57:26 Webinar
New IDC Research: Building Data Infrastructure for AI
Join us to see the research, hear what's changing, and rethink your data strategy to get ready for AI.
This webinar first aired on March 5, 2026
The first 5 minute(s) of our recorded Webinars are open; however, if you are enjoying them, we’ll ask for a little information to finish watching.
Click to View Transcript
Hello, everyone, and welcome to our webinar, Build Data Infrastructure for AI, featuring new research from IDC. I'm Erin Stevens, senior product marketing manager for AI here at Everpure, and I am so thrilled today to be joined by Ashish Nadkarni, and Sabur Miyan. Ashish Nadkarni is group VP and gen- general manager of worldwide infrastructure at IDC.
He leads the research teams covering compute, storage, cloud, and performance-intensive computing, including the AI-ready data storage infrastructure that will be the study that we'll be discussing today. This worldwide survey captured insights from over 1,000 IT decision-makers across 10 countries and 24 industries.
I got to see a readout of these results from IDC and really wanted to make sure we got a chance to share it with all of you, 'cause there's some just really interesting findings in there. So really excited to have Ashish here to, to share that. Ashish, welcome. Thank you. Awesome.
And then we also have Sabur Miyan, who is the founder and CEO, CEO of STN. They are a partner and customer of Everpure. We've done some amazing work with them. STN is an AI-native infrastructure and managed services provider powering AI companies and enterprises running production critical workloads.
With nearly two decades of experience across data centers, cloud and infra, regulated industries, Sabur brings a rare combination of deep technical leadership and commercial strategy. He launched STN to really re-architect how organizations consume modern compute, delivering production-grade GPU cloud, compliant private infrastructure, and white
Colgrove managed services built for today's AI workloads. His focus on innovation, strategic partnerships, and operational excellence continues to position STN at the forefront of the AI infrastructure economy, and we are so happy to have you here, Sabur. Thank you so much for joining us.
Thank you. I appreciate it. Awesome. All right. Well, let's dive into it. We're going to have a great discussion today. I'm super excited for it.
You know, this is kind of the year that we're seeing AI agents really take the stage, and become ubiquitous in, in so many applications and so many parts of the business. But how do we make sure that they're fast, they're SafeMode, they're reliable, really that we bring them into an operational state, and, you know, kind of move away from that experimentation, to oper- oper- oper To, to actually being
operationalized, and at that enterprise-grade capacity. What we're seeing is that a lot of enterprises are really kind of hitting a wall, and it's not really about the compute or the models. Those are working great in a POC. But when we try to bring them to production, there might be some interesting findings in
IDC's research about, about what that real- really is, right? So again, today we're going to be sharing brand-new IDC research on what's working and what's not when it comes to, really bringing AI to life in practice. And then we'll be talking to Sabur and, and getting his commentary on kind of, what, what it looks like when you're building real AI infrastructure and production.
So we'll kind of have the, the macro view from IDC and, and talk about the which is going to be a really, you know, interesting, really interesting look into what's happening at the macro level. And then Sabur will be able to kind of give us a little peek at, as to, to what this, what this means in, in the real world and, and how they're doing things at STN.
All right. So first, we're going to talk a little bit about the AI production gap. So what's really keeping, organizations and enterpri- enterprises from scaling AI, in production? And, to talk about this, first, we're going to turn to IDC's research.
And I'll, I'll start, start with you, Ashish. You know, organizations are investing in AI, but a lot of projects aren't into production. So let's talk a little bit about that. You recently surveyed, surveyed IT decision-makers. What did you find about the current state of AI projects that are actually
making it to production? So thank you for having me here. I'm excited to talk about, some of the results of our finding, the, the survey that we've done. And, I mean, look, it's not a pretty picture out there.
The fact of the matter is, is that, enterprises are really doing a lot of experimentation, but very slow uptick in production. And, you know, the gap, between experimentation and operationalization or pr- you know, going, taking something into production is it, it appears that it's really is a data infrastructure problem.
So, for example, you know, we do see a lot of companies doing, AI/ML proof of concepts. You know, there's a lot number of GenAI models that created You know, for example that the average is Or our sur- study finds that the average is, like, eight, and the ones in production is, like, five, which is-Even wider for agentic AI if you thi-- if you look
at, you know, under the covers and look at what the details are. Agentic AI is, you know, still in its early stages, so people are still sort of doing a lot of experimentation with their data infrastructure. A lot of manual tasks are being performed behind the scenes. You know, only 35% of respondents really have AI/ML workloads in production.
And, and probably the most underlying reason for it is, is, is the fact that there's a lot of manual processes with data preparation, data cleaning. You know, there's a lot of nuances there that create challenges in taking something into production. Governance is a big issue. There are no guardrails on what data can be used in enterprises and such.
And then we find, you know, the actual proper deployment type. So for example, a lot of companies operate, a hybrid approach or a multicloud approach, and only 6% are exclusively on-prem because of restrictions around, the use of those environments in a sort of a public setting. 10% are public cloud, but those are using data sets that are not proprietary
know, in some cases has been heavily cleaned out. And, and so the vast majority are using some or the other combination. Now, we all know that deploying hybrid cloud on paper sounds great, but when you really try to do it in real life, in production environments, it becomes challenging. And then finally, I think we should talk about the complexity of LLMs themselves.
You know, a growing majority, we would say probably just over 50% tell us that fewer LLMs. I mean, LLMs can be quite difficult to manage in a production environment. So you kind of have to use a mix of cloud-based resources or, you know, open source with some kind of fine-tuning, or hybrid custom-built.
It's sort of all over the place. But the underlying theme with all of this is not having the proper data be the biggest headache or the biggest source of problems for companies as they roll out production AI in production. And if I don't mind adding, to, to Ashish's point, one of the biggest things that we're
seeing in, in that exact space is, is, is people don't actually even know what data they need to feed in. Where they are, where it is, how it's being managed, how it's being tagged. So now that you have this agentic AI push, folks want to feed it a lot of data without any kind of guardrails.
But they don't know where to what to feed, or how to feed. And when it starts happening, you kind of realize, "Oh my God, my data's spread across Salesforce to AWS to, you know, on-premises. Oh, guess what? My, suite data cannot be accessed because there's, you know, different contractual requirements." So managing that entire aspect
is probably one of the most cumbersome processes that we're, we're seeing. Yeah, absolutely. And just to sort of add to that, the data types are different too. So there's unstructured data, structured data, semi-structured data. It's all over the place.
And, putting it all into a common repository that can be fed into, that can be used as a source for the AI engine, if you will, is, the next problem, so. Correct. Yeah, absolutely. Such great points. I, you know, what I I feel like what I'm hearing is that, you know, a lot of organizations are lagging on agentic AI, and
what it comes down to is that that data is really the bedrock of being successful with AI. And the current state it's in today, just like, you know, when folks were, trying to, you know, to do predictive modeling or to do advanced analytics, data's everywhere, right? And it becomes an even bigger problem with AI and, and trying to, to feed, feed that into the models, especially software.
I, I heard you say in a way that, you know, is SafeMode and, you know, keeps built in, right? So, a, a, a lot to consider there, as, think about success. But ultimately, AI success really depends on, on that data infrastructure for sure. Mm-hmm. Mm-hmm. So, that's a great, segue into, one of the next, stats that I pulled from the,
the IDC research, which is really di- diving into why AI projects fail. And you know, we have this stat here where 94, percent of the respondents said that data quality was very important or important to AI success. But 89% acknowledged that they have data quality problems, right? So you can see immediately this mis- this mismatch, right?
It's critically important, but we're just not there yet, right? Mm-hmm. So Ashish, what were the biggest data-related reasons that AI projects failed to reach production? Yeah, that's a great, question. And, you know, just to put a finer point on it, it's not just important,
factor that decides ultimately if your projects succeed or fail. So if you don't address this, this problem or these set of problems, then, it's safe to say your, your projects have a very good chance of not getting into production some substandard results in production. So to kind of build on that, you know, 94% just say data is very important, and then the
remaining, 89% sort of report, some level of data quality, some level of problems with data quality. And then sort of going back to what Sabur was saying, you know, it is something that underestimate, I think, and then come to realize it's a problem, right? So for example, when AI, POCs, fail to take off, you know, or reach
production, you know, people kind of then start looking at where is the data quality problem. And then you look at-What are the, you know, sort of the source, if you will. So time to generate data quality is a big problem, so by the time the data is ready, it's already expired, if you will.
Or you might have inadequate data management tools. So back to Subhor's example, you might have data sitting in different systems connected to each other. So putting them into a single repository or common layer, you need tools for it. You can't just manually write, scripts and such.
Data silos are the biggest bane of existence for companies. We know that. AI kind of brings it out, whether they are logical in nature or whether they are physical in nature. I think overall, if you, if you think about AI and what AI is doing is it's really opening out some of the problems that are there
in enterprises today, but people have out ways to get around them to keep their business running. AI means now you have to kind of bring it all together and operate it as a set of cogs in a wheel, and if your data is not one of those cogs, then AI isn't going anywhere. Yeah, absolutely.
And, you know, just to bring it back to real world, Subhor, I think we'll, we'll pivot and, and talk a little bit about your experience at STN. And, and just for everyone listening, truly we're so delighted to have Subhor here because, you know, the, the fact that they're running, a GPU as a service cloud means that they're out on the cutting edge, and they both can offer services to, to help,
help as you're transitioning into this, world of, of deploying AI/ML in production. But also, they've, they've they're kind of out ahead and can kind of talk about their experience in, in deploying infrastructure to support AI/ML workloads. So it's just such a unique, unique perspective here. Subhor, I wonder if you could, you know, talk a little bit about GPU One, and then talk
about how when you were scaling GPU One, did you, hit data infrastructure walls, and, and what can the folks listening here learn from that? Absolutely. So right before that, do you mind if I cover- Yeah, go ahead. Please do because it actually hits real, real close to, to home for, for us who, So we've been managing and supporting enterprise
clients for a long time, WEKA before the AI, revolution, if you wanna call it, came around. So our genesis has been, how do we take a enterprise customer that is m- that is in the business of not infrastructure, but in the business of shipping jeans all throughout the worldwide, how do you actually go ahead and support that logistics portion of being able to actually say somebody ordered, something in Thailand, and it has to be procured
in Germany to be able to deliver there in a fully, you know, logistic manne- manner. What, what has happened over the last, I'd say, 10 years now is, we started focusing too much on WEKA as in business, and, me as a business owner, I can, I can relate to this. We started really focusing on our ability to get the latest and the greatest system and storage and network and yada, yada, yada, to be able to enable that.
However, the more f- most fundamental, functionality of anything that we are building is being able to, to have data come in, data being processed, and an output happening. That is essentially the, the whole thing that, that, that we as technologists have been trying to figure out, and that, that idea has kind of gone away. Now, what, AI/ML has allowed us to do, the agentic side has allowed us to do,
or pursuit of super intelligence at this point in time has allowed enterprises focus that Go, go back and focus on what truly is the data ingest and data output. And that is, has, has, has, has gotten so much, attention now on trying to figure out how can we actually bring everything back in-house, and manage it.
And truly, the, the position of a s- a, a CDO is, is something that we're starting to see, chief data officer, out there because th- they're more focused on the ins and out of the business and tying it in together. Now, to your question, what kind of, infrastructure walls we've hit. Well, you know, we operate pretty dense clusters.
We also have, a very huge crunch in time to market from our clients. And time to market is it's, it's a race. Everybody's out there to try to create and something or the other in, in, in that's, that matters to them, whether it is, you know, the ability to have robots run around all over the place or being to actually go ahead and automate all
across the board. IDC's data quality, you know, really resonated with us. From our perspective, data quality and production AI isn't just about accuracy or labeling. It's about whether the infrastructure can reliably deliver the right data at the right time and scale. Now, when we're talking about managing our
customers, most of the time they start small because the investment is too high. Mm-hmm. So they're starting small, dab- starting to dabble with eight nodes, two nodes, four nodes, six nodes, you know, a small cluster. And it quickly, it quickly scales. And data structure has not been able to, to keep up with that because they'll start
feeding a lot of data into the LLM because the LLM can take it, or the LLM is learning at the same time. And keeping up with that demand has probably been, been one of the most challenging things that we've had to solve.Mm-hmm. Absolutely. I mean, once those models take off and, and
the usefulness is proven, right, it just, it's kind of a, a floodgate is, is opened. And, you know, I think your point on that race, you know, time to, time to actually get to results, right? Time to first token, I, I hear a lot, right? Um- Yeah you know, that is, that is such a, a big factor here, I think.
Businesses are feeling that pressure, that to stay competitive they have to be using AI. And so, it becomes really critical. And, and it is, you know, I, I think it is a story we keep hearing repeated, that, you know, you have, you have the compute clusters, you have the, the models built. You open up the floodgates, and the data infrastructure can't support it, right?
And that, that ends up being your bottleneck. And so I think that there's this great opportunity for the IT leaders who are listening in today to be forward-thinking and consider, you know, how they can actually get out ahead of that, kind of in the way that, that, you and your team, team have done, Sabur, for sure.
Ashish, I wonder if you can comment a little bit on, what Sabur is describing here and, you know, kind of bringing it back to the macro level and what the data shows. Mm-hmm. How, how widespread do you think that this is? So the bad news is it's pretty widespread. So, you know, going back to sort of the, the landscape, so siloed da- so I'm, I'm just
letting you how bad this is, letting you know how bad this is. Siloed data, 55%. Multiple data storage platforms, almost 50%. Out-of-date data, 48%. Redundant data, 47.
In- inaccurate data, 43%. And then governance issues, 42%. And then, you know, you ask them some of the reasons why this exists, these problems exist. Only 23% actually told us that their storage hardware works well for all their AI projects. 41% actually said it works well but has gaps, and 28% said it just
And so if you tie that off, all to, you know, kind of like, "Hey, your infrastructure isn't keeping up," these gaps, they might work well for your enterprise workloads, for your business workloads, because they're very project oriented, they're very siloed in nature. But when you're talking about AI, that requires a completely different way of
approaching infrastructure. It- it's l- you know, you, you have to rethink the tech stack. You have to rethink how you approach data infrastructure. If you don't address it, then you're not have too much success with AI in production. Yeah, absolutely. Sabur, from your perspective, what are, what
are the attributes of data infrastructure or data storage platforms, that, that fix this issue, right? And that from STaaS's perspective were, were sort of non-negotiables as, as you were looking into, various types of data infrastructure. Well, I think, the s- the silo problem is a pretty
big thing. Specifically when we're talking about trying to move data around or consolidate data, that has been a very major challenge. People have accumulated a lot of data. The one thing, and I'm not sure if Ashish knows this off the top of his head, but one thing that we have done as a, as a humanity is, is, is, is create and store data.
And honestly- Mm-hmm I think that's w- been one of the biggest problems that we have, we are facing at this point in time, is we are really good at creating data. We are really good at storing data. Most storage systems are designed for that as well, which is write lots of data, because that's what they're doing. That's what we want to do,
with infrequent access. However, AI/ML and LLM and fine-tuning and training, or even foundation model you know, take Perplexity for example, you know, scraping the internet, or Google, scraping all of YouTube. That is a lot of inf- a lot of reading.
Okay? Mm-hmm. We created it, and we continue to create it, and that's continuously going to go up, but, but now it's all about reading it. How can we read and understand and tag it? And, then on top of that, bring it really, really, really, really close to the GPUs. Mm-hmm. So the first challenge that we've had to
handle is, okay, we're gonna grab, things from AWS S3, and we're also going to grab stuff from Azure Blob. Oh, and guess what? I also have this, storage array that's, been serving all of my, you know, SAP HANA sitting right here. Let's tie it all together. Cool. W- we figured out how to do that. However, now we have to actually feed that
into one of our data centers to be able to actually get it close to a V300 cluster that can actually learn from it, create it, and create whatever the, the researchers want to do. So that was the first thing. So speed is very important.
Simplicity is probably the second portion. These are complex data structures, or unstructured. And being able to actually consolidate into a singular platform is one of the key things. And I think one of the main things, that I would focus on is, is, and, and push everybody who's listening, push to have a singular protocol.
And by singular protocol, stick with either NFS or stick with S3, something that is more generic, that can actually translate across multiple different storage platforms and multiple different interfaces, because that is probably one of the mo- the biggest things that we have to do as an interim step before we can actually even import most of your data. And then the third ple- piece would probably be compliance and, isolation.
I'll, I'll, I'll, I'll give you my example as an S- as, as a CEO in STaaS and, and we're building STNAI in-house. I do not want my sales operations people or my engineering folks to be Yeah. Absolutely. That, but when I'm training-Or fine-tuning, all of that data is fed into it.
So creating that governance, creating that, that com- compliance Well, well, not even just compliance, creating the audit trade trail that, nobody's going to ask STNI about, somebody's, you know, personal data is probably one of the key, well, really big challenges that we're working on trying to solve. And then the last piece, linear scalability.
Scaling is big. We want to be able to go from, you know, and, and I'll talk about GPUs, but really I wanna talk about petabytes, right? We wanna be able to go from petabytes to exascale without having to rip and replace the entire infrastructure on a five-year cycle.
And by the way, five-year cycles don't exist anymore for data structures. There It's, it's a living, breathing data pool that's sitting on infrast- on, on, on, on storage systems that are being actively read. Yeah, yeah. You need that living system. We can't think about five-year rip and replace cycles anymore, right?
It has to- Nope grow and expand and flex for what the new demands in three months will be, right? Not in five years. Oh, yeah. So yeah, absolutely. Yeah. Absolutely. Three months. Yeah. We're, we're, we're living, l- we're
living WEKA by WEKA at this point in time, Erin. WEKA to WEKA. WEKA to WEKA. Oh my goodness. Yep, yep. We don't even have the luxury of months anymore. Nope. Yeah. I just wanted to, provide to, to Subhoor's comment about how much data is generated.
So IDC, we have a product called Data Sphere. We are estimating it is going to be short of 400 sextillion bytes, so that's like 400-ish ter- zettabytes. That's It's crazy, you know. And I Like, start counting zeros.
There's a lot of zeros to count, so Yep. Yeah, absolutely. Yeah. Absolutely. So, Subhoor, I'd, I'd love if you could kind of just take it one step further. I think you've painted such a vivid picture of, of, you know, what, what you all are doing to, to build the data infrastructure to, to support, what STN is doing in practice.
I wonder if you could kind of, you know, tell us a little bit more about how you're building that production grade AI infrastructure and about how you're really bringing AI to life with that. AI to life. That's a good one. Well, our, motto is probably the best way to put it, our mantra, is to be able to provide the best
technology stack that we can to our clients, whether they are enterprise or whether they're researchers. It That's our, our, our bread and butter. We are infrastructure folks. That's what we do. So creating the best private cloud services is kind of our, you know, MO.
And I, I see, you have a slide up there for 99.999%. That's five nines and six nines and node availability, and it takes a lot to actually achieve that. Most for, most of what we see specifically in GPU one or GPU-as-a-service business right now, SLAs are about 99.5%.
And, and Pure part of that is because researchers are, you know, can take a break if something breaks, right? So they're able to actually go ahead and sustain some sort of an outage. However, that's not going to be the case as we progress in the next eight, 12 to 18 months, and I'm not sure what you're seeing, Ashish, but in the next 12 to 18 months, most of the,
AI agents are going to be running around li- w- wild, and the bigger thing is going to be around inference, right? So being able to read the data, feed the data, and, and inference upon it, and that requires the shift back from 99.5% SLAs. And, and that's why we've never deviated from our original mantra of providing the best
availability and best, infrastructure across the board, whether it's GPU, inference, training, production, actually something that's not talked about anymore because we're so gung ho about GPUs. CPU. CPU is still required. It's just ov- get, gets overshadowed by the amount of spend that we have out there, on the
GPU side of the house, but none of if you don't have the CPU infrastructure. So being able to bring that to, to, to market is one of our key traits, and we've sustained that across the board. And on top of that, especially customers that are paying, as, as a private cloud services for GPU-as-a-service, which is done on dollar per GPU per hour, it is, it is actually wasted funds, fa- wasted money if a GPU is
down for an hour or two hours or three hours or in, in an ActiveCluster. Even if it's a small cluster or a large cluster, the percentage matters. You want to be able to have your infrastructure up and, up and running as much as possible, and we've done some pretty cool, amazing things around that, by, I'm, I'm sure you're familiar with the liquid cooling side of the house as, as, as these things run
pretty hot nowadays. So, when we are we approach building a data center, for handling GPUs, we make sure that we have redundant, crack units, we have redundant, reservoirs on top of what the data center itself is providing. We're, we're, we're building that on skids next to the infrastructure so we can sustain
longer, longer, longer outages. I believe the last one that we did was, to sustain three to four megawatt build-out of, V200s, and we could sustain 19 minutes of liquid running through, without any kind of, you know, external generators to kick in or UPSes to kick in, or media systems to kick in. So that's our focus.
Awesome. Awesome. And, Ashish, I, I think in a moment we'll, we'll switch gears and talk a little bit about the Everpure and, and STN partnership. Before we do that, was there anything that you wanted to, kind of comment on, from what Subhoor shared about how, how they're bringing AI to life in practice, that's
what's happening, with the clients that, you work with?So I'll say, you know, what Sabur is talking about is exactly what the market needs right now, and I'm glad that he's, he's helping customers out. I mean, look, the fact is that AI infrastructure is a very expensive it can't be just left to waste because of, inefficiencies in how you manage your data,
inefficiencies in how you prepare your data. You have to make sure that those resources are put to good use because they're not just expensive, but they, there's, you know, otherwise you'll have a lot of tech debt that you have to deal with. So you have to use them at the appropriate time, and to get those insights out in an appropriate way, it, there is a sense of
urgency to it. So I, I think that's what enterprises need to keep in mind, you know. A server and a GPU in, in those servers are not gonna just last you six years. You have to make sure you make the most out of it. Yeah. Yeah, absolutely.
Absolutely. Sabur, I'd love to talk a little bit more about, how Everpure and STN have, have worked together to, you know, to bring, bring AI to life. And, you know, I think, in sometimes, you know, we're over here on the, the data storage and data infrastructure side, and we're, we're a few steps removed from kind
of ultimately what we're trying to help customers do. I think you're, you know, kind of one step closer, right? We're helping you help your customers who are the ones that are doing, you know, like you said, they're trying to ship jeans, worldwide, or they are doing fraud trying to, you know, set up customer service agents in, in healthcare,
things like that, right? But, you know, really I'd love to talk about how we've partnered together to, to bring, bring those projects to life, and, and how we've worked together to solve some of the challenges that, that Ashish, mentioned, in the, in the research report that, that so many IT leaders are facing, right?
The, the performance at GPU speed needed, that operational simplicity, and then also the, the compliance-ready infrastructure. Can you tell us a little bit about the partnership and how does the Everpure platform fit into GPU One, and, and what has that meant specifically for your know, really what matters here?
Ah, absolutely. So, Everpure and STN relationship goes back to our, founding, 80 years ago. We, the founders, me, Anamar, and Tom have been fans of Pure, and back then it was Pure Storage. I, I know we're gonna talk a little bit about the rebranding, Everpure, but, but, we've been a fan of Everpure for a long, long,
long time. The ability to simplify data management ha- goes back to the initial kind of ge- you know, creation of, the Pure Storage system. We used to have what we used to have was storage engineers. We had a bunch of guys or gals running around making sure there's knobs tuned the application owners are able to, to have the best performance or move IO around.
I can't remember the last time I had a conversation around IO, like, "I need this much IO." 'cause everybody wants all IO, right? But, anyways, I'll, I'll fast-forward to now. You know, we've standardized our AI data layer on FlashBlade because it aligned with our requirements. Specifically with EXA now, in particular, we,
we, we really love the new architecture of FlashBlade//EXA. One of the key things from data tagging perspective, being able to actually move data around, being able to actually make s- data, the stale data active and hot and move it around and be able to present it to different types of GPUs, is, is has brought to, to, to, to market it.
And, and for folks that don't know, I'll explain a little bit on the technical side what the, what the, what, what, what why this is different than everything else, that we're seeing out in the market. Metadata. So Pure engineers, and specifically with EXA, has taken amazing technology on FlashBlade, which we used to use for all kinds of HPC deployments Forever, and essentially created
and moved the metadata, which is the tagging portion, which is the ability to, to know where, what your data structure is, and split them into two different environments that work together. So when I'm addressing a large cluster, I don't have to go read a whole bunch of small files because I have access to the metadata.
So the metadata allows us to go ahead and index everything that's already in the data nodes and be able to retrieve the response immediately and know exactly which block or which file to go ahead and pull. So having that combination at really fast 400 GB/s per data node or multiple data nodes is immensely important when we're talking about
mission-critical, workloads. Yeah, absolutely. That metadata and, and that, you know, the, the way it can help speed up, delivering data to the AI models is, is really critical for, for the AI success. That's, that's fantastic.
Yeah, I, I wonder if you could comment a little bit more on, the experience with EXA, kind of, you know, what, what, you know, you sta- started on FlashBlade. What was the experience like going to EXA from that? You know, ta- talk a little bit more about that part of your journey. Oh, sure. Absolutely.
So, FlashBlade, like I said, has been our standard across all clusters for a very long time and, and FlashBlade//EXA builds on top of FlashBlade, right? And it-The, the, the journey was that we were able to saturate the, the Okay, let's backup. All of our clusters are built with, with multi 200 GB NICs, which is about per node around 400, about four 200
gigs, so it's about 800 GB total. That's 800 gigabits per second. That roughly translates to 90, 95 GB per second- that each node is capable of. In a cluster of, let's say, 100 nodes or 200 nodes, that's a lot of data for a singular or multi- you know, a singular storage system to pull.
And, and that's where, where we were, we were, kind of getting into a bottleneck of the, the FlashBlade being able to actually handle all of that, no matter how much more, throughput we gave it. So having the infrastructure, or having the storage subsystem with EXA separate allowed a much better scaling option for us to be able to handle multi- multiple hundreds of,
nodes all communicating in the same cluster. And con- and, and because we're trying upon the same data and trying to read same data all simultaneously across m- and, and loading it up into GPU and NVRAM, that was one of the key, key differentiating factors for us to kind of now standardize almost everything on EXA across the board.
And now I can tell you, I can I, I haven't b- I haven't reached a saturation point yet. The largest cluster that is hitting a single FlashBlade//EXA is 192, nodes. And I have I'm, I'm getting consistent 95 GB/s every time I need data. Doesn't matter- Great how large the job is. Yeah. That's, that's fantastic to hear.
Yeah, it's, it's exciting to know that, that it's working out in practice, right? And I know that, you know, you and your team were, really part of, you know, part of the early testing, as well with EXA and, you know, making sure that, that everything as, as it was supposed to, so it's, it's great to hear that. From, from what I've been told, I believe we were the be- we were testing
when we had a production need, and the beta product, went pretty quickly into production for us. And, I can tell you, testing for this particular client that almost everybody knows about, because this, that's, the GPU world, the experience that they had with, with the storage subsystem that we provided them was unparalleled.
Awesome. Awesome. That's fantastic. I love it. Ah, amazing. Well, so, you know, we're, we're sort of getting to, getting to the, the later part of the webinar, so I'd love to talk, before, before we, we lose you both, I wanna hear a little bit of your expertise on where
you think the market is headed next, and I, I think that would be really valuable. Just for everyone who is on the line, we're live here. If you have questions for Subhor or Ashish, please put them in the Q&A section. I am, reserving some time at the end to ask some of those questions live, so, I Or to answer some of those questions live.
You ask them live, we'll answer them live. So please do go ahead and, and, put those in if, if you have anything you'd like to, to, ask, these, these experts here. All right. So let's, let's talk about where, where we're headed, the, the future of AI data infrastructure.
So Ashish, you know, based on the research, what's coming next? What should infrastructure leaders be preparing for next? So, one thing is for sure, that the tech stack is going to be completely reimagined and re-envisioned for AI at scale in production. So the future is gonna belong to unified AI-ready data platforms that are part of an AI
factory set up, an infrastructure, if you will, right? You probably heard that term AI factory being mentioned more and more. I think the idea is they're all optimized to work with each other. You know, I think this is where, I would say Everpure and the partnerships that you guys have with companies like Cisco, I think are going to really be important because the, the
stack needs to be defined and optimized and such. I would say that, you know, the other part that doesn't get talked about is people still don't wanna give up on the enterprise features of that stack. Mm-hmm. So they want performance, they want data mobility, they want the management tools, they want the ecosystem, they want
the security and resiliency. So all of these things are important to that AI factory stack, to that tech stack. Then we sort of w- go into the whole agentic and how agentic is influencing it. So right now, nobody's going to tell you that agentic is, mature enough to be enterprise-wide, right? It's still we see 35% or so.
But it's gonna accelerate. Where is it gonna accelerate? I think it's about the data management side, it's about the workflow, it's the automation. It's, it's, it's taking care of things that are very manual today and need to be automated so that the time to insights can be compressed, and you can get your
insights quickly, and then you can realize the business value out of it. I think, data access, governance, you know, these are table stakes. They are going to be part and parcel of the stack. Nobody's gonna have to scratch their head to wonder if these things are there or not. Mm-hmm. It's going to be table stakes.
Um-Then the other thing is the value of AI as a function within the storage itself, right? Today, a lot of, storage vendors will bolt on AI to it. But what we are hearing loud and clear from these respondents is, "Hey, we want AI built into storage." So whether that's for performance optimization, threat detection, capacity optimization, all of these things that were done in a sort of a very rudimentary
machine learning-like fashion, people want advanced AI capabilities to manage that. Even things like data localization. All of these things are part and parcel of AI making storage better. And then the last thing I would m- mention is that data management platforms are a high investment priority for organizations, and I, you know, note here that your intent to
acquire 1touch I think fits right within that sweet spot of what people want. They don't want data management to be a bolt-on to that AI factory. They want data management to be part and parcel of it, integrated, not piecemeal. So, you know, and you're probably seeing a theme here, right? The theme is no more piecemeal.
We want an integrated stack. We want that stack to be complete, and we want it to be- Awesome. Awesome. I think we lost you right at the end there, but, but hear- hearing you loud and clear on we, we don't, we don't want piecemeal solutions, right?
And, and I do think, You know, I've, I've been here at, Everpure for a few years now, and I've, I've actually covered a, a couple different areas, including, you know, when we sort of, you know, expanded our, our partnership with Nutanix and have, you know, looked into new virtualization options, and now I'm on AI. And it strikes me that, any time that we are, you know, looking to, help customers
understand how the Everpure platform can, can help them succeed, that the same things continue to come up, right? It's the, you know, it's the unified platform that makes it really easy mission critical workloads as- alongside your AI without having to rip and replace or go buy a, you know, a new point solution, right?
It's things like the governance built in. It's that enterprise grade, security and reliability, you know, all, all of those pieces. And, I also love what you talked about with, you know, AI actually built into system itself.
We really have not talked about that on these webinars at all. I feel like we could do a whole, you know, a, you know, a whole, whole webinar on that piece as well, 'cause there's some really cool stuff, happening there, that it would be, would be great to talk about too. But yeah. And then that data management
piece, so critical. It absolutely needs to be a part of what the, storage system provides and offers and shouldn't be kind of a bolt-on later, for sure. All right, Sabur, you know, you're out on the bleeding edge building AI infrastructure every day.
What trends are you seeing that infrastructure leaders should pay attention to? Well, I'll, I'll springboard right off of the last point that you just made. Our AI agents are retagging everything that we have in life right now as we speak. Mm-hmm. It is reorganizing our data structures through and through. So agents are capable of being able to comprehend what exactly does
the business need for what side of the business, and re-architect, redesign all of that, now. Now. Not tomorrow, but now. And that's what we're gonna see moving Now, the question really comes down to it is how do you actually go ahead and create the boundaries, for the AI agents?
I will tell you, I was surprised. We created boundaries, for some of our, our engine- our, our, our, AI agents, and, they worked around it, which was surprising to us. And when we caught them, they apologized. So it is a pretty interesting time of, of how these agents are being able
to, feed the data that we've given them and be able to actually go ahead and circumvent some of those boundaries that we've created and still be able to 'cause I think this is more focused on results driven portion, right? They're results driven. They don't have the wherewithal to say, this is where I need to stop." and then I
think that needs to continuously develop, but we will see where that goes. Now, your original question about where the trends go, inference. I cannot stress upon the fact that as all of these companies release products that are mostly based off of, agents or LLMs, they will require inference. And inference is very, very important to be tuned to your data.
And for enterprises, it's about compliancy, it's about security. It is about making sure that threat actors are not able to access data. That is probably one of the biggest, things that we see for enterprises before enterprises can really adopt. Yeah. Yeah, absolutely.
Absolutely. All right. Well, I do see some questions coming so from Guillermo, he, he has a question about, wanting to know and how efficient AI can be with respect to energy consumption. By 2030, it's estimated that ener- energy consumption by AI will be 11 times higher than
in 2023, and data centers could consume up to 4.5 of global electricity. Um-Energy usage is obviously a, a really critical, discussion point for the use of AI, right? And in terms of sustainability and, and responsibility, and also just feasibility, right? You know, I, I don't know that you can comment on, the, you know, the global state of, of
energy in the world, but maybe, you know, Sabur you could comment a little bit on, what you're seeing, with your work in terms of making AI more efficient and, you know, what you're doing in terms of, of energy use and, and dealing with, dealing with the power needs there. G- do you mind commenting on that a bit?
Yes. Great question, Guillermo. I actually have a lot to say about that. We hear a lot about the, the, the fact that the grid is, depleting very fast. The cost of power is going up, for community members across the board. And, and it's going to continue going up.
There's a couple of things that I will point out there. A lot of it has, has to do with the fact that the policies that we've created have not been modified for the energy susta- for the energy needs that we have as, as population, not only just in the US, but in the world. So we're, we're, we're, we're chasing, right?
So we're always behind the curve. So that's the number one po- point. So it seems like it's a lot of, a lot of energy. Overall, data centers are projected by 2027, at the end of 2027, to consume, I believe around And, Ashish, you, you, you might know these numbers better than
me, but, right now, it's less than 2% and it's going to go up to 4% of the power consumption, by the end of 2027, on everything that has been committed to be built. And, if you look at it from a grand scheme of what AI is solving, that is not a major number. Now, this last point I'll talk about the energy portion is, is,
is the output continues to get efficient. So right now we're at that stage where there's a lot of power required for a lot of GPUs because there's a lot of training. There's a lot of fi- a lot of things happening, and we don't see the output yet. The output will actually B- the way we're seeing it is not linear to the more power you
feed it, the more GPUs you feed it, the output is, is exactly the same. No, it's the other way around. It is going to get more and more and more efficient. Over a period of time, this is This will not be a, a major problem because we're not pumping out a whole bunch of GPUs just to do a lot of training.
There are, there are, there are point in times that we have to keep in, keep track, requires power. But after that, it is just, it is just building upon that. So gradually this is going to be a non-issue. Great. And Ashish, I wonder if you could comment on, on what you're seeing as kind of the, the macro trends for this as well.
Well, I'll go back to my other, so comment about, resources are precious have to make the most, most out of them. You know, so buying GPUs, yes, generationally they are gonna become efficient and over time, you know, compute is going to become a lot more, economical. But till that time, you have to make sure that whatever resources you are using are
best put to use and there's a return on investment that you're planning for. Yep. Yep. Absolutely. Absolutely. So foresee some, some, better efficiency with compute going forward, but also a need for this to be an active, policy discussion as well in terms of, how we're making sure that the, the energy that is needed is, going to be
available in a, in a clean-room and sustainable way for sure. I feel like we could do a whole webinar on this topic too. Um- Oh, easy. We might have to Yeah. Might have to add that to the list. Awesome. Great, great question, Guillermo.
And, just looking at the time, you know, I think, it's, it's probably time to, fire my, my last question over to, both of you, Sabur and Ashish. So in closing today, we'll start with you, Ashish. What's, what, what's the one thing that you want every infrastructure leader on this, this webinar today, to take away, from this discussion?
So I'll just say that AI-ready data infrastructure is the foundation. Without that foundation nothing's gonna be possible or scalable. I mean, your peers, CIOs, you know, stay s- say emphatically that, data quality is critical to AI success. But lack of suitable infrastructure, a lot of lack of suitable,
platforms that are integrated, will not fix the data quality problem or make it better, right? So you have to have the right And investing now, planning now and investing now is really going to set you up for the future. Underestimating it now is going to make it pretty problematic down the line because this
situation isn't gonna improve. So my piece of advice would be don't underestimate data quality, don't underestimate security performance. These are all, table stakes, and they will ultimately decide what success is and what failure isn't.
So Great. Yeah. Sabur, what's, what's your, what's your one take? My one take is, you know, AI success isn't really about GPUs. Uh-I- it's about whether your infrastructure behaves predictably at the scale that you need, and, whether or not, you know, the pilot AI that you're working around with is, goes into
production, right? That's one of the key things, and it requires discipline, it requires, storage, it requires networking, security, day-to-day operations, management, scaling. And, and, you know, the decisions that you make as an infrastructure, you know, owner is really the difference is whether or not, you know, it's, it's, it's AI is, is a,
is a point on a slide, or whether or not your AI is actually making the slide. Yeah, absolutely. Mm-hmm. Absolutely. Yeah, it seems like the research is clear. AI success really starts with, with your data infrastructure. Thank you so much, Ashish, thank you, Sabur, both for, for taking the time today to, to
share your insights and your perspective. It was absolutely wonderful having you here. Thank you, Erin. My pleasure. Thank you. Take care. Thank you so much. Bye. All right. So, in closing, if you are
attending GTC in a couple of weeks, and, you know, would like to, to have a live discussion about, how Everpure works with partners like STN and NVIDIA, come see us, at our booth, at number 935. I'll be there probably the whole time. So, so do come see us.
The STN team is also going to be there as well. Sabur will be speaking on the Deep Tech Nexus panel on Monday, March 16th, hosting an invite-only event called AI Loft at NVIDIA GTC that I believe we're sponsoring as well. This is going to bring together AI builders, enterprise teams, infrastructure leaders to really foster these conversations so you can, you know, talk to your peers, share, share
those notes, live. So, yeah. Sabur says DM him, on LinkedIn if, if interested in attending any of that or connecting, so, so please do that. And of course, we also have, Pure Accelerate coming up, in June.
So registration's open for that right now. Again, would love to see you all there. And if you can't make it to see us in person, you can always, always join us, on the, the community. Lots of lively discussions happening there, amongst, folks here at Everpure, but also,
more importantly, discussions between all of you, helping each other, do, do your work, work better and, and, and, sharing those insights with one another. Thank you so much for joining us, and we'll see you next time.
Erin Stevens
Senior AI Product Marketing Manager, Everpure
Ashish Nadkarni
Group VP and General Manager, Worldwide Infrastructure, IDC
Sabur Mian
Founder and CEO, STN
What's behind AI agents that are fast, safe, and reliable? It starts with smarter data infrastructure. Everpure and IDC have been studying what separates organizations that are pulling ahead in AI from those that are stalling. Join us to see the research, hear what's changing, and rethink your data strategy to get ready for AI. Plus, hear from AI pioneer STN on how they're bringing AI to life with Everpure.
In this webinar, we’ll share:
- Drivers of AI performance today
- New IDC research on how leaders are building for AI
- What separates fast-movers from the rest with a look at how STN is leading the way
- A preview of where Everpure and IDC see the market heading
We Also Recommend...
Personalize for Me
My Updates
No updates at this time.