Welcome, everyone. Thanks for taking the time to join the session today. We'll get started in just a couple minutes. Again, we're gonna wait a minute or two to get some more folks onboarded. During the session, if you want to submit questions throughout, please use the Q&A feature. We'll do our best to answer those questions throughout the presentation, and then we will have a dedicated Q&A section at the end. Again, we'll get started in just a couple of minutes. Appreciate your patience as we wait to begin. Welcome, everyone. Thanks for taking the time to join the session. We'll get started in just a minute. Throughout the session, if you have questions, please submit them using the Q&A feature. We'll do our best to answer during the presentation, and then we'll have a dedicated Q&A section at the end.
Again, we'll use the Q&A feature to submit those questions for the live Q&A. Thanks for your patience. We'll get started in just a moment.
Recording in progress.
Hello, and welcome to today's webcast, titled The AI Era. I'm Russell Fishman, Senior Director of Product Management for NetApp Solutions, and I'll be your host for the session today. We have an awesome agenda that I believe you'll find extremely valuable in continuing to grow your understanding of how the most successful organizations are adopting AI, and what lessons you can apply to your own journeys. To take us through this, I'm extremely privileged to be joined by experts from a number of leading AI companies, including NetApp, NVIDIA, and IDC. With me today are Jonsi Stefansson, Senior Vice President and Chief Technical Officer at NetApp, Ritu Jyoti, Group Vice President of Worldwide AI and Automation Research at IDC, and Tony Paikeday, Senior Director of Marketing at NVIDIA. Ritu and Jonsi will discuss results and key takeaways from IDC's recent AI Transformation Study.
Then Tony and I will then spend some time covering our latest joint solutions announcements, and we'll wrap up with some Q&A. So let me start by introducing the study and associated report that we'll be discussing today, Scaling AI Initiatives Responsibly: The Critical Role of an Intelligent Data Infrastructure. To scale AI initiatives responsibly, organizations need an intelligent data infrastructure. This intelligent data infrastructure must have the flexibility to access any data anywhere, with active data management to enable superior data security, protection, and governance, and adaptive operations to maximize performance and efficiency of their infrastructure and applications, all while optimizing cost and sustainability. Taken together, these can maximize AI knowledge, worker productivity, and propel organizations to more consistent success as they use AI to achieve more for their businesses.
With that, let me hand over to Ritu and Jonsi to take us through the key takeaways from the report.
Thank you, Russell. It's wonderful to be a part of this discussion today. We all have witnessed Cambrian explosion of generative AI technologies in the last 12-18 months, and today we can easily say that generative AI is poised to be one of the greatest disruptors to impact business and society. With all these exponential changes in technology, a business environment seems like metaphorically, a fluid, unpredictable, raging ocean. In fact, as per IDC research, 37.4% of the respondents have reported that generative AI will disrupt their competitive position. So it begs the question that what exactly is happening? And who could have answered it better than Larry Summers? I love his response.
He says: "AI is coming for the cognitive class, and there is a substantial chance that AI will be a threat to IQ versus EQ." So essentially what he's saying is that this is going to impact every knowledge worker and thereby every organization on this planet. We at IDC, as per our AI Spending Guide, we are predicting that the worldwide AI spend will exceed $512 billion, and Gen AI is a critical catalyst to drive this exponential growth. What we are seeing is that the use of AI and machine learning is occurring in a wide range of solutions and applications, be it your ERP solutions, the manufacturing software, content management, collaboration, end user productivity. It's critical to notice that IDC expects that AI will be the most disruptive influence, changing entire industries over the next decade and for future.
But the most interesting point that I'd like to zero on, and is that the manufacturing software, content management, collaboration, end user productivity. And it's critical to notice that IDC expects that AI will be the most disruptive influence, changing entire industries over the next decade and for future. But the most interesting point that I'd like to zero on in is that the changing landscape of generative AI market opportunity. As I noted, Gen AI is going to be a big catalyst for the overall AI adoption. But when you kind of take a look into the finer details as to how the Gen AI market is evolving, in the earlier days in 2023 and 2024, a huge chunk of investment is going towards infrastructure.
But the real meat and the value that an end user will be able to kind of accomplish will drive the forecast towards apps, platforms, and services. So given all of these factors that are going in the industry, in December of 2023 and January of 2024, IDC conducted both qualitative and quantitative research with global AI decision-makers to assess the state of AI initiatives at their organizations. And as part of the study, we actually looked into 51 distinct inputs to assess the maturity of the organizations, be it AI strategy, whether they're you know kind of ready from a data perspective, or whether they were ready from the infrastructure perspective, whether it's the storage latency, the scalability, the access, the movement of the data from a distributed environments, or whether it's data governance and related processes. Security, as you and I know, we all...
That security is top of everyone's mind, and so whether it is data security or data privacy, how efficient are they in handling their datasets? Cost. Cost is the number one inhibitor, so we also kind of poked around in terms of cost efficiencies, storage optimization, storage right sizing, processes, and tooling for data science as well as developer productivity. And as part of this study, at the end of it, we actually classified them into four different stages of maturity, as you can see on the slide, right from AI Emerging to AI Masters. And our assessment kind of gave us a very interesting insight that it's kind of equally split between AI Emergents and AI Pioneers. When you combine the two, they're almost 50/50 split, but there's a very small percentage of the organizations who are at the master stage.
As we kind of start looking into what are the reasons the AI failures are caused, and the number one reason, be it an AI for an AI emergent or an AI master, it's actually cost-
Sorry, folks, the technical difficulties. Getting back up and running now.
You can easily see that the percentage of improvements needed, whether it is minor or major or a complete overhaul, there's a stark contrast between the masters and the emergents. And the point that we were trying to stress on is, like, the masters have been on this journey for a longer duration of time, and we always hear it from the end user, that those who have been on the AI journey, especially some of the organizations in financial services or retail or manufacturing and healthcare, they are better prepared to kind of harp on to the game-changing, transformative, you know, opportunities with Gen AI. So as you can see from this particular chart, there is some very interesting insight that organizations are not dealing with data which is stored in one particular you know, location.
It is actually distributed across different data locations, you know, be it on cloud, it can be different types of data, and especially hybrid and, you know, private cloud scenarios kind of shape up. So ability to integrate the organization's private data with the data that is stored on public cloud, seamlessly integrate, and especially in the case of when we are talking about Gen AI, you know, organizations are working towards contextualizing the data with their own private data, and the technique of RAG is really, really front and center for all of that. In all of these cases, the underlying data architecture readiness is critical. I would love to hear from Jonsi. I'd like to invite him to get his perspectives on what he is hearing from his customers and his, his point of view from a NetApp perspective.
Yeah, I fundamentally believe that AI workload is going to stay hybrid forever. You know, companies will have to be able to freely migrate and replicate data to the public cloud or a service provider of choice when it's needed. That's why we've actually put a lot of work in integrating our first-party storage offerings with native services in the public clouds, like Vertex AI and Google, like SageMaker and Bedrock and AWS, like Azure AI services for Azure, of course, and we have our own AI toolkit that allows you to basically chat directly to... with your data. And honestly, nobody does data management like NetApp. And that's the number one reason why all three of the hyperscalers chose NetApp as the world's first external native storage offering in the public clouds. It's for the data management capabilities and ease of migration....
Our integrations with NVIDIA NeMo Retriever is basically RAG enabling every single deployment of ONTAP, whether it's on-premise or our first-party storage offerings with the hyperscalers. You can really take proper advantage of your hybrid cloud or multi-cloud strategy for AI. I mean, ultimately, your trained model is only as good as your data.
Yeah. Thank you, Jonsi . See, I couldn't agree more on that because there's no AI without data. And, you know, many organizations that I speak to, they kind of complain about that, you know, they have been working on data issues for the last decade or so, but it's still an ongoing problem, and it becomes really, really critical in the age of AI, when AI is becoming more mainstream. So with that, let's go to the next slide. So the second imperative that we are kind of looking at for an organization to scale AI initiatives responsibly is to actually become responsible at the core. Every organization needs to define their principles that they need to adhere to.
They need to have an AI governance committee that can actually ensure that they're being adhered to, and also every individual in the organization needs to be educated and made aware of what the critical requirements or principles are. In addition to that, they need to have the right sets of tools and technologies, because responsible AI is not just at one part of the stack, it kind of does through the entire AI life cycle, right from the infrastructure to the model to the application and the end user layers. Every organization needs to understand that this is not a destination, but a journey, and they need to kind of continue to learn from their experiences, iterate upon it, and continue to enhance and improve their responsible adherence.
So if we kind of, you know, dug further into the insights that we got from the, you know, the survey that we are talking about here, one thing that stood out for us is that if you take a look at this right side of the chart here, that the standardized policies are in place and rigorously enforced by independent group within the organization. And I want to underline that rigorously enforced. It's not enough to just have policies and principles, but also kind of, you know, the right set of tools and technologies to kind of observe it and do this on an ongoing process. That's a stark contrast between what AI masters are and in terms of AI emergents.
If you look at the AI emergents, they're still in the process of developing policies and procedures are being developed, but they don't have the standardized processes to put them into enforcement, and that's the key difference. All policies must be standardized and also constantly observed and monitored and improved upon. When you look into, you know, some of the critical or, you know, aspects when we talk about data security and privacy, because it's really front and center, and we started the conversation about that the most of the failures are happening because of the data issues. I set the stage on how organizations need to become responsible at the core.
The important aspects to be looked into via AI master versus AI emergent, they are way more advanced in terms of mitigating bias or data sovereignty concerns, and likewise, when it comes to data security and privacy. Let me spend a couple of minutes on talking about what are the best practices that the AI masters have figured out vis-a-vis the AI emergence. Essentially, first of all, when you're talking about bias or data sovereignty, they really need to know and put in the right set of data, prepare the data. Depending upon the use case that they are working for, they use the right set of tools and technologies to create the right balance of the data. They might be using synthetic data in some cases.
If there are some data sovereignty rules, you know, Federated Learning could be put into place so that you do not really kind of move the data outside the geographies. When it comes to data security and privacy, it's extremely critical, you know, especially in the era of GenAI, everyone's talking about your data is your data. It should not be in, in, you know, kind of a IP leakage. Organizations should be asked for their permission if they're willing to kind of share their data, to kind of, you know, use it for the further training of the model. So those checks and balances are very strictly adhered to, and they are making sure that all the regulatory policies are kind of also included.
If you recall the slide that I was sharing about responsible at the core, that's what I was talking about, because the regulatory policies are evolving. They really have the right sets of tools and technologies to make sure that these are in place. But at the same time, they're also kind of making sure that they are partnering with the right technology suppliers. And these technology suppliers need to have the right responsible principles and the responsible guardrails in order to give them the sense of confidence.
There is a balance between speed and good governance. In the long run, clear guardrails and knowing your data will set you free to innovate and succeed with AI. Your data today is the number one attack surface for hackers, and we are constantly innovating in this space and being able to train our own models to essentially monitor itself and have an automatic reaction to any ab normalities, is the only way to go, and can only be achieved with AI. Because a human is very unlikely to be able to catch it in time, let alone fix it before the damage is done. And that's why we are very confident in offering our ransomware guarantee to our customer. And the other thing is, you know, you really need to map out your data. I mean, that's a key...
That's a key aspect of being able to take advantage of your AI hybrid or multi-cloud strategy. You might have PII, HIPAA, GDPR, PCI or GDPR data that can't just be uploaded to any LLM, you know, what needs to be obfuscated before or prepped before, and so on and so forth. So, I mean, the only way to really, you know, speed up your AI journey or your AI advancements is to know your data, map your data, and secure your data.
Well said, Jonsi. I think, you know, all these tools are really, really going to be game-changing in enabling the efficiency, the tech acceleration of time to market. Organizations, you know, especially the masters, they're way more advanced in working with, you know, superior solutions, provided by folks like you all. So with that, next slide, please. So now we come to the next, the third imperative for organizations to scale their AI initiatives, and that is all about resource efficiency. So as AI workflows become increasingly integral to the various industries, it is important for us to acknowledge the implication on compute and storage infrastructure, data and energy resources, and their associated costs.
Here we have a very interesting insight from our IDC research, where we are looking into that the AI data center energy consumption, they will grow at 44.75% CAGR from 23 TWh in 2022 to 146 TWh in 2027. That's pretty alarming. We really need to come together as a, as an industry to look into what are the ways and how we can do better optimization and better utilization of the resources. We are not saying that, you know, you need to limit the user GPU, but we have to consider the most resource-efficient technique for your use case and make pragmatic choices.
We also have to look into better management of the infrastructure needed to support AI and GenAI, and while focusing on timely business outcomes, but then we have to look into how do we do the cost optimization, end-to-end, you know, performance optimization, energy utilization, while not forgetting that there's scarcity of GPUs currently. So when you think about resource efficiency metrics that needs to be standardized, when we asked in this particular survey that how they are kind of looking into whether they have clearly defined metrics for assessing the efficiency of resource, the answer is very, very startling. Because if you think about the masters, they have completed and standardized it across AI projects, whereas the AI emergents, only 9% of them have completed. And in terms of not starting yet, the number is very, very high.
That's why we are calling out and sharing that these best practices from the AI masters as to how they can actually look into. You know, first of all, you always, you can monitor something, you can get better at something only when you start kind of recording as to what really are the important metrics that you have to go after. That difference can be clearly seen here between the AI Emergents and AI master. So with... I'd like to invite Jonsi for his comments on what he's hearing from his side of the house.
You know, the report actually calls it out, just like every single customer meeting. I mean, every AI is on the top of the agenda for any meeting you go to these days. You know, the cloud conferences are no longer cloud conferences, they are AI conferences. You know, it was amazing to see NVIDIA at GTC, and you met basically everybody in the industry there. But the funny thing is, everybody seems to be sort of dead set on creating more silos instead of going for standardizations. They should be able to use the same technology, same data management capabilities, as for other workloads. I mean, for example, you can't just sacrifice, you know, your standards, your governance, your compliancy for just speed. Speed alone is not gonna make you go faster.
What makes you go faster is knowing your environment, having standards like the masters do, and this is where the emergents are actually struggling. And we see that, with companies that we are helping, you know, get them started. So from my perspective, standardization is key. You know, and another important thing is, you know, the data unification. Choose a vendor that can offer you all the storage protocols, you know, used for any workloads, not just AI. I mean, having block, file, object, all managed by a single pane of glass or all having the same. And being able to access different, you know, like, file and object duality, is key for integrations with all the MLOps platforms out there.
Because you can't just expect them to, you know, this only works for file, this is only-- you only need block here, and, object is key for, you know, archive. But, you know, like, NetApp helps our customers become the data pipeline. The data pipeline for AI allows you for different phases within your AI journey. You need different performance metrics, or characteristics. So, we can basically call and move your data to high-performant controllers when it's needed and move it to a cost-efficient, object storage when you, when you don't need to be training or fine-tuning your model. Or, like, Nemo Retriever, you can actually do, like, ingest in place. So, like, collaborating and working with ISVs that are all in this space is also a key.
You know, you need to be able to play with everybody, basically. But don't sacrifice your standards.
Yeah. I couldn't agree more with you on that, Janci, because it's... You know, when I speak to the end users, they typically tell me that they have been spending so much time in kind of moving things and, you know, incomplete connections between the different disparate systems, and that's really, really kind of slowing them down. So very well said. Building on to this resource efficiency, a topic we are just kind of, you know, talking about right now, it's very, very critical that we actually have build, buy, tune, kind of a discussion as well, because this has a critical impact on your resource efficiency. You know, being in the industry for a long time and watching this whole AI, the traditional AI, explosion, I've seen that intuitively, many organizations, they take a lot of pride in going and building things themselves.
where the market is kind of, you know, rapidly evolving, the pace of innovation in GenAI, it's going to be impossible if organizations think about doing build for everything. So make sure that you actually have the right set of decision-making, and its impact on what kind of decision you're making, whether it's on your return of investment. Do you really have the right skill set? Do you have the right data sets? Do you have the right resources? Is it really kind of a game-changing, truly differentiated, or is it a nice-to-have kind of a scenario? Do you have, you know, the area where you can think about whether you're going to really move the needles extensively by doing something, by building yourself?
But in majority of the cases, you might be good enough by fine-tuning or contextualizing it and automated prompt engineering. But there are scenarios in which you kind of take something off the shelf and actually tune it, or to kind of better cater to your unique needs. So please keep in mind that this decision will actually have a direct impact on your resource efficiencies as well, so try to simplify the decision. So building on to that, you know, it's important to kind of think about that, you know, when you are looking into all of these decisions, you really have to focus on improvements. What kind of improvements you need to get to the right size storage for AI, right?
It could be, you know, that, consider the storage that can keep your GPUs utilized and burst into cloud for GPU access, because we've been talking about, hybrid scenario to be very, very kind of, you know, prevalent. You could be thinking of looking into storage that can be data and energy efficient. It can unify data and reduce data silos. That is one of the biggest problems that I hear from all the end users. And also develop best practices around mitigating the, you know, what tiers of storage you are using, how you're kind of seamlessly moving it. Where applicable, you can use federated learning with all the right data security and privacy requirements. So, it's... It has a lot of interesting insights as to how you can actually optimize your storage to drive better resource efficiency for AI initiatives.
I would love to hear from you, Jonsi, as to what are you hearing from your side of the house?
We, of course, optimize our storage for NVIDIA GPUs, and we are laser-focused on maximizing that performance and utilization for our customers that are, you know, our joint customers. But, I mean, ultimately, like I said earlier, fundamentally, AI workloads are going to be hybrid. And, for model trainings, some customers just simply don't have the power budget to support, you know, model training in their own data centers, or even potentially don't have the need, need the GPUs available for that particular task. So you need to be able to burst and connect into, like, NVIDIA DGX Cloud. And that's sort of the design principle of the concept that we call intelligent data infrastructure. You know, to be everywhere where data lives, protect and securely deploy data wherever the customer wants it to be.
We are constantly striving for sustainability. You know, less power consumption for our controllers, less carbon footprint. You know, data or storage efficiency within our own portfolio is of super importance, like compression, de-duplication and all of that. That all lowers the or increases the efficiency. And we'll do our part to facilitate the, you know, or to get AI done as efficiently, considering performance, cost, and sustainability. That's part of our mission statement.
Thank you, Jonsi. As we can all agree with this, that the right sizing will also mean a reduction in data silos and incorporation of-
Mm-hmm.
Very comprehensive, hybrid, unified and multi-cloud storage approaches, and giving the organizations the flexibility that they need. So we have had such interesting discussions about all the different imperatives that an organization need to kind of, you know, scale their AI initiatives. I'd like to close with this very interesting data point here. So if you think about here, that the, we are really interested. All the AI initiatives is all being driven to drive superior business outcomes and kind of, you know, organizations to unlock their defensible moats and continue to thrive. When you're thinking about this, if you take a look at the data point here between AI master and AI emergence, AI masters, they experience the lowest twelve-month improvement in business outcomes. So it begs the question that, why is it? The reason is because they have accumulated benefits from prior stages of their journey.
They're working in way more complex initiatives. So it can be a little bit, you know, kind of, confusing, but it's the, they're still experiencing significant, significant improvements. They're on a prolonged journey. Whereas the AI emergence, they are kind of, you know, just starting on their journeys, and they are experiencing the greatest, you know, 12-month improvement in business outcomes because they are looking at much more simpler, but also they don't have that accumulation, accumulated benefits from their prior stages of the journey. So it can be a bit misleading, but the really, details are, needs to be focused upon as to what differentiates the AI emergence versus AI masters. So in closing, I'd like to kind of stress on the fact that AI is not a nice-to-have, it's not a choice, but it's not an option.
Every organization who is at an AI emergent, they really need to kind of accelerate their journey towards being an AI master. We have some very interesting insights and guidance based on the discussions that I had with Yancey right now, as well as a part of our detailed study that we'll be sharing with you all. So take a look at that, and we encourage you all to kind of engage with all of us to learn further. So thank you, everyone. It was wonderful to discuss our insights from our collective study, and with that, I'm gonna pass it over to Russell.
Wow! Those were some great and actionable insights. My sincere thanks to both Ritu and Jonsi for that peek into what we can learn, in particular, from best practices of more AI masters leading the charge in this era of AI. As Ritu mentioned, the report that we've been discussing is available today through netapp.com, and for those of you who registered for the webcast, you'll get an email with a link to the report, too. But before I move on, I wanted to take a minute to summarize what we have been talking about today and the fundamental and critical role that data plays in helping organizations scale AI to the whole enterprise. Let's start with the reality that data is everywhere. It doesn't confine itself to traditional definitions of on-premises or cloud. It's generated everywhere and is needed anywhere.
That's why we believe organizations that want to accelerate their success in AI need a flexible data and storage architecture that is, at its core, seamless, hybrid, and multi-cloud. Data can be held back by the proliferation of so-called specialized storage and data silos, which add hugely to the complexity of making AI real. We believe the future of AI will be delivered by a unified data environment that is optimized for AI throughout its life cycle. Organizations are leveraging their most precious data to drive this AI revolution. This data isn't just subject to legal and regulatory concerns, but it's also commercially very sensitive intellectual property, and yet, the very folks that are charged with innovating at such an incredible pace around AI, such as the data scientists and data engineers, aren't typically that focused on the responsible use of this data.
That's why we believe that organizations must demand an intelligent data infrastructure that delivers built-in governance, security, model and data traceability, and provides automatic defense against malicious attacks. AI requires these incredible resource needs, so although it may seem like table stakes, having a unified data environment that can seamlessly deliver the performance required for any AI workload is critical, all while ensuring it's being done efficiently from both cost and environmental perspectives. Finally, what use is a unified intelligent data infrastructure if it doesn't directly help the users that are actually tasked with working together to deliver AI? That's why we believe in a data environment that is tightly integrated with MLOps platforms that are used daily to innovate in AI, focused on delivering improved productivity and maximizing the use of these hard-to-find roles. Now, with that, let's switch gears and talk about our partnership with NVIDIA.
We've just come out of the NVIDIA GTC event in March, and together, NVIDIA and NetApp made a number of exciting joint announcements, which we're gonna recap now. First up is this great quote from Jensen Huang, the CEO of NVIDIA. This talks about one of the main reasons why we continue to expand our long-standing, more than six-year relationship around AI. Data is the fuel that is driving this current explosion of generative AI solutions, and NetApp's leadership in intelligent data infrastructure for AI makes us a perfect complement to NVIDIA and their incredible rate of innovation in the space. In particular, our unstructured file data capabilities are foundational to solutions like Retrieval Augmented Generation, or RAG, which was front and center at GTC. And to talk more about this, joining me from NVIDIA is Tony Paikeday, senior director of marketing. Tony?
Hey, thanks, Russell. So, you know, I've had the privilege of partnering with NetApp for, I'd say, the last eight years, beginning with the creation of our first integrated infrastructure solutions for enterprises. And, you know, since that time, our two companies have embarked on the journey of taking something that used to be exclusively the realm of maybe science and academia, I mean, no offense to either of those, and democratizing AI, you know, for every enterprise developer and data scientist who needed a new, better platform on which to innovate. So over the next few minutes, I'm excited that you and I get to share some of the most recent advancements in our joint portfolio as it unveiled at the GTC conference.
Thanks for that, Tony. Yeah, so let's get on to that, that recap of those GTC announcements. First up, I wanted to cover the new NetApp AIPod. So leveraging the NVIDIA DGX BasePOD architecture and the DGX SuperPOD product, NetApp AIPod is a set of reference architectures designed to help our customers accelerate their adoption of a broad range of enterprise-class AI training environments. AIPod builds on and is a culmination of over six years of joint innovation between NetApp and NVIDIA. Combining NetApp's AFF A-Series and C-Series storage solutions with NVIDIA's DGX delivers the performance customers need, non-destructive scalability, and a unified approach to data management throughout the AI life cycle. Customers enjoy a simplified deployment and operational experience with built-in data security and threat protection.
With the previously announced NVIDIA DGX SuperPOD with NetApp E-Series, customers get the ultimate in speed and density HPC and ultra-performance workloads from the same partner they trust with their AI pipeline. So Tony, if you had to pick, what are the three most important things you're hearing from your customers when it comes to architectures purpose-built for AI?
You know, Russell, this is why I'm so excited about our partnership. You know, our two companies are single-minded in our focus to help businesses overcome the challenges of AI platforms. Customers repeatedly tell us, "First, help me eliminate the complexity of designing infrastructure and navigating that delicate balance of compute, storage, networking, software, all working together as a full stack solution." I guess the second thing is, putting all the componentry together can be hard. Deployment time frames get elongated, delaying the point at which, you know, one's developers can actually be productive.
So, you know, they're saying, "Give me a faster way to deploy, backed by partners who have the competency to do it all." I guess the third thing is, driving performance and higher utilization of infrastructure can seem like a black box to those who don't have the deep bench of HPC expertise. So they say, you know, "Give me solutions that offer linearly predictable performance, that scales, that's IT manageable, that's backed by enterprise-grade support and AI experts who can help solve problems." So we're excited about AIPod because it addresses all three of these concerns.
Yeah. Appreciate those insights, Tony. Those that were able to catch the keynote in NVIDIA GTC may have seen this picture on the main screen behind Jensen. We are super excited about highlighting the work NetApp has done to demonstrate how to integrate our ONTAP storage OS into the recently announced NVIDIA NeMo Retriever. Building on our strength in the enterprise data center, we've enabled our customers to simply and rapidly integrate their existing unstructured file data into a generative AI knowledge management solution, powered by the NVIDIA AI Enterprise software suite. We talk about helping our customers talk to their data. With this solution, we're accelerating our joint customers' journey to extract latent value from that data. So Tony, how do you see retrieval augmented generation impacting the GenAI landscape and our customers?
You know, Russell, for years we've partnered on helping leading edge companies train mission-critical models, from scratch, and we've done really well at it. But we're entering into this new era of AI because of a couple of things. First, rather than serving up answers and content that already exists, the future of AI is generative, creating answers and content that really never existed before, using natural language as the API for this capability. I guess the second thing is, with the advent of RAG, as you pointed out, we're now enabling businesses to tap into oceans of unstructured data they already have and literally speak with their data, talking with their vast PDF files, technical documents, customer transcripts, financial reports, you name it. So who better to help them unlock this capability than the partner that they already trust with their data, namely NetApp?
So unlike the paradigm of the last decade, AI is no longer confined to those who are committed to creating their own foundation model from scratch every so often. Now, literally anybody can get incredibly accurate up to the mid- and mid answers, leveraging data that they're already sitting on, unlocked by this joint solution, leveraging NVIDIA NeMo Retriever and NetApp.
Yeah, I agree, Tony. We see this as ... I think we both see this as truly revolutionary rather than just evolutionary. So now let's turn our attention to how NVIDIA and NetApp are working together to help organizations operationalize their AI workloads at scale. NVIDIA recently announced their updated OVX platform. Optimized for the enterprise, it targets key functions such as inferencing, fine-tuning, and light training at scale, and is ideal for workloads, including generative AI and techniques such as RAG, as we just discussed. NetApp's intelligent data infrastructure capabilities and enterprise pedigree, combined with holistic support for the entirety of the AI data lifecycle, means that we are perfectly suited to NVIDIA OVX, and we recently announced NetApp ONTAP is certified as a storage partner for the OVX platform. So Tony, we are super excited about partnering with NVIDIA on OVX.
Can you share how it complements the rest of an organization's end-to-end AI solution?
Yeah, Russell, you know, we really see two modes of AI development emerging in the enterprise. I mean, there's the training and customization of foundation models that require massive computational power to deliver a production-ready model, maybe in hours or days, instead of what used to be, like, weeks or months. But additionally, we see the rise of RAGs, as we discussed, leveraging pre-trained models, working in tandem with an embedding model and a vector database to augment those models with live enterprise data for more timely and accurate answers. The second mode requires a platform that can live where the data is created, maybe at the department or the business unit level or even the edge of the enterprise, alleviating the pressure of having to continually retrain a large or very large model.
The OVX infrastructure fills this need, with a platform that can be deployed cost-effectively where the data lives, optimized for fine-tuning and RAG deployment, with NetApp intelligent data infrastructure, enabling effortless mobility of datasets and models to wherever they're leveraged.
Thanks, Tony. Yeah, we certainly see this huge drive towards operationalization. It's the next big frontier. It's really where we're gonna start seeing folks generate real value out of their AI investments. So we're now gonna move on to the Q&A section. So participants, as mentioned earlier in the session, please go ahead, ask your questions using the Q&A feature in Zoom. I'll go ahead and pick them up to ask the presenters. And for the questions, I'm pleased to say that we are also being joined by Will Vick, who's the Global Director of Technical Sales and Strategy, NVIDIA. So let's see what questions we have for the panel. Okay, so let's start with this one. Can you explain the definition of unified storage used in the study and its relationship to accelerate time to business value?
Let's, Jonsi, would you be okay taking that one?
Yeah. Can you repeat it? Just broke up a little bit.
Yeah, sorry. Can you explain the definition of unified storage used in the study and its relationship to accelerate time to business value?
Yeah, I mean, when you are able to unify your storage underneath all of your workloads, I mean, that is a huge benefit to AI. You know, no longer having these silos. I mean, as you see in the report, the biggest issue is access to data, and avoiding silos, and having it 100% secure, and mapping and knowing your data. That to me is a unified storage. And not only the unified storage, it's also about being able to move the data when it's needed. When you need additional GPUs, you can go into DGX Cloud. When you need to or in any of the public clouds, it's all running on NVIDIA.
From my perspective, you know, having a unified storage strategy for your workloads, and particularly for your AI workloads, is key.
Thanks for that, Jonsi. Ritu, anything you want to add here?
Yeah, thank you, Russell. So, you know, as I shared in the study that we did, the number one reason for AI project failure was infrastructure limiting data access, right? So infrastructure that breaks down data silos and allows faster access of multiple data types. We are no longer just working with structured data types stored in very structured formats or databases, but a huge amount, especially in the era of GenAI, we're dealing with unstructured data and, you know, and semi-structured data. So any infrastructure that helps to kind of provide flexibility and access and faster access to multiple data types and accelerates the workflow, will be really a game changer for the businesses. So I just want to kind of stress on that.
Yeah. Thanks for that. Thanks for that, Ritu. I've got an interesting question here that maybe Will can help us with from NVIDIA. Can something like DGX Cloud or Equinix solve the issue of power consumption in legacy data centers? And I'm assuming the person who asked this question is referring to some of the recent announcements of the Blackwell training environments, which have some quite juicy power requirements. Will?
Yeah, absolutely. And, you know, having this kind of ready, let's say, ready data center infrastructure is kind of how we work with Equinix, as an example. To be able to leverage that when you don't have the power, you know, capabilities within your own data center, or let's say you're building out, let's say, a net new environment, and you need a data center immediately. You know, working with them makes it very easy and attachable from an API point of view of how you're delivering AI, but also how you're connecting to the cloud, as well as other resources to move data between, right? So it's all about efficiency and reducing the total cost of ownership between the resources you have today.
Yeah. Thanks, Will. Jonsi, anything to add on this?
I mean, just to emphasize what Will said. I mean, a lot of companies aren't ready to change, you know, redesign their data centers to InfiniBand. I mean, they're probably gonna go with Ethernet-based solutions, that there's a vibe that I, you know, that we are partnering with NVIDIA on as well. But, you know, the key is to be able to burst into Equinix, DGX Cloud, and the public clouds when it's needed. And, not only that, but you have to be able to support all these different protocols, you know, file, object, block. It all has to be sort of managed and being able to distribute or burst into these service providers that have the power design and have the GPUs available for your needs.
Thanks, Jonsi. Ritu, from your perspective, what are you seeing on the IDC side of things?
Yeah, I think I'd like to add. I think Will and Jonsi captured it pretty well and succinctly. The only thing that I'd like to add is that we are seeing, you know, colocation as a very viable option for many of the training instances because of data security, data privacy, and, you know, data gravity. A lot of customers do have, you know, situations where they have their, you know, training of the traditional AI. We have kind of forgotten about traditional AI in the era of GenAI, but they do do a lot of the in-house training and inferencing from there.
For all of those reasons, not duplicating the copies of the data, getting it faster access from the colocation facility within their own private firewalls, is a very viable option and a very, you know, solvable problem for the power consumption and resource efficiency as well.
... Awesome. Thanks. Thanks, Ritu. I've got another good question here that I'm gonna, sort of ask, Jonsi, if he can answer: What can NetApp contribute to responsible governance of data for AI?
Well, I mean, we put a lot of emphasis on our governance and compliancy within NetApp, and that, of course, is key going forward. Because you have a lot of requirements that you need to, you know, prep the data, obfuscate the data before or prior to moving it, or uploading it, and using it to train your models or even fine-tuning it. So I don't think a lot of our competitors, frankly, are able to offer the same guarantees as we do when it comes to compliancy, governance, and security.
Thanks, Jonsi. From your perspective, Ritu?
Yeah. So there are a couple of aspects. I think Jonsi covered it extremely well from your perspective, but, you know, from a market industry perspective, there are a couple of things that we are seeing that from a resource efficiency perspective, we actually want to kind of, you know, bring the data to the- AI to the data, right? In the many situations when it comes to training large, you know, buckets of data, even in the case of edge AI inferencing, when you actually want to bring AI to the data. So there are some very interesting angles where massive chunks of data for data privacy, data has gravity, you don't want to move the data, and it also compounds and adds to the resource efficiency aspect.
Thanks, Ritu. Okay, here's another good one. I think this one would be for you as well, Ritu. So regarding resource efficiency, how should I think about moving data for AI training? When should I do it, and when is it better to bring the AI to the data?
Yeah, yeah. I think I kind of, you know, jumped to this question because I was leading to some of the questions in the, you know, Q&A. So I think it is fundamental to remember certain kind of guidelines as to when. First of all, in the era of GenAI and with all these pre-trained models, you don't have to use massive, massive chunks of data just for the heck of it. Use the high-quality data set, make sure that you have the right data security and data privacy around it, and make sure you have the right whole life cycle of the data governance aspect of it. But then, when you're training, you know, and the high-quality data sets, because I kind of already alluded to it, the data has gravity.
You don't have to move the whole chunk of data because it will have overheads in terms of, you know, security, privacy, data movement has costs associated with it. And with all the data security requirements, if you kind of bring AI to the data, your own training, you can actually help, you know, manage it much more efficiently. At the same time, you know, edge AI, we all know the most important advantage for that is to actually kind of bring AI to the data for the inferencing, for privacy reasons, resource, you know, less power consumption, less bandwidth usage in actually moving the data. So there are a whole bunch of, you know, combinations, and I also kind of alluded to when we were talking about the co-location aspects.
There are certain scenarios when you already have your existing data sets running in the co-location facilities. It makes a lot of sense to keep it there. Russell, I think we also had a question in the live session about Edge, so feel free to ask to all of us. I'll be happy to provide some perspective on that.
Oh, I need to find... Which, which one is this one, Ritu, if you can see it?
I think the question was about, let me just, I had noticed that we can answer that. You know, it was on the,
Yeah, it was, it was basically what the panel is seeing with the Edge AI in terms of requirements-
Exactly.
Deployment models, adoption, growth, data management.
Yes. Yes, yes.
Yeah, go ahead.
You're right, Will. You want to take it first, Will? Go ahead, and then I can answer some perspective from IDC perspective as well.
Yeah, for sure. From an Edge AI perspective, we see this growing tremendously, you know, these years. I would say the first... I've been with NVIDIA for 5.5 years. I would say the first couple of years were more of like, how do you go build training, and how do you start supplementing these things in reality? And so now we actually see the huge trend of upscaling a lot of Edge AI capabilities, let's say, whether it's in a point of presence for different web apps that are being serving for different audiences, for different languages across the globe. We see a lot of, you know, customers migrating towards that.
And as well, from an adoption standpoint, you know, customers are leveraging a lot of the GenAI capabilities to start building this out, from what we see, right? You're getting a lot of new funding. You know, how are you addressing this market for your own good and, let's say, diversification plan? But then, therefore, how do you go take this, and put it into market to service your customers? So we see this quite a bit being one of the new struggles, but as well, new capabilities that they're giving customers to go drive this. And again, it's all about, you know, a lot of the time, environmental locality, internal constraints that could be historical bias, regulation, data sovereignty, things like that as well. So, all pointing to that as well.
... Thank you, Will. So let me just add some perspective, what we are seeing from IDC angle as well. So we all know that there are some critical benefits of edge AI, right? But I do want to call out that there's some nuanced difference between traditional AI and GenAI. In the GenAI spectrum right now, we are in the very early stages. We haven't seen so much of edge AI so far. Having said that, there are, it's just being done, and it's in the pockets, right? And, but having, and majority of the inferencing that is happening right now is in the cloud, but we expect that to happen in the edge as well for GenAI.
But now coming back to more traditional AI, we all know that the most important advantage of Edge AI is that it brings high-performance computing capabilities to the edge, where sensors and IoT devices are located. And it—we are seeing that, you know, if the data processing in the cloud takes seconds, data processing at the edge, where it's mission-critical workloads, like autonomous vehicles, you know, the edge needs to make decisions much faster if data is being processed at the edge. And these decisions, they impact human lives in many cases, and near real-time processing is critical. There are aspects of privacy that we see, which plays a big, big critical role in healthcare and all. We see that customers do a lot of edge training as well as edge inferencing.
There was a very interesting use case that we had seen where the training was done at the edge devices for submarines, right? I'm not, in the interest of time, I'm not going to go into the details, but in all of these cases, data availability, data accessibility was very critical, and it also resulted in reduction in internet bandwidth and cost and less power. And there are many, many interesting edge use cases. But I would like to kind of close with on this particular topic, is that I alluded to the word federated learning. So many use cases we see where data privacy is front and center, the things are being done, the edge AI models, they operate on the edge device, and only the results are transferred to the main central model to kind of take it in a cyclical loop.
I'm just talking about a very simplistic federated learning, the different options of it. So those are the different models that we see. And we'll see this completely transform in the GenAI era, in the near future, but not so much yet right now. Okay? Over to you, Russell.
Awesome. Thanks, Ritu. There's a question here I'm gonna answer myself. Is there a list of MLOps platforms that NetApp integrates with? And the answer to that is we integrate with a whole bunch of MLOps platforms. We certainly believe that the value that's intrinsically available through NetApp's ONTAP Everywhere approach is can be exposed up to data scientists and data engineers in a way that significantly improves their day-to-day experience from a productivity perspective, really simplifies dataset management for these folks. Some examples, Domino Data Lab, Run:ai, Iguazio. In the cloud, on AWS, SageMaker, and Bedrock. In GCP, Vertex. There's also some integrations we've done on Azure, too.
So, you know, across the three main hyperscalers and across a bunch of enterprise control plane solutions, NetApp has got a very broad ecosystem of solutions that we support in this space. Obviously, Run:ai, particularly interesting because of NVIDIA just announced the acquisition of Run:ai as well. Okay. Let's just see if there's one more here. One second. Oh, yeah. Okay, so there's one last question, I think, folks, before we wrap this up: In regards to competitive landscape, where do we see NetApp versus their competitors with these AI workflows? Do we have an advantage? Are there things competitors have that NetApp does not? Hmm, that's pretty direct. I think, Jonsi, this is a good one for you.
Yeah, I mean, what uniquely sets us apart from our competitors is we are the only ones that can say that we are everywhere. You know, we don't have to dictate where the customer is deploying their workloads. So... The other thing is, you know, especially when, like, when we are talking about the edge, and edge is coming, because a lot of the, like Ritu said, a lot of the inferencing is going to be happening on the edge, at the edge. There, the data management capabilities, how do you link up, you know, core to cloud from edge to core to cloud based on the needs that you have?
And being able to redirect the data where it needs to be at any given moment is a very unique NetApp thing. When it comes to our sort of competitors, I mean, a lot of when you go with InfiniBand and you have, like, the parallel file systems and everything that is out there, of course, in some cases, that's where our competitors might shine, but they don't have the data management capabilities. And in all honesty, they create more silos than and create more problems than they actually solve. But we have, of course, E-Series with BeeGFS as a solution there as well. But then again, you lose the data management capabilities. That is key for you know, going faster.
Thanks, Jonsi, and I'm super proud of the things that we've done to say make dataset management in an intrinsically hybrid world of AI data life cycle super easy. So okay, with that, I think, that's all the time we have for questions. To wrap up, I wanted to recap how an intelligent data infrastructure delivered by NetApp can assist in overcoming many of the challenges and pain points on the road to more successfully ramping AI, and in particular, GenAI initiatives. First, a flexible data and storage architecture that is, at its core, seamless, hybrid, and multi-cloud, delivering unified data access across a variety of data types. Second, a data management architecture that delivers built-in governance and security, model and data traceability, and automatic defense against malicious attacks.
Third, all the performance needed for even the most demanding AI workloads, while optimizing for cost and sustainability. Fourth, a data environment that is tightly integrated with MLOps platforms that are used daily to innovate in AI, focused on improving the day-to-day lives and productivity of the key people responsible for delivering AI in your organizations, whether they be data scientists, data engineers, developers, or even IT. And finally, a laser focus on the targeted business outcomes rather than the technologies that drive them. We have a fundamental belief, as I mentioned earlier, that data is the fuel that is driving AI innovation, and unlocking your data's potential is your key to taking advantage of the opportunity ahead of us. With that, I'd like to thank both our guests and our attendees for your time today.
Remember, the maturity study that we covered during this webcast will be sent to those who registered for the session, as well as being available on NetApp.com. Additionally, on our website, you'll also find some of our executives' perspectives on AI, as well as more detail on our joint solutions with NVIDIA. Thank you, and have a great rest of your day!