Stop Scaling GPUs. Fix Your Data First | Prakash Darji, Everpure

Today, every enterprise wants to move faster on AI. Leaders buy bigger models and more GPUs, but that doesn't solve the real problem. It's not compute. It's data, and most enterprise data is fragmented and not ready for AI. And that's the gap Everpure is closing by rethinking storage as active intelligence instead of passive infrastructure. And today we have with us once again Prakash Darji, GM of digital experience business unit at Everpure, to break down why data, not compute, is holding AI back. Prakash, it's great to have you on the show. >> Thanks for having me back. >> First of all, let's talk about Everpure. What is Everpure all about? >> Yeah, look, largely we we started our journey um delivering the best storage and made it a unified data plane that allowed supporting all workloads and applications from archive to AI, right? We've generally come from that lens. But as we take a look at where the industry is going and the emerging problems, that largely started in an application-centric world. People like the unit of architectures for the last 40-50 years was yeah, software companies build applications, I uh vendors buy them, and IT departments implement them. Right? So we see a world and now where if you see agents are working on data and applications become workflows, and you have this emerging challenge, we think it needs to be rethought with this idea of data primacy. And this idea of data primacy is such that at with all of your data, how do you go ahead and understand what is needed and what is relevant for your applications and for your agents for AI. So the rethink is instead of just saying, "Hey, I'm going to take everything I have and give it to a model, is that going to yield a good outcome?" I think 6-7 months ago, people were like, "Build factories. Give me all your data." And you largely ended up with expensive bills and hallucina- hallucination. Now, we're saying, "You know what? If you know what is relevant to solving the problem, small language models do just fine. Agents work on it with relevant data, producing better results at much much better cost paradigms, but it all starts with the input to AI, which is data. >> Did you just say the term data primacy? >> Yeah, the word data primacy is like, "Okay, if historically, you had applications as the center center of the universe. This is my CRM. This is my finance system. This is my supply chain system." You had applications, and they all shared data, right? Like, finance data, sales data needs to make it into finance to book. Right? It's a common thing. Like, you have orders and invoices. They need to work together. Um and that integration problem became things that people needed to work with. Well, if you think about a data primacy lens, what if you had one canonical system of record with like your data sets, your sales, finance, billing, orders, all of it in one model where applications were workflows that just contribute to that model. So, data primacy is actually inverting the pyramid, but it only works if you change the paradigm to first say, "Build a consistent view of your enterprise in data format, where applications become workflows contributing to that, versus you integrating multiple applications." That is the idea of data primacy. >> When it comes to data, we have been saying for a long time that data is the new I think data is the new solar. It's evergreen. Uh but when it comes to AI, folks often get caught up with compute, GPUs, tokens. Talk a bit about why is data the real bottleneck, not compute, and how are organizations getting it wrong? >> Let me use an example, right? I think we all know that, you know, people have procurement processes. They buy things. And what You have people that approve invoices and approve purchase orders. So, what if you wanted to create a procurement agent that reviewed and approved purchase orders for a company. Right? This is a simple scenario that probably most enterprises have to deal with. What information or what data do you need to solve that problem? Right? Where does it exist in your enterprise? And for example, an enterprise probably has previous purchase orders sitting in systems, finance systems. They probably also sit on spreadsheets of users on desktops. They, you know, there are people who do credit checks. There's AP teams, accounts payable, there's accounts receivable. You have all these different teams that participate in the process. But if you say, "Okay, to really do this procurement agent, I don't know. My data's somewhere. I have a lot of invoice-oriented data." I could go ahead and say, "You know what? Give access to this agent to go train and build a model like build and, you know, a vector vectors and apply weights and graphs to allow the agent to make uh to inference on my data, but I'll give it all my enterprise data. All the way from sales orders to finance to supply chain to like everything. One, it'll be expensive. And two, your outcome won't be as good as if I said I only needed to give that agent relevant data. Now, what is relevance? Well, it's probably my primary records for invoices. Right? And purchase orders. It's past due payments. Right? Accounts payable records. It's probably credit checks. Um it may be some customer relationship. But that's about it. Like So, it's the same thing that a human that was improving PO, approving POs on a day in day out basis would look at. They'd go run this thing. They'd go check that. They'd be like, "Oh, is this vendor good?" They'd do all these things to approve the PO. You want the agent to do the same thing with the access to the same information. Right? The human doesn't look at everything. They don't look at irrelevant data. They don't brute force compute it. They look at relevant data. And because of it, they make better decisions. And the same thing is true for agents. So, what if you could go ahead and scan and find and categorize all of your data to say what is primary source and what is relevant? With Everflow data intelligence, that is what we do. And that is the capability we're providing to contextualize what is relevant information for input into AI. Right? And then, you can apply whatever model you can build the agent with There's tons of companies building models, and models keep changing. There's tons of companies trying to build agents. Build it on my platform. Build it on that platform. Right? And most companies are saying, "Just copy all your data right? From all of your sources to each agent." And I'm like, "Well, that's the dumbest thing I've ever heard." Right? You can't go ahead and say move all of your source data from everything into every use case that an agent is doing. You're just going to have data explosion. So, you need to work with the data sources where they're at, categorize them, and allow the agents to work with them. And that is the approach that, you know, we are talking about as we provide the input data AI. Your compute becomes smaller, your models become smaller, everything becomes smaller because you're optimizing for SLMs and purpose-built agents. >> What you're saying is that there's a shift from application-centric to data-centric. What does that actually look like for enterprises day-to-day operations, developer teams, procurement when they are operationalizing AI? >> Well, so look, I would I'm going to talk about two different types of companies, right? First, let's say I'm a new startup today. That's the easiest because it's like, oh, well, you know what? I'm going to ensure that my engineering team, my data engineering team, manages my enterprise data as one canonical schema and system of record. And I'm going to use Claude or, you know, some type of workflow generator to create workflows that populate the schema for orders, for deliveries, for billing, for supply chain, for support. You're treating everything as part of one canonical model. Right? Um and I think that approach only really works for a new company with no legacy. Right? [laughter] Like, if you've been buying apps for 40 years and you've I worked at SAP working on financial applications way back when and SAP Hana, etc. You've got a ton of applications or workflows sitting owning your enterprise data. You can't get to data primacy overnight. Right? What you can do though is you can say, "Okay, I'll probably start with it for my agents. Like, I'm going to build this as a separate schema for my agents, and I might have to replicate my data as a starting point. But, as agents take over some workflows, could I generate for things that are simple, simple workflows? I wouldn't say complex MRP runs in ERP, but I would say, you know, creating support cases. Right? A simple use case, or creating sales opportunities, where it's, you know, pretty much a web form on a table. Like, those things are things you could say, "You know what? Generate me a workflow, and I'm going to decommission my old workflow, and decommission the data replication, and start chipping away at my enterprise landscape to move towards data primacy. Right? Which is this idea of this unified canonical schema, cuz everything I decommission saves me an integration work. Right? All that like, a bulk of an enterprise's implementation and application is integrating the APIs and the connections between the apps and the data. Right? So, by moving to data primacy, your major value, besides unlocking AI efficiently, is reducing your burden of integrating applications. >> And for a long time, we have seen that storage was passive infrastructure. It was never the sexy, shiny object. Now, AI expected to behave differently, provide context, be available where AI runs, egress, ingress, that also matters. Data warehouse, data lakes, there is a big debate about that. So, talk a bit about how is AI changing how enterprises look at data and data platforms. >> Well, look, if I start with this year's kind of storage question, right? Storage and data were historically separated. Like, where storage platforms were infrastructure that was meant to be reliable and durable and long-lasting and you know, it was like mission-critical. It's like an x-ray It's like blood monitor machine in a hospital. Like if it goes out people die, right? Storage is if it doesn't exist you're toast, right? Like it is a critical asset and the design principle is always about reliability, stability, etc. And certain applications needed that like transactional applications and analytical applications were like, okay, if my reporting is down for a few days, it's not a big deal, but like, you know, if I can't take orders, it's a big deal. Right? So, I think the design paradigm for infrastructure was always built with a high resiliency approach in mind. And now if you extend that idea to agents doing work where you're now saying it's not transactions and analytics, but it's bringing it together where agents need both. It needs to make a decision to book a transaction or approve an order, right? That same procurement agent I talked about will need to do both the analysis saying, is this worth approving? And then it'll actually book the transaction to approve the PO. Right? So, it's both doing transactions and analytics together, right? Like most agents that take action will have to do that thing. So, it creates this high resiliency paradigm across data where storage and data have to come closer together. Right? And if you hear different companies like you know um you have analytical companies like Databricks and Snowflake and you have transactional companies like SAP and Oracle um and Salesforce all trying to navigate this change. But their first answer is always move all of your non-data that's not in my platform over to my platform and do it with me. And they can't all be true. You know what I mean? You can't say put all my CRM data in SAP and all my SAP data in CRM just for a sales agent and a finance agent. Like it doesn't work. So I think you have to look at this heterogeneously, cross application, and cross data platform provider. And that's where storage has always been, right? Storage has been about all data. Structured, unstructured, et cetera. And that's the That's the reason why we believe we have a unique entry point coming into the space. >> You introduced Data Stream and Data Intelligence at Accelerate. How do these two fit into the data-centric vision that you just shared with us? >> Yeah, look let's just talk about like the end-to-end of AI. At the very beginning, it's what data inputs do you have? That's where Data Intelligence comes in. It finds, discovers, classifies, and creates a knowledge graph of data relationships. It can say, "This is all your finance data related to creating needing to go into a procurement agent model." Right? That's the find and discover side of the problem. Then, if you want to go and create a pipeline and et cetera, Data Stream is there to vectorize the data so you can prep it in a vector-ready format for AI inferencing and use. Right? Data Intelligence then can apply attribute-based access controls because previously you had role-based controls for information um uh on the data sets. And then you got customers can use that data, right? So how we see this working together as part of solving end-to-end AI problems for end customers. >> With the generative AI rising, what role does governance, what role does security play in regulated industries? How is Everplate addressing this challenge? I think one of the most important things that is going on right now in the world of AI is how do you govern and secure? Because when vector databases came out, your straining needed access to all information and they used vector databases don't have role-based access controls. That's not how they work, all right? [laughter] So, largely you have to think about security differently. I'm no longer saying you can see these things in your SAP ERP. If the agents are working at the data, you need data access security and attribute-based access controls to say, "Hey, these are the attributes is these are all the finance or procurement-related data, regardless of the workflow or the app, that this user needs access to or this agent needs access to, right?" So, uh I believe that becomes very important. And then if you lay on like GDPR and DSAR and right to forgotten and like governance, cuz security and governance and compliance are two different things, right? I was talking about data access security, but when you move into governance and compliance, what if there's PII information in your model, right? Like so, this is personally identifiable and you could be fined if it, you know, propagates to the wrong place. How do you know identify it? So, with data intelligence, we actually can tag and find and classify sensitive information and ensure that it's only uh within the bounds of your governed policies and security. And we can actually prove it with audit reports. So, if you get, you know, uh SEC auditor or a EU auditor looking at your environment, you can actually produce an audit report for compliance to, you know, GDPR or DSAR or those types of things. So, that is the approach that we're taking in terms of it, but it's really important and I think it's an often overlooked problem in the world of AI. >> Now, we are going through this whole geopolitical crisis, which is also accelerating data sovereignty concerns, model sovereignty blocks, organizations are cautious about data access. How are you looking at this changing landscape? And of course, we have things like new clouds as well. So, will data privacy handle these challenges or it is out of the scope of data privacy? >> Yeah, look, I think the one yes it will, but like I'll I'll kind of offer a philosophical comment first, right? One, I do think it's unfortunate. Like, you know, I think governance and policy should be protecting consumers like less about, you know, nationalism and governments, right? But like we are inherently heading into a more nationalistic and sovereign era where people are trying to treat models that way. Like you saw that with the US blocking Fable 5 and then allowing it. Um you know, a Chinese built uh building GLM and kind of making that open source. There's like this view on models and then there's this view on data and controls, right? What are the things? And I actually think the more important thing less like the models are not the moat. I think eventually the models become commoditized and shared and open source. But I do think the data moats are things that people are like the inputs to the model are the things that governments are trying to protect now. Right? And you'll see that with like a lot of the original European regulation saying this is the data that can leave. Like right now I think they're mixing the data's in the models and kind of confused policies, but eventually as you straighten that out, um you know, what is the data that can't that is [snorts] unique to whatever your problem is. And it might be PII now, but it might be trade secrets or it might be how to enrich uranium. You know, like there's all sorts of things that become sensitive, right? From an information class standpoint. So, I do think data inherently in a world of AI, 100% of all data should have an information class attached to it. Right? It has to. And you have to operate within that information class. Right? It's no longer, oh, let me just find I'm a credit card company in a highly regulated in company, let me find where the credit card number is. I don't think that's enough in a world of AI now. Right? So, um that Now, if it's more nationalist, that's just a policy. If it's less nationalist, another policy, but it all starts with ensuring that you understand the in classification of that data. And with our data intelligence, that's what we do. We discover data wherever it sits, on premise, on pure arrays, in cloud, on mainframes, in SaaS systems, we can discover it and classify it. So, you can assign it to an information security policy for governance and compliance reasons. >> Now, let's talk about culture versus technology. People are the hardest part. Technology is the easy part. Now, with agent TKI that is making decisions, bad actors are finding new ways to manipulate it. There may be images which human will see just an image, but it may have code that AI can read and it can execute some action. So, talk about how much role does culture play, processes play, people play in rethinking data. What culture should companies build so data privacy becomes part of their DNA, not something that was bolted on top of it? >> Yeah, look, it's interesting and I'll give you this more from our own culture right here at Pure. Um the pace of change in what's happening in AI is so fast. Right? Like where six, seven months ago, we would have been like, "Okay, what's going to happen?" And now we're like, "Hey, this thing writes really good code." And you know, it's changed the productivity of our engineers in terms of what they do. But we wouldn't have gotten this knowledge if we didn't allow people to experiment and learn. Right? So, the first thing is experiment and learn. The second thing from a culture standpoint is you need to understand that some people will blow your budget by doing really stupid things because they're uneducated. And that might be part of the learning process. But you do need with the cost of models and tokens to understand where and how observability becomes everything, what you're using, and what models you're choosing. Otherwise, you'll like blow your token costs like well beyond the ROI of doing anything. So, I think culturally right now, you have to first understand what the model's capable of, what data and context is relevant, allow people smart treat people as smart and empowered, but know that as you learn, as we all learn as humans, we make mistakes. And mistakes are okay as long as you learn from them and don't repeat them. Right? And the guardrails you need to set isn't like limiting, you know, a quota per engineer, but it's to prevent something where you like, you know, a company can go bankrupt overnight with a recursive job that's burning tokens while they're sleeping. Right? You don't want that thing to happen. So, how do you build guardrails to protect against the mistakes, not prevent the people from learning? Right? And eventually this will stabilize. Like I don't think we're going to be talking about token costs when people can just download a GLM model or run it offline on a thing and you know, apply it to their enterprise data, like you're going to think about this very differently. Um when you're downloading the weights and applying it kind of and running it on your own thing because then you're just using the asset you have and not paying per token, right? So like the models of where we're at is just a kind of current place in time. But the thing that lasts is one, build a culture where you trust people, allow them to experiment, build safeguards where even if they make a mistake, the mistake isn't crippling to the company. And the three, as things stabilize, then drive efficiency. I think the companies that like are like, "Hey, I'm going to prevent because I'm afraid of cost or I'm going to prevent." is just preventing the learning. Right? So have a learning curiosity culture, I think is important. And having a safety culture where you get mistakes are safe is also important. >> Now, let's talk about how should organizations look at a whole token context window, of course, cheaper models are there, open weight models are there, smaller models are there. Sometime we get caught up the weight models. How should organizations approach infrastructure and architecture to leverage the these new models, these new technologies to get full advantage of AI without getting burned and avoid an AI bubble burst? >> Man, that's a so such a large question. Okay, so generally, I'm I look at historical patterns and I think you mentioned a lot of good ones. Like early days of internet, early days of social media, early days of telcos, early days of SMS text, right? And what is unique not about each scenario, but the pattern we can learn from history and any new innovation in our disruption cycle is at the beginning of the cycle you end up with sensationalist headlines saying we're going to put a rocket on the moon and we're going to like, you know, completely change the world and whatever whatever. So, the early innovators to raise money have to promise the stupid. Right? Like, what they're saying has to be so sensationally far away to be able to raise the amount of capital you need to raise to move 10 20 yards forward. Right? They're going to promise the 100 yards forward and we'll never get that We've never ever gotten to the initial promise of any transformation that has ever been promised in history. Right? However, sometimes we get to the 20-yard line, sometimes we get to the 50-yard line, and sometimes we get to the 60-yard line of that promise. Right? And we start seeing a slow down in a stabilization at some point. Sometimes it's starting to stabilize towards 20 and sometimes it's starting to stabilize towards 60. Like, and the internet was really interesting, right? Like we think of routers and switches and networking standards are all kind of there, but like Cisco was built on consolidating standards of 15 to 20 to hundreds of different you know, network routing things that create the backbone of the internet today, right? And the same way, like you talked about cell phones or whatever, right? Like, if we look at every technology, there was a point where it started stabilizing and optimizing. Right? And then you're getting incremental bandwidth gains and whatever, but the business model stabilized. Right? So, I would state that right now we're in the in unstable growth period towards the future envision of what they're promising, which is AGI. Right? Will we ever get to AGI? Generally, probably not this round. Right? Maybe that's the next transformation. This round will We may get 30, 40, some like some percentage of the way there. And when it starts stabilizing, what what happens? It kind of turns into a utility you can run, operate, and manage under constraints. And then I boil it down to do you want to own those constraints and run and operate and manage it, or do you want to outsource it? So, the infrastructure thing is do you want to own your data, or do you want someone else to own you? That's the question. Right? And I'll I'll use the example in the early days of SaaS, Salesforce, NetSuite, like a bunch of companies, Workday, LinkedIn, you know, were built SaaS companies were built before I as AWS, Azure, and Google like came out. And they built their own infrastructure, and they owned their own data, and they got to profitable. Like 70, 80, 90% profitable. Then when you had like the Snowflakes or other kind of SaaS companies come out, and they all crossed billions of dollars, right? Then when you had the next gen SaaS companies come out after infrastructure as a service, north of 50% of their cogs goes to the infrastructure provider, AWS, Azure, or Google, and it scales linearly, so they never can improve profitability because their resource consumption, because they've outsourced their infrastructure. They never get to profitable business models. So, if you believe that, you know, you don't need to own your assets at some point, like how do you build compounded profitability on a stable market if you're going to allow someone to scale your costs linearly? That existed with SaaS companies on I as and that's going to exist with OpenAI and models to be, and that's your decision That's your decision point on, you know, whether you want to use your own roll your own infrastructure or not is a business decision. Right? But we believe that you know to build profitable businesses you likely will need that. Right? You need infrastructure that doesn't scale your costs linearly like the infrastructure as a service world does. >> Prakash, thank you so much for walking us through VMware Enterprise the I get stuck and how it pure is fixing it. Once again, thank you for joining us and I look forward to connect with you again. Thank you. >> Yeah, thanks for having me.