Audio shared by Stephen Galicia: Hey, good morning, everybody, and welcome to a discussion about getting to business value with your enterprise AI projects. I'm Tom Shields, and my job today is to set up the conversation by introducing the agenda and the speakers, and then I'll return at the end to moderate the Q&A. So let's go! Audio shared by Stephen Galicia: So here's the agenda we'll cover today. We'll start off with a quick overview of how we partner with NVIDIA to build best-of-breed customer solutions to advance adoption of AI efficiently and securely. Audio shared by Stephen Galicia: Then we'll cover some really interesting research from IDC that has dug deep into why AI projects succeed and why they fail, and outlined how data infrastructure needs to evolve to help enterprises achieve more success. Audio shared by Stephen Galicia: We'll then bring on NetApp and NVIDIA speakers to react to the research and hear how that resonates with our experience in the field and aligns with our own thinking on the evolution of data infrastructure. Audio shared by Stephen Galicia: We'll talk about how NetApp and NVIDIA solutions are evolving to keep customers, to get customers to business value in their AI journey. And then finally, we'll discuss a great joint customer win that illustrates some of the points that have been discussed. Audio shared by Stephen Galicia: Now, as we go through the conversation during the session, we'll have folks online to answer questions in real time, and we ask you to work through the Q&A channel in Zoom. Audio shared by Stephen Galicia: And then we'll save a few to address at the end with the team. Audio shared by Stephen Galicia: With that, I want to introduce our speakers today. Audio shared by Stephen Galicia: First, Tony Chidiak, and he's Head of Worldwide AI Sales at NetApp. Audio shared by Stephen Galicia: Morning, Tony. Morning. Audio shared by Stephen Galicia: And then we have Anne Hecht, who's Senior Director of Product Marketing for Enterprise at NVIDIA. Audio shared by Stephen Galicia: Hello, Anne. Hey, Tony. Hey, Tom. Phil, good to be here with you all. Good to have you. And then finally, a special guest, Audio shared by Stephen Galicia: We've got Phil Goodwin, who's a research vice president at IDC, and he's going to talk us through some of his research. So welcome, Phil. Thanks, Tom. Nice to be here with these speakers. Audio shared by Stephen Galicia: Great. Well, welcome everybody, and now I'd like to turn it over to Tony to give a quick overview of the Alliance's work in the service of our joint customers. Audio shared by Stephen Galicia: Tony. Audio shared by Stephen Galicia: Great, thank you, Tom, and it's great to be here with everyone today. Audio shared by Stephen Galicia: Over the past 7 years, our collaboration with NVIDIA has been really an extraordinary, successful journey, and in many ways, it does just feel like we're just getting started. I remember the early days, when NVIDIA launched the DGX, and we're rolling out the Audio shared by Stephen Galicia: DGX test drives with ONTAP, and just getting the enterprises to adopt, you know, these more power-hungry, but very,compute-intensive machines. And this was, of course, even well before the SuperPod came to existence. Audio shared by Stephen Galicia: Fast forward to where we're at, it's deployed all across the world. You know, in these 7 years, we've witnessed the evolution of what was predictive AI, which has Audio shared by Stephen Galicia: of course, been around for many, several years, but then this second wave of Gen AI, which just opened up the enterprise, the art of possibility, and also opened up Wall Street to just see the art of the possible. Audio shared by Stephen Galicia: And now we've moved into this third wave of agric AI and reasoning models. Audio shared by Stephen Galicia: Today, over a thousand plus customers rely on our joint solutions, and this number is rapidly growing. It's been an exciting process. And we've had this global partner ecosystem that has embraced many of our collaborative solutions. Audio shared by Stephen Galicia: And with half of the world's running, files running on NetApp, it's provided customers a seamless way to accelerate their AI initiatives using a Audio shared by Stephen Galicia: storage solution in ONTAP that is tightly integrated with NVIDIA. And perhaps the most exciting part of this is we're not slowing down at all. Audio shared by Stephen Galicia: That's great, Tony. I'm so excited to be here and be a part of this great webinar, talk about the work that we've done, the work we're doing right now, and what it's going to mean for our customers, today and in the future as this wave of AI sort of takes over and really transforms business. Audio shared by Stephen Galicia: And how we generate profits and revenue throughout industries. Audio shared by Stephen Galicia: Great, and before we say too much and get going on all the solutions, I want to hand it over to Phil to present the IDC research on the future of AI-ready data infrastructure. So, Phil, go ahead. Audio shared by Stephen Galicia: Hey, thanks, Tony. I appreciate the introduction. And I also want to express my appreciation to both NetApp and NVIDIA for inviting us to share this research, which I think is very important research in this AI journey that IT organizations are taking. Audio shared by Stephen Galicia: As Tom said, what we want to accomplish today is to help use this research to illustrate why AI projects are succeeding, as well as why they're failing. Kind of the lessons learned type of thing, the things that we have from our conversations with IT organizations. Audio shared by Stephen Galicia: We'll be talking about the compute infrastructure and some of the imperatives in that environment, but also particularly about the data storage environment and how that contributes to success or failure of a project, with special attention, really, on the data portion of that, which we find to be, to a great degree, the crux of that success. Audio shared by Stephen Galicia: I'll then close with what I think is some practical guidance for organizations on how you can get started on improving your data storage infrastructure. Whether this is a relatively new project for you, or it's something that you've been doing for a while, we want to help to improve that outcome. Audio shared by Stephen Galicia: The way we conducted this research is to first of all, talk to IT organizations and find out where they were. Audio shared by Stephen Galicia: When I mentioned a moment ago that the majority of POCs are failing, AI POCs are failing, that number currently is about 56%. Audio shared by Stephen Galicia: So, every time one of these projects fails, it costs the organization time, it costs them money, and it costs them the opportunity to gain some kind of competitive advantage in the marketplace. Audio shared by Stephen Galicia: Areas that we've found that are the root cause of these failures often relate to the infrastructure, and that's because legacy infrastructure simply isn't designed to support AI workloads. Audio shared by Stephen Galicia: Moreover, the data management that's necessary is often missing from this infrastructure as well, and we found that to be the critical success factor for organizations, and we'll share that with you a little bit more. Audio shared by Stephen Galicia: There's the need to balance the performance between the storage and the GPU. We've always had that, you know, you have to balance storage and compute, and that's especially important in the AI world as well. Audio shared by Stephen Galicia: So really what it comes down to is how do you get an infrastructure that can feed the right data at the right time, at the right speed, to optimize the performance of these systems, and to give the Audio shared by Stephen Galicia: the outcome that organizations need in terms of the modules that they have for their AI organizations. Audio shared by Stephen Galicia: If we look at some of the reasons that AI projects fail, we started to develop this bubble chart of elements that contributed to those failures. And when you first look at this chart, you go, man, that's pretty complex, and the truth is, it is. Audio shared by Stephen Galicia: But you can simplify it by starting to combine or bucketize these different reasons. Certainly, people is a big and important one at all times. Audio shared by Stephen Galicia: But what we found is they started to categorize in terms of performance and in terms of data management. Things like data silos became real problems for IT organizations. Audio shared by Stephen Galicia: Other areas that were cited were governance, security, recovery. Those kinds of things were highlighted in our research for improvement among IT organizations. Audio shared by Stephen Galicia: If we look at, from the, kind of the highest level of how organizations are implementing these LLMs, what we really determined is there's not a dominant model that one can follow. It's really an all of the above. Organizations use Audio shared by Stephen Galicia: commercially available LLMs, they use their own LLMs, they use the public cloud, they use the private cloud, and really what it comes down to is how do you balance the requirements of the specific workload? Whenever possible, organizations do try to go with a commercial LLM. Audio shared by Stephen Galicia: is generally more cost-effective, and generally gives faster results. But there are times when higher cost, higher complexity, data privacy, compliance concerns require that organizations deploy these LLMs. Audio shared by Stephen Galicia: On their own premises. So really what it comes down to is how do you balance those factors? How do you determine which compute environment is going to be the best? Audio shared by Stephen Galicia: The other element is, where is that data? Because you need to have the data and the compute together. And what that implies is breaking down silos, being able to have that single view of data across the organization, and being able to leverage that data at the right place, at the right time, for the AI application. Audio shared by Stephen Galicia: We recommend that IT organizations look at their compute environment holistically. What are those specific requirements in terms of performance, in terms of Audio shared by Stephen Galicia: size, and so forth, and to be able to balance those factors that we were talking about earlier. In some cases, you'll find that it's the cloud and the OPEX model that's most efficient, especially when you're doing POCs, or when perhaps you have a bursty requirement where resources need to be applied rapidly, but then taken away. Audio shared by Stephen Galicia: The cloud is often the correct environment to do that. But what organizations are now doing is really looking across the entire enterprise, whether that is in the cloud or on-premises, in order to meet those compute requirements. Audio shared by Stephen Galicia: I'd like now to turn specifically to the research that we did on AI-ready data storage infrastructure. And to conduct this research, we used two primary tools. The first is what's called a QA board. Audio shared by Stephen Galicia: Which is really a way of delving deeply into specific issues. It's like a focus group where we had 12 highly qualified, highly screened individuals Audio shared by Stephen Galicia: Where we could ask open-ended questions over a period of 3 days. We could ask follow-up questions, we could delve deeply into it, and it really gave us a depth of research. At the same time, we conducted a worldwide survey of over a thousand participants. Audio shared by Stephen Galicia: In order to get that breadth of research, where we could find out where things are going on geogra… from a geographical standpoint, or from an industry standpoint. Audio shared by Stephen Galicia: But before I get into those results, what I'd like to do is share with you what is our definition of AI-ready data storage infrastructure. And it starts with the hardware and the software and the services to prepare, ingest, store, manage, protect. Audio shared by Stephen Galicia: Secure, govern, and move data within the infrastructure environment. Audio shared by Stephen Galicia: At IDC, we call that data logistics, and you can use the analogy of package logistics of moving data from, or moving a good from the time it's manufactured to the time it's used by the end user. Audio shared by Stephen Galicia: The concept is very similar for data. How do we move that through the environment? And of course, that requires service levels to support those AI workloads, so that's another important part of the definition, as well as the data services that need to be applied to it. Audio shared by Stephen Galicia: To emphasize the importance of this, I'd like to share with you one of the quotes that we had from one of the participants in the QAL board, and this was a data architect in the U.S. Audio shared by Stephen Galicia: And he said, poor data quality, lack of data timeliness, and data silos preventing data usage were major contributors to AI project failures. Additionally, the underlying file system either lacked performance or adequate scale, and inadequate data governance further exacerbated these issues. Audio shared by Stephen Galicia: So let's see what else they told us from the co-op board. What we learned is the key success factors that they cited were high-performance environments with low latency, high throughput, as well as a seamless cloud experience, where you can Audio shared by Stephen Galicia: create that broad view of the environment and take advantage of both on-premises as well as the cloud. The reality is the vast majority of organizations are hybrid multi-cloud. Audio shared by Stephen Galicia: They use more than one cloud. They use private cloud. They use Amazon, they use Microsoft, they use Google, and many,others. Audio shared by Stephen Galicia: In addition, what they told us was the things that caused failure in their environments were inadequate data governance. Audio shared by Stephen Galicia: Unacceptable latency, inadequate metadata indexing, and data silos. As well as key challenges around compliance and data regulation. Audio shared by Stephen Galicia: What I want to call your attention to here is that quote on the left side of the slide, because I think it brings up a very important point. And here, an infrastructure architect from the UK told us, we struggled to identify a single source of truth in our data. The inadequate data governance was a significant challenge. Audio shared by Stephen Galicia: Data fragmentation across different fields made collection and curation difficult. Audio shared by Stephen Galicia: In our other research, we find that IT organizations deal with an average of 13 different copies of data. And whether that data is on primary storage, or secondary storage, or wherever it might be, finding that single source of truth Audio shared by Stephen Galicia: Is really a critical success factor to making sure that you feed the appropriate data into these learning modules so they get the proper output. Audio shared by Stephen Galicia: If we turn our attention now to the survey, again, this is worldwide, more than 1,000 participants responded, we found that 52.1% cited data quality as the most important factor in AI project success. Audio shared by Stephen Galicia: On the other side of the coin, 77.3% indicated that poor data quality played either a primary role or an important role in the failure of their AI proof-of-concept projects. Audio shared by Stephen Galicia: The top 3 reasons for these, for the poor data quality was, first of all, the time needed to generate the quality of data. How do you gather it? How do you find it? How do you identify it? Those are all important parts of how you get that data through the logistical Audio shared by Stephen Galicia: through the logistical channel, in order to be able to feed these LLMs. Then also, the final one was data silos. Audio shared by Stephen Galicia: And here, this was exemplified by an infrastructure architect from a manufacturing company in the US who said, data silos preventing usage, inadequate data governance, and insufficient data volume were major challenges. Audio shared by Stephen Galicia: Additionally, the underlying file system lacked performance and scalability, impacting AI project outcomes. Audio shared by Stephen Galicia: So let's take a look at what this infrastructure might look like. And if you think back to the definition that I gave you,can see there on the bottom you have ingest, store, manage, secure, protect, and access. Audio shared by Stephen Galicia: And that's really the data logistics journey the data needs to go through this hybrid multi-cloud environment, whether the data is created internally through applications or ingested through an internal source. Audio shared by Stephen Galicia: Obviously, it's got to be stored somewhere, it's got to be on physical,media somewhere. But some of the things that we're starting to see creep into these AI-ready infrastructure environments are things like parallel file systems. And they're becoming more important because they have the kind of performance that is necessary to serve the data. Audio shared by Stephen Galicia: Organizations deal with structured, unstructured, semi-structured, synthetic, parquet data, all kinds of different data. Audio shared by Stephen Galicia: That then needs to go through the journey of things like data exploration, of grouping, classification, indexing, tagging, so on and so forth. Audio shared by Stephen Galicia: Other critical elements are things like data trust, where you have knowledge of provenance, intellectual property protection and sovereignty, but specifically around data security, around encryption, RBAC, ABAC, Audio shared by Stephen Galicia: cybersecurity and so forth, as well as the data assurance to be able to recover that data at a stateful level, whether you get bad data because it was ingested from a wrong source, or because a bad actor ingested it, you need to be able to recover it back out of there. Really, when you look at this journey. Audio shared by Stephen Galicia: The left side of the environment, something we've been dealing with for quite a while, but getting further to the right are things that we're now dealing with more and more in the AI-ready environment. Audio shared by Stephen Galicia: Another diagram that I'll share with you is this AI-ready data storage infrastructure ontology. An ontology simply describes the relationship between elements, and in this case, I've expressed it as kind of a fancy Venn diagram. But you can see on the left the storage software and the storage hardware. Audio shared by Stephen Galicia: We've had that for a long time. Those elements that are in there, data protection, workload migration, capacity management, so forth, again, they've been there a long time. Audio shared by Stephen Galicia: But as you modernize your data storage infrastructure to meet AI workload needs, you're going to be moving further to the right of this diagram. And it is that data exploration that I just talked about a moment ago. Audio shared by Stephen Galicia: As well as adding things like LLMs, embedded vector DBs, agents, agentic AI, and so forth. Audio shared by Stephen Galicia: with the governance and the data trust that's necessary to support that. And it's all within the context of that bottom box, which is the hybrid multi-cloud environment that organizations deal with. Audio shared by Stephen Galicia: As well as bringing in AI tools from the marketplace, including OpenAI and Copilot as a couple of examples. Audio shared by Stephen Galicia: So, to bring it all together, what I'd like to offer is some practical guidance from IDC that we believe will help you to get more POCs into production and to get better business results Audio shared by Stephen Galicia: out of your systems. And I've broken this down into three phases. The first phase is what I would call the investigation phase, where you need to identify exactly what it is in those AI workloads that need to be addressed. Audio shared by Stephen Galicia: What is the business outcome that you're expecting? What are the results that would define success of the project? And then characterize that workload in terms of what are the performance elements that it needs, what kind of storage, what kind of compute, and then balance those requirements. Audio shared by Stephen Galicia: You need to factor in your data estate. Where is the data residing? Is it in silos? Is there governance that needs to be considered, or sovereignty, or other issues around security that organizations need to address? Audio shared by Stephen Galicia: Then it's what I call the architectural phase, and this is where you define what the data service requirements are. Determine data quality, timeliness, and all those things that go into the data logistics that we talked about earlier to get that data to the right place at the right time. Audio shared by Stephen Galicia: Critically important is identifying that single source of truth. Audio shared by Stephen Galicia: If you don't do this, it's going to be very difficult to really have optimal results. Getting that single source of truth, I think, is an incredibly important part. And also, treat data like a product. Treat it like a catalog within your organization, so that you can better leverage it and have better cooperation within the organization. Audio shared by Stephen Galicia: And then the last part is the implementation, and this is where organizations are spending time modernizing their storage infrastructure to get better outcomes. And this is where that journey that we talked about of moving more to the right, where you're building in more data management capabilities Audio shared by Stephen Galicia: Not just simply storing the data in order to get that AI success. Audio shared by Stephen Galicia: Then the last piece of advice is to look for embedded AI. Audio shared by Stephen Galicia: in the storage systems. And these are things that will help with dynamic provisioning, with troubleshooting, with error detection or threat detection, mitigation, and so forth, that will make your life easier. And so with that, I would like to turn it back over to Tony and Ann for their comments regarding their specific solutions. Audio shared by Stephen Galicia: Bill, thank you, for that fantastic overview. It certainly aligns with what we observe today, working with customers on a daily basis. And, you know, if I had to take all that data and put it into 3 key challenges that I hear, summarize it. Audio shared by Stephen Galicia: You know, number one would be modernizing enterprise storage for AI. The second one, delivering AI-ready data in a hybrid environment. And the third one, ensuring data quality, ultimately through that pipeline that has the right security and governance along the way. Audio shared by Stephen Galicia: So, why don't we dive into these areas and look at modernizing the enterprise, storage for AI? You know, across enterprises, we see customers grappling with legacy architectures that maybe they established 5, 10, or even 15 years ago. Audio shared by Stephen Galicia: These architectures have resulted in several interconnected challenges. And even a customer who recently summed it up, they don't have data management, they have data management chaos, where their own valuable data for AI initiatives is either not obtainable or not usable in the current form. Audio shared by Stephen Galicia: And then you run into different BUs, business units, where they often manage their data estates separately. Some are in the cloud. Audio shared by Stephen Galicia: Others on-premise or in different co-location facilities. These fragmentations create massive blind spots of data location, and even Audio shared by Stephen Galicia: the data existence itself. So, to maximize AI potential, the enterprises really need this global, comprehensive view of all their data, and then also ensure it's AI-ready. Audio shared by Stephen Galicia: If I jump to the next one, delivering AI-ready data in a hybrid environment, and why don't you go ahead and expand on this one, as I know you have a lot of experience on this. Audio shared by Stephen Galicia: Yeah, you know, Tony, just as you mentioned, you know, AI has really evolved, and with each advancement, the need for data and efficient data pipelines has increased. Audio shared by Stephen Galicia: You know, images. Audio shared by Stephen Galicia: audio clips, even videos. It transmits that data into twins. Audio shared by Stephen Galicia: in different,platforms. And training data… training data Audio shared by Stephen Galicia: use open models, and there's millions of them out there. Meta's Llama model family, Google's Gemma model family, Mistral, even NVIDIA. We publish the Llama Nemotron model family. There's just millions of options for enterprises to leverage that are open. Audio shared by Stephen Galicia: But what they need to do with these models is post-training. And skipping this step is really where a lot of, I think, enterprises get into trouble, because they skip this step, and deploying a model that hasn't gone through any post-training is like hiring a new employee. Audio shared by Stephen Galicia: And expecting them to be productive day one without doing any onboarding or job training. Audio shared by Stephen Galicia: So post-training enables the model to learn on a subset of tokens that are relevant to the use case and the business that'll be doing. Audio shared by Stephen Galicia: These could be tokens with domain-specific information for an application, like law. Audio shared by Stephen Galicia: Translation. Audio shared by Stephen Galicia: And make sure the model generates that Phil was talking about. Audio shared by Stephen Galicia: to an inquiry. Audio shared by Stephen Galicia: Where does that data come from? And that data is all over the enterprise. It's distributed. It's in different environments and on different systems. It's in the ERP system, the CRM system, it could be in the line of business data center, data stack. Audio shared by Stephen Galicia: It could be stored on-prem or in the cloud, and being able to access that data efficiently, wherever it's stored, is critical to that post-training process and enabling really domain-specific, accurate AI for an enterprise. And that's why we're working with NetApp, because you help with that process and managing that data chaos. Audio shared by Stephen Galicia: Yeah, no, absolutely, Anne, thanks. And obviously, working with the NVIDIA NIMS and the re-ranking model that we put in there as well also helps drive to that affordability piece on the tokens. Okay, the last bucket, ensuring data quality, security, governance through the AI pipeline. You know, bring back that data management chaos comment. If your data management is chaos, I can guarantee you your AI data pipeline will be even worse. Audio shared by Stephen Galicia: We're seeing things like outdated data, different versioning, lacking governance, policy capabilities, or security teams even just making binary yes-no decisions. Because they don't have the right setup, they're just completely inhibiting or stopping AI initiatives and projects. Audio shared by Stephen Galicia: And really, at the end of the day, so much of this is due to not having the robust enterprise features and capabilities at the data layer. Audio shared by Stephen Galicia: At the end of the day, enterprise needs to build towards an effective end-to-end data lifecycle management. And it not only delivers on the data access and classification, but equally prioritizes the privacy, protection, security, and governance that you need. Audio shared by Stephen Galicia: So let's now jump to the next slide and get into the NetApp and NVIDIA solutions. Audio shared by Stephen Galicia: I referenced at the top our 1,000 plus joint customers and touched on joint solutions, such as Audio shared by Stephen Galicia: the DGX SuperPod, NetApp AIPod, the NVIDIA AI Enterprise that we have across hyperscalers, and then the NCP reference architecture. But I really want to pivot and discuss NetApp being the intelligent data infrastructure that delivers for AI workloads. Audio shared by Stephen Galicia: you know, to keep it very simple, NetApp's role is to feed the GPUs with high-quality data, ensuring successful AI initiatives. Audio shared by Stephen Galicia: And while, you know, raw performance is vital and often talked about. Audio shared by Stephen Galicia: Getting high-quality data to the right place, with the right governance, security for computing, is even more critical, and that's what Phil highlighted, and why so many of these projects are failing. And this is where the intelligent data infrastructure really shines. Audio shared by Stephen Galicia: We're having so much success with our enterprise customers today, and really, it comes down to our 30 years with continued innovation. You know, NetApp has delivered Audio shared by Stephen Galicia: features and capabilities that ultimately future-proof the demands of AI workloads. So when I look at our joint solutions, yes, we're both sitting here referencing boxes and servers, so to speak. Audio shared by Stephen Galicia: But from the data and infrastructure side, you're getting a stack that allows you to be successful. And just to hit on some of these that are key. Audio shared by Stephen Galicia: You know, from the ONTAPS perspective, it's a unified data management through a single OS, and it's built to work across your entire data, giving a holistic view and access across any environment and protocol. Audio shared by Stephen Galicia: It's the data triceability of ONTAP and governance that gets you with our industry-leading snapshots and role-based access policies. And then efficient data movement. We'll talk a lot about hybrid, but whether you need high-speed caching or in a robust data replication mechanism, we're ensuring it's available Audio shared by Stephen Galicia: to the AI pipeline in a high-quality fashion. Audio shared by Stephen Galicia: And then, you know, it's not just NVIDIA and not just NetApp. There's a robust ISV ecosystem and integrations that serve up the stack for users on the application layer, and this is a critical piece that we've really, spent a lot of time and attention and engineering on. Audio shared by Stephen Galicia: That includes… includes key partners, like Domino Data Labs, Snowflake, Dremio. Of course, NVIDIA has their own software stack that we're plugged into, Kubeflow, Spark. It's very important that we are connected with that ecosystem Audio shared by Stephen Galicia: And then lastly, the hybrids with scalability and flexibility that works seamlessly across on cloud, on-prem, or colos. Audio shared by Stephen Galicia: You probably have more to hit on with the cloud piece of it. Audio shared by Stephen Galicia: Yeah, you know, our journey is really interesting, Tony. We first started working with NetApp on our DGX SuperPod initiative, and that was really focused on, and you mentioned it earlier, helping customers with that pre-training journey, optimizing the data. Audio shared by Stephen Galicia: For pre-training models. And today, what we're trying to do together is really help with the post-training, which I mentioned earlier can be really complex because data is stored Audio shared by Stephen Galicia: throughout many different environments, and managing that data, optimizing it for post-training, as well as for inferencing, can be very challenging for enterprises. And recently, the introduction of Magentic AI and reasoning models has created even a new set of challenges for enterprises. Audio shared by Stephen Galicia: In addition to just, you know, generating input-output tokens that a chatbot does, you know, from a prompt and response. Audio shared by Stephen Galicia: These new models also generate thousands of reasoning tokens over minutes and even hours of processing as they think through, you know, how to solve these complex problems. And these reasoning tokens allow for better responses, just like a person can formulate a better answer, given the time to work on that problem. Audio shared by Stephen Galicia: But they also generate an enormous amount of new data that needs to be stored and managed. Audio shared by Stephen Galicia: Similar to a person, the reasoning model can gather information and do analysis across multiple datasets, assuming it's got access to those data sets. And again, that's where NetApp comes in. How does that agent, how does that reasoning model get accessed safely Audio shared by Stephen Galicia: to that data and use that data. That's a really important part of this whole pipeline, and that's what we really want to do and build out with NetApp and our partnership. Audio shared by Stephen Galicia: Because this increase in tokens from a Gentic and a reasoning agent, you know, a reasoning AI, it requires, you know, a thousand, hundred times more compute. Audio shared by Stephen Galicia: And we optimize that compute, you know, we optimize the GPU acceleration, we bring in networking, an optimized networking solution with our Spectrum X. Our AI software, you mentioned some of it, Tony, you know, our RAG and our, Audio shared by Stephen Galicia: AI reasoning models. Audio shared by Stephen Galicia: But this is only part of the stack. You know, the other part of optimization for this to be really an enterprise-grade solution is that it needs to have a new level of governance and supervision for these pipelines to really be effective in the enterprise. That's where I think our work integrating ONTAP Audio shared by Stephen Galicia: has been really impactful and important, because with NetApp, we're able to provide secure data exploration across these different locations, across cloud and on-prem. Audio shared by Stephen Galicia: And with NetApp's support for first-party services in all these public clouds, users are also able to get a unified experience. So these reasoning agents can access data from different locations, they can rationalize that data, they can validate its accuracy, they can dedupe it. Audio shared by Stephen Galicia: For efficient data processing, and then make sure thatdata pipeline is honoring the data governance policies of the enterprise in which it's operating. Audio shared by Stephen Galicia: Yeah, no, absolutely. Now, let's jump to addressing AI-ready data in a hybrid cloud environment. So this is one that's fast-moving, evolving. What are you seeing with customers as they look for different landing spots for their compute? Audio shared by Stephen Galicia: Yeah, this is such an interesting area that's just exploding right now. You know, every enterprise Audio shared by Stephen Galicia: I talk to has compute running in multiple environments. On-prem, cloud. Oftentimes, they have more than one hyperscaler cloud partner that they're working with, and they're also starting to look at what we call the neoclouds. Audio shared by Stephen Galicia: And that's a really interesting new ecosystem that's developing really worldwide. And what's prompting this in some regions is really partnerships between public Audio shared by Stephen Galicia: and private sector, as governments recognize the importance of AI for their own countries and for sovereign initiatives locally. And these initiatives include investment in making AI infrastructure available within a region, within a country, as well as training models. Audio shared by Stephen Galicia: So, for example, many, you know, regions are investing in developing and post-training for models Audio shared by Stephen Galicia: to honor local language, even dialects. Some regions, it's hundreds of models where you have many,dialects across, you know, a very diverse population. Audio shared by Stephen Galicia: And that's just really exciting to see, because it means that AI is going to be very accessible to literally everyone. It's going to talk everyone's language, which means it can really serve everyone. And I know that's NetApp's goal as well. Audio shared by Stephen Galicia: And of course, data infrastructure and management will be critical to all of this, as neoclouds are built out. And that's where we're also partnering. Audio shared by Stephen Galicia: with NetApp is making sure that the data provenance, data sovereignty, security, access is also available in these neoclouds, because I know you're working with many of them as well to stand up storage infrastructure. Audio shared by Stephen Galicia: Yeah, no, absolutely agree with all that, and yeah, we're working with several NCPs and customers who are shifting in this direction, and it's really been an exciting one, both for the sovereign pieces and the capabilities that we have, but also just to simplify the process of moving the data to the neoclouds, and even back to maybe what either Audio shared by Stephen Galicia: 1P Cloud Services or, on-prem. Audio shared by Stephen Galicia: That's right. Yeah, exactly. With the majority of enterprises, they're using NetApp today, and, you know, we view the Neoclouds as really an extension to our solution, and giving them, you know, the capabilities and securities that they have on-prem, and that they're accustomed to. Audio shared by Stephen Galicia: And the three hyperscalers, things like the robust multi-tenancy capabilities, seamless and efficient data movement, and we have the ability with the SnapMe or FlexCache to move it to and from. Audio shared by Stephen Galicia: And then, of course, the security element, our ransomware protection and overall built-in security. They're able to leverage all these, no matter where that environment is, so… Audio shared by Stephen Galicia: With ONTAP deployed on-premise, in the clouds, and now the neoclouds, customers really have these multiple landing spots for their compute needs, and they can really take their data to where the AI and compute needs to go and happen, and not hinder their AI workflow overall. Audio shared by Stephen Galicia: Let's jump to the last piece, because this is one that's exciting, and ensuring the data quality, security, governance of the AI pipeline. I think it really hits on, you know, the NVIDIA data, AI data platform, but really the convergence that's happening between the compute and storage layer, how that's just getting tighter, and it's going to open up new doors for the enterprise. Audio shared by Stephen Galicia: Yeah, that's right, Tony, and this is one of the areas where Audio shared by Stephen Galicia: we're working on now, and really, it's going to start really paying off and delivering in the near future here. So, earlier this year, we published a, it's a reference design, it's customizable for this fully integrated AI data platform. And it brings together enterprise storage systems. Audio shared by Stephen Galicia: As well as NVIDIA-accelerated computing, networking, and software and AI models to power AI agents capable of reasoning. Audio shared by Stephen Galicia: with real-time business insights using a RAG architecture, retrieval augmented Architecture. And so these turnkey platforms enable organizations are important because they're able to tap into all this unstructured data that's throughout the enterprise, you know, and unstructured data is, you know, residing in all the reports and all Audio shared by Stephen Galicia: Product manuals, instruction manuals, presentations, just like this one. Audio shared by Stephen Galicia: That serve as the knowledge and the backbone of an enterprise's knowledge center. Audio shared by Stephen Galicia: These integrated stocks, you know, integrated offerings, are built on our Blackwell platform, RTX Pro servers in particular. They… they include our Nemo Retriever models, which you mentioned before, which are really state-of-the-art RAG, and embedding and retrieval and ingestion. Audio shared by Stephen Galicia: For… for really great performance and accuracy. Audio shared by Stephen Galicia: of these, you know, reasoning chatbots. They embed and index the structured data, making it instantly searchable through a vector search interface that can be accessed by an end user, a person, as well as additional agents that might be, you know, part of a bigger agentic workflow. Audio shared by Stephen Galicia: So, we're partnering with NetApp to bring this to market and build out, you know, a full AI data platform on the NetApp platform that will integrate ONTAP to bring, really, AI in a really streamlined way, in a very secure way, to enterprises. It's really exciting. Audio shared by Stephen Galicia: Yeah, no, this is an exciting one. We'll leverage that mature, data management of ONTAP, and then, of course, NVIDIA NEM servers, services, Nemo, and all the other software applications to tap into, but really providing a nearly turnkey experience. Audio shared by Stephen Galicia: From the metadata management to the vector DVs, it's going to ensure high data quality, and utilizing things like NetApp SnapMirror and SnapDiff technology, it will not only move the data into REG and Agentic pipelines. Audio shared by Stephen Galicia: addressing some of the data quality issues, but it will also keep that data updated instantly. So it's going to really change the game for customers, and there's really 3 areas, I think, to highlight for me on this is… Audio shared by Stephen Galicia: The accessibility of data, whether directly or through an API into the AI workflow, will become instant. And not only will the accessibility be instant, but the understanding of what they're accessing Audio shared by Stephen Galicia: We'll, while keeping it updated and fresh, real time. Audio shared by Stephen Galicia: The second one is the governance piece, another critical element that will come into play by way of easily setting policies, anonymizing the data, and making sure data scientists who should or should not have access to this data or use it as part of the pipeline, they're using it in the appropriate way. And it's really just a set-it-and-leave-it manner. They're not having to go back and Audio shared by Stephen Galicia: you know, ask for permission. It's already set in place. And then lastly, on security. This is all happening now at the storage layer, so they're fully taking advantage of the industry-leading features that ONTAP delivers today. Audio shared by Stephen Galicia: And I have to say, this is really probably one of the exciting ones, to just deepen the connection between compute and storage, and I really cannot wait to see how customers are able to benefit here. Audio shared by Stephen Galicia: Yeah, I agree. It's exciting. Lots of great things in the hopper for enterprises and our customers to take advantage of. Oh, you've got one to talk about right now. Audio shared by Stephen Galicia: Yes, I do, here it is. Yes, so let's talk to a joint customer win story, and this one really captures a lot of what we discussed today, where you have a customer, Audio shared by Stephen Galicia: who started in the public cloud with a lot of their AI initiatives, but required more compute and more data. And a lot of their data that they needed was actually sitting on-prem already. So, ultimately, when they started this project, it was… the data was everywhere, and they needed to deploy Audio shared by Stephen Galicia: at the end of the day, what turned out to be a hybrid AI factory that worked across different applications in the cloud, on-prem, and environments in a seamless way. So. Audio shared by Stephen Galicia: It was a pharmaceutical, international pharmaceutical company, and if we just walk through the win here, you know, the desired business outcome is to utilize AI to reduce their costs of getting drugs to market, and also increase their probability. And for those who don't know. Audio shared by Stephen Galicia: Typically, to introduce a new drug to market takes somewhere around 10 to 12 years, and close to $2.5 billion spend, so there's a lot of opportunity to shorten that cycle, as well as reduce costs. Audio shared by Stephen Galicia: There's technical challenges. There's long life cycles, poor visibility, and then data is everywhere. Again, that idea to want to Audio shared by Stephen Galicia: leverage some of the cloud-native applications they were using, but also look at on-prem, whether it be for security reasons or different data sources they have there. Their desired technical outcome was delivering a seamless hybrid AI environment. Audio shared by Stephen Galicia: And, you know, where NetApp and NVIDIA are critical to the success is Audio shared by Stephen Galicia: Again, having different landing spots for the compute and data to go. Audio shared by Stephen Galicia: So really, delivering services that spanned all environments was critical to this. Audio shared by Stephen Galicia: And also, you know, with these projects, it spans… they're complex. It spans multiple personas, multiple… Audio shared by Stephen Galicia: BEUs that need to be aligned when they want to deploy these. So, across this part of it, you have IT, the data engineers, data scientists, of course. Audio shared by Stephen Galicia: chief product officers, data protection officers, and then all this, for this specific case, was sponsored by the chief data officer. So a lot of moving parts, a lot of people to bring together. Now, what was exciting is the outcome. This customer has successfully deployed an AI factory Audio shared by Stephen Galicia: That's standardizing on ONTAP, spanning both on-premise and cloud environments. Audio shared by Stephen Galicia: They utilize the NVIDIA Clara application frameworks. Their core training is taking place on-prem, but of course, again, they're able to leverage that cloud with bursting for applications, the ability to move that data back and forth seamlessly, and they use this by leveraging NetApp Flex Cache technologies to move data as needed. Audio shared by Stephen Galicia: You know, ultimately, one of the big things is they've reduced Audio shared by Stephen Galicia: their OPEX costs by really making things more efficient in how they're using the compute and the data. They've increased visibility to the data, and of course, they're actively reducing their time through AI to develop the drugs. Audio shared by Stephen Galicia: By bringing AI into the workflow, ultimately. Audio shared by Stephen Galicia: This is a great story, Tony, and I love the,framing up of it as an AI factory. At NVIDIA, we talk about AI factory as really a North Star, like a center of excellence for how an enterprise is doing. Audio shared by Stephen Galicia: is doing AI. And, we… obviously, we talk about AI infrastructure as well, and AI infrastructure, and at least in the way I… I think about it, is it's all of the compute, the networking. Audio shared by Stephen Galicia: the systems to make AI possible, but you don't really have a factory until you're producing results, and you have to have data to have that. Audio shared by Stephen Galicia: And that's really critical for, I think, customers and enterprises to understand, is that data component of their AI journey is just so important. Otherwise, you're just building out infrastructure. You're investing in infra, but you're never going to generate results unless you also look at the Audio shared by Stephen Galicia: You know, optimizing the whole data platform, the data pipeline, and then, obviously, then you can generate, actually, results in business Audio shared by Stephen Galicia: Business, you know, return on investment by generating tokens and reports and insights that you never had before. Audio shared by Stephen Galicia: So, clearly, you know, that's an example of this customer, what they did, and it's because we have this sort of unified, you know, I think, version or vision of how we deploy our technology. This customer is on the cloud and on-prem. Audio shared by Stephen Galicia: So is NVIDIA stack, so we pres… our full stack is available on hyperscalers, neoclouds, and on-prem. Same with ONTAP, same with NetApp. So, it makes it really easy for our customers to meet us wherever they are, and have a full integrated stack, bringing their data. Audio shared by Stephen Galicia: an optimized AI platform with NVIDIA to really take advantage of this sort of AI era, this AI revolution that's happening. Audio shared by Stephen Galicia: Yes, absolutely. And with that, Tom, I will turn it over to you. Audio shared by Stephen Galicia: Hey, so… Audio shared by Stephen Galicia: Man, I was just loving the conversation. Phil, incredible insights on the research that you provided, and I think really raising awareness on all it takes Audio shared by Stephen Galicia: To really create a data infrastructure that considers All the steps. Audio shared by Stephen Galicia: the data needs to go through in its journey to make AI effective. Audio shared by Stephen Galicia: And then, just the solutions description and what's coming, you know, from NetApp and NVIDIA, you know, the closeness of the partnership in solving the data challenges and… Audio shared by Stephen Galicia: that are getting in the way of people getting to production business value. And I just want to kind of sum up what I think I heard, and talk about what you can expect from NetApp and NVIDIA around data infrastructure for AI going forward. Audio shared by Stephen Galicia: we're certainly going to be focusing on performance. Performance… Audio shared by Stephen Galicia: Both on-premise, in the cloud, everywhere that you do your AI projects. There'll be a lot of that coming at you. But really, I hope you heard the data story here, that what you can expect from NetApp Audio shared by Stephen Galicia: Around the data is more visibility and understanding and access to data across your state. Audio shared by Stephen Galicia: A focus on the capabilities, as Phil mentioned, on data currency and maintaining that single source of truth. Audio shared by Stephen Galicia: And then, specifically, data services focused on security and governance for the AI workflow. Audio shared by Stephen Galicia: Look for that coming from NetApp going forward. It's really all about getting enterprises to production business value with their AI investments, and we realize that data quality is at the core of that. Audio shared by Stephen Galicia: And so with that, I want to point you to a couple of resources. We're going to put some links in chat. If you want to dig deeper on Phil's research on AI-ready data storage infrastructure, I think you'll find that fascinating. Audio shared by Stephen Galicia: The link will be in the chat. If you want to go deeper on NetApp NVIDIA solutions, please Audio shared by Stephen Galicia: Check out the second link. Audio shared by Stephen Galicia: And then finally, we will be making some really exciting announcements. Audio shared by Stephen Galicia: around our solutions and ecosystem, as it is going to evolve, and you can find out all about that at NetApp's annual Insight Conference. It'll be coming up here in October, and we invite you to either attend or tune in, and you can see the link there. So I want to thank again the panelists. Audio shared by Stephen Galicia: Tony, Ann, and Phil, and I thank everyone that has joined us today. Audio shared by Stephen Galicia: Thank you. It was a lot of fun, guys. Yeah, thank you. And Ann, Phil, you'll be at Insight, right?That's right, see you all there. All right, love it. You bet. And now we'd like to bring our audience into the conversation. So we're going to take a couple of questions that you've been piping in over Q&A. So let's get to it. Tom Shields: Okay, do we have everybody? Tony Chidiac: I'm here. Tom Shields: All right, cool. Great. We got the gang back, and we got a few questions that I wanted to ask in front of everyone, and just for all's benefit, for folks that are sticking around. So this first one is for Ann, and it's coming in saying, hey, NVIDIA partners with everyone. We know they're good at that. What's the difference about the NetApp partnership? Anne Hecht: Sure, Steve, great question. You know, and I think, Tony, you hit on a lot of the key points, and I think the core competency and the value you bring to your enterprise that we also appreciate, and our joint customers appreciate, around data quality and the tooling you've done there to… to make sure that Anne Hecht: you know, that processing of that data from ingestion to delivery, to an AI model to an agent is a high-quality, accurate result. It's, you know, incredibly important work, obviously, in the era of AI. Anne Hecht: The governance of that for an enterprise is really important, because there's a whole new set of, you know, policy, supervision. Anne Hecht: transparency that needs to be put in place to make sure that AI is trustworthy. And with 30 years, over 30 years now, right, of working with enterprises, I think NetApp really understands Anne Hecht: what those enterprise requirements are, to make sure that the AI is delivering quality results and not, you know, potentially adding risk to the business. And then, of course, there's the overall security, that you bring to the whole… to the pipeline and of the data and the privacy. Anne Hecht: So, our teams work really well together. We've done, integration work for, like, almost the last, I think, almost 10 years now, starting with… with trying to deliver, like, results, like I said earlier, on the area of training. And,so that partnership is really strong and deep. Tom Shields: Thank you very much for that, Ann. Really appreciate that. So, here's one more. What's the one recommendation you have to get past some of the pitfalls that companies are encountering in making their data work for them in production environments? What sticks out the most? Who wants to take this? Tony, you want to… Tony Chidiac: Yep,I'll go ahead. I think one of the big ones is… I mean, all the AI, everything that the enterprises are doing at the end of the day is to get to a business outcome that's profitable. So one of the biggest pieces to get there, okay, what's the objective? But then two. Tony Chidiac: what proprietary data do you have in place to make this happen? And in order to do that, you need to really treat your data Tony Chidiac: like a product, and part of our big design thinking is that single data catalog to generate and have that access and visibility as to what your data is, and to keep it fresh. So I think the data component is huge, and then, you know, now implementing it at the actual storage layer gives you Tony Chidiac: Different value that's really going to help feed the ecosystem, like the granular metadata management pieces, keeping it real-time updates so your data is not getting stale. Tony Chidiac: the integration that comes at the storage features, and then, of course, security and governance. I hit on, you know, just certain projects being stalled because they don't meet the security and governance. So being able to do this at the storage layer is going to be really critical to ultimately get to those business outcomes that BU's and Tony Chidiac: chief data officers, where they're trying to go and build out, AI in the enterprise. Tom Shields: Cool, thanks, Tony. And I just want to open that up to some of the other panelists here. So, you know, these catalogs exist today, and we're talking about doing it at the storage layer. Having a data catalog at the storage layer is part of our design thinking. Tom Shields: But Phil, I'd love to get your impression. Now, why is having it at the data storage layer an important thing and useful to Tom Shields: Customers in getting… getting their projects over the finish line. Phillip Goodwin: Yeah, I think the key thing there, Tom, is that there are certain things you can see at the infrastructure layer, at the storage layer that you can't see at the application layer. And to feed on what Tony was talking about, you know, we think the only way to really break down those silos and get to that single source of truth Phillip Goodwin: is to address, the data at the storage layer. Being able to do the data classification things, being able to Phillip Goodwin: provide the appropriate governance and security and things like that. But to be able to sort out where all those different pieces of data are, whether they're in the cloud or on-premises, or wherever they might happen to be, so that you can get that correct copy of data and that single source of truth. Tom Shields: Great,Do you have any thoughts on this, Ann? Anne Hecht: Yeah, I mean, I think, the… the opportunity is there to accelerate this whole pipeline, because, as Phil and Tony pointed out, there's Anne Hecht: These AI models, the reasoning agents, whether you're doing post-training or actually in production, you need an immense amount of data, and that data is moving across your networks, and that's where I think acceleration Anne Hecht: comes in, and when applied in a really efficient manner, you… you can actually lower your cost of operating that AI pipeline, because you're adding efficiency to it, as well as providing faster response times to the agent or the end user who's querying on that data stack. Tom Shields: Cool. Well, we've got time for one more question, and what I want to do is revisit something that came in during the session here earlier, and it was answered, but I want to go Tom Shields: broader with it. The question is, what are some of the essential capabilities for a storage system to be qualified to have embedded AI? And so, Phil, I'd love to get your perspective first, because, you know, this was definitely one of the capabilities that you asked Tom Shields: or recommended that, teams look at if they were evaluating a data storage infrastructure. Yeah, I think one of the… Phillip Goodwin: Yeah, sure, Tom. I think one of the huge parts of that is Phillip Goodwin: really, efficiency of the IT staff, and, and… Phillip Goodwin: problems dealing with lack of skills, or simply not having enough people and things like that is something that's plagued IT since the beginning of time. Phillip Goodwin: So, being able to do things like automated provisioning, troubleshooting, problem resolution, threat detection and response, you know, being able to do those things Phillip Goodwin: in real time through Agentec AI, I think are some of the things that would differentiate, AI-ready infrastructure from, more legacy infrastructure. Tom Shields: Good, thank you for that. And, Tony, you want to add anything? Tom Shields: I guess not. Ann? You? Anne Hecht: No, I think that's great. I mean, it's, I think Phil nailed it. Tom Shields: All right, well, we'll leave it there. That's about all the time we have, so, we hope to see you all, either tuning in or joining us at Insight Conference, where we'll take a deeper dive on how NetApp can help customers with their AI data challenges. Tom Shields: to get their projects to production business value faster and more consistently. Thanks again, everybody.