we hear is,that, you know, the data scientists might literally go onto GitHub and download the latest version of something, right? Um, at the start of a project. Right. And so the idea that a centralized storage team would dictate the version and which components they can use to do a particular project is a complete anathema to them. I mean, it blows their mind thatwould even be a topic of conversation. So we are really, you know, concerned about ensuring that there continues to be flexibility and agility. But at the same time, we know that the role of data and storage is changing. Um, and so we want to evolve, you know, we're evolving with it. And probably the most, you know, without getting into all the specifics, all the different tools, but the one that kind of comes to mind is probably agentic and reasoning models particularly. Right. And again, I don't know to what extent your listeners havemanaged to do this, but if you haven't, I highly recommend going to one of the public models like, you know, obviously from open AI or Gemini from Google or going to one of the AWS models and,trying it out and seeing how Agentic AI works. The first thing you're going to see from these more advanced reasoning models, at least, is that they are very good at breaking down problems into manageable chunks. Right now, traditional. Oh, so the first generation of Rag systems have always taken the view that I, you know, as a Rag system, I need to, uh, maintain an index in the vector database and the vector embeddings of all the data in your environment, and keep that next to the model. Right. These reasoning models, though, take a slightly different approach. What we're doing is we're breaking down the problem into subquestions essentially, that I can then federate out now that instead of now, uh, indexing and bringing all that data into a single rag model, as long as I know where the data is approximately, I can actually send that question, right, thatsemantic query to the foreign storage system, that foreign storage system can then respond to that one semantic query without that,source system having to have indexed all of the data. So what you start to see is these federated storage systems, taking the role of almost like intelligent queryable engines that are able to respond to semantic queries. And that is a that is obviously a very different thing to what storage has done. Traditionally we've served files. Now we're responding to potentially, you know, natural language queries, um, to make that happen. Storage is evolving, right. So the idea of, you know, having an LLM and an embedding model and a vector database and a query engine and the API's to support that, bringing that close to the data honestly is a way more efficient way of doing it. Right. You just think conceptually of the idea of having to bring all that data into a centralized model rather than federate it out, which makes it just way more efficient to run, but also way more efficient to keep that data up to date. Right. I want to ask you something that, um, I'm hearing a lot from customers and that, you know, I usually advise customers against building kind of, um, you know, their own, but leveraging. So how does NetApp help those customers that don't want to build their own potentially, uh, infrastructure or their own, uh, capabilities around AI? How does NetApp help them take advantage of AI? Yeah. So thatthat's a good question. So there'sa couple of things like um, firstly we believe that AI should be accessible, right. And so look for that if you're an organization that wants to build your own AI solution, um, you have a use case you're targeting. Maybe you're working with a consulting organization. Maybe you have the skills internally. We want to at least simplify the experience of building an AI platform, right. So at the most basic level, that could be infrastructure up to the sort of platforms and frameworks and software components, making sure those are all tightly integrated, that we have best practices that you're not taking the risk on when you're deploying it. Hey, is it going to work? And more importantly, is it going to deliver the performance I need for the accelerated compute I've bought? Right. Sothat is absolutely one level of it. But I think you're right for pointing out that that's not a use case. That's a platform you need to go build and listen. There's a whole bunch of organizations that are willing and able and capable to go do that. But there's many more that aren't right thatif you think most organizations, as AI matures, are going to just want to buy the outcome, you know, you talked about, I think we mentioned inferencing as being the value phase of AI. Why bother with training? Why don't I just let someone else go do that for me. And I can just move to,use it and get the value out of it. And so we're starting to see a lot of focus around that. NetApp actually just announced something in this space. We announced, um, this thing called um, I pod mini, right. AndNetApp I pod mini is, uh, you know, one of the first sort of full stack appliance type solutions that you're going to see on the market. Uh, the idea is,that rather than buying a bag of bits that kind of work together, but ultimately you have to go and build your AI solution on top of. We're bringing a, uh, you know, a enterprise class chatbot for enterprise search and knowledge management as a full stack, right? So the idea being is you can. Purpose built environment basically, or purpose built. Platform. So, you know, you buy it through one of our partners, um, it gets deployed. It has a very reasonable entryprice, uh, fraction of what I think we've seen from other systems. But more importantly, it just works. It provides enterprise class chatbots that you basically point at your data. It's ingesting that data. It's presenting a chatbot for your users to utilize within your organization. Um, it provides an API that allows you to integrate into existing enterprise applications. So let's say you have an enterprise developer that you don't they don't necessarily have data science expertise. You can now use this,chatbot as an API into existing applications as well. So,that, you know, and I'm not saying it's the be all and end all of these sorts of solutions, but I think we're going to see more of that way, more of this kind of it just works. And,up to now in the cloud that you can get those things right. There hasn't been many opportunities to do that on premises. I think we're starting to see a lot. I mean, it's not just going to be NetApp you're going to see others in the industry do that. Um, butyou're actually going to see that. By the way, interestingly, the one that I just mentioned is initially available with Intel CPUs, right? Not even GPUs. So you kind of make this comment earlier aboutyou know, accelerated compute. It's not just GPUs. We are going to see a proliferation of other types of accelerated compute. I mean, if you look at yourcell phone, it has AI in it and it's not using a GPU to do that. It's using some form of Neural Processing unit, or TPU, depending on which device you have. Um, we're going to see a lot more of that sort of stuff coming up. You kind of answered my question. I was going to ask you kind of what you're seeing on the horizon or what you feel like, you know, the next step is going to be. So if I were to summarize it, what you're telling me is the near future or the near long term, near mid term future is going to be all about these purpose built platforms that are use case specific, that are more turnkey as opposed to a broader, more general platform. Did I get that correctly? Yeah. I mean, and actually I'd say I'd go, yeah, I'd go even go one step further. So what I described was kind of a generic sort of horizontal use case. It delivers something very specific. Butit but it's horizontal. Uh, but the reality is, you quite rightly pointed out, is that we're going to see a lot more, uh, industry specific and verticalized solutions moving forward. And, you know, um, for,those that of your viewers who've listened around things like LMS and obviously this move to LMS. So some kind of quantized models that are just way more efficient operating, but then they come pre-trained in something. So it's the difference between pulling someone off the street that knows everything, including who won the I don't know, the 1988 World Series, which is, while a wonderful thing, is not necessarily something you need to know if you're working in a customer service environment. Um, sothat's a problem. But then let's say I want an agent or a chatbot that's going to help me in a, in a legal environment. Right? I would probably not want to do all the training myself as the end user. It'd be great if that chatbot came firstly, not worrying about who won the 1988 World Series, but also had some of that legal, uh, experience knowledge. I'm always going to have to add my own knowledge, right? So whatever's specific to my particular organization. But you want that gap to be small andwhat I'll tell you is,that, um, you know, slms are way more effective, way efficient way of,delivering a particular outcome. Um, so we're going to see a proliferation of smaller, more tuned, more focused slms. And the less you can rely on techniques like Rag, the more efficient it is, right? It's more performance. So thiskind of trade off between how much I'm going to go and train or retrain and how much I'm going to just kind of buy a pre-trained model, that that's where we're going to see a lot of calibration, I think, over the next,few years. Russell, uh, thisconversation has been amazing. And you've just sparked a lot of, um, a lot of questions in my mind. I'd love to do a follow up podcast, maybe, where we talk about very futuristic things that maybe nobody's thinking about, or maybe somebody is thinking about. But one of the questions that came to my mind, and I'm not expecting you to be able to answer that is, um, you know, the more we mobilize, the more we have smaller form factors. I wonder when robotics and robotics are already here, but when they become a little more accessible, I wonder if we'll see an ONTAP within these robots as well. That is maybe designed to perform specific tasks, because as you were talking, I couldn't help myself but say, you know, AI powered robotics. This feels like a really good use case forsome of them as well, where, uh, you know, you're going to need storage, you're going to need intelligence, you're going to need a lot of these things to perform, again, very specific tasks. And I'd love to have a follow on conversation potentially on whether or not that's even possible. Or am I just a little bit in. The in the. I love it, I love it just yeah, let's make it try and make it not too Terminator y. Yeah. We're not going to build Skynet, I hope. Not yet. Yeah, that's a slightly. And you know, we laugh and then, you know, myAnd then we build it. My wife will say things like, you know, I do think robots are going to take over at some point. Look, no, in all seriousness, that would be a great conversation to have. And, uh, I will say this, um, I moved at a rate of pace that I don't think we've seen in the industry for maybe 20,years in terms of how revolutionary it is. Yeah, since the internet, I think that's a fair statement. Yes. For 25, 30 years. Right. So what I will tell you is,that anyone who tells you they know for sure what's going to happen in the next two years is probably lying. Right. I would say this to you. I'd say, look up. ChatGPT became big two and a half years ago. And no one saw it coming. And in two and a half years, the demands on infrastructure and specifically on data have changed immeasurably. Yeah. Right. So what are you know, I would say that, you know, what all of your listeners should be thinking about is not just today, but how do they ensure that whatever they put in place is giving them an agile environment that can adapt to whatever's going to come at them. And that's a really open ended way of saying you need to make sure that agility and flexibility is at the center of all of your strategies around AI, right? That's fair. And it'sjust not what people have thought. People by Oracle or VMware. And it's like I put it in it works for five years. I depreciate the asset, I move on, right? That is not how AIS plan is going at all. So the question is, are you working with the right folks to help you navigate that? Anyway, I'd love to have that conversation. Russell, it's been a great chatting with you today. Um, I learned a bunch. I loved the conversation. We're definitely going to have to do this again soon. But in the meantime, thank you so much for making time, uh, today. And, uh, we'll see you soon. Thank you so much. Take care. Bye