Vrashank Jain de Dell Technologies sur le problème des données qui pourrait faire échouer votre IA

Transcription

Welcome to East Speaks, where we explore the technology shaping the future of enterprise IT. I'm your host, Corey Nolles, and in today's episode, we'll explore the data challenges that can make or break AI initiatives, how enterprises can simplify their data environments, and what organizations need to consider as AI workloads continue to evolve. Joining us today to talk about these challenges is Versank Jin, lead product manager for Dell's AI data platform.

Versank works at the intersection of product innovation and enterprise strategy, helping organizations rethink how they manage data to support AI at scale. Versank, welcome to the podcast. >> Corey, it's lovely to be here. Thank you. >> Awesome. Awesome. We're so glad to have you. I I guess to start off, a lot of organizations say they want to do AI, but they quickly run into data issues. From your perspective working on Dell's AI data platform, what's the most common data problem that stalls AI initiatives today? >> I think a lot of people do say, you know, data is the problem, but uh we need to add more color to that question.

So, the number one thing that I really see stalling AI projects isn't the model. It's not a model problem anymore. It's really data readiness. And it's not just the data. It's actually the readiness and the ability to go and find it. So every organization, especially large ones, have uh enormous amounts of data, but it's fragmented across systems. Some on prem, some in the cloud, some in databases, some in Salesforce and Service Now and and streaming sources, etc.

None of these were ever designed to talk to each other. And you've got structured data living in a warehouse. You have unstructured data. think images, video, emails sitting in a object storage or a file share. You got logs coming in from all over the place from your edge and so none of it is in a state where the model can actually consume it. And the second layer I feel is the metadata or the lineage. This is becoming way more important now in the last maybe year or so than ever before.

Teams spend weeks just trying to understand what data they have, whether it's clean, whether they're even allowed to use it for a given AI use case. And by the time they've answered those questions, the project has lost momentum. And what separates people from fast to slow are the ones who usually have a great way of understanding governance and cataloging of all their data before they really started to chase AI. >> That makes sense. And that's definitely a thing that has to be has to be fixed up front.

So for companies that have that data scattered all across clouds and on-prim systems, edge devices, what makes building a reliable AI data pipeline so difficult? >> Yeah, I think the the first thing we have to do is make sure that we distinguish between what are AI pipelines and AI data pipelines. AI pipelines can sometimes be confused with agent workflows that are now becoming very popular with, you know, uh uh anthropic and and uh lang chain and and those types of things.

Yeah, >> the AI data pipeline very specifically is about how to make sure that the models get the right amount of data at the right time at the right place. >> And the core challenge I think there is because is the fact that we've thought about AI data pipelines traditionally in the same way that we thought of ETL pipelines which are batch oriented. They are relatively predictable. They can afford latency. And AI pipelines on the other hand are are much more throughput hungry.

They are latency sensitive. they they need data in a specific format and are often different from how it's stored operationally. >> And so when you add multicloud and edge to that mix because the data is all over the place, you're not just dealing with data silos. You're basically dealing with network physics now. Moving large data sets from a cloud bucket to a GPU cluster on prem for training. That has a real cost and latency implication. And the consistency problem is significant, too.

I mean, if you're training data and you're inferencing, inference time data are coming from different systems that aren't synchronized, your model starts behaving in ways that are really hard to debug, especially because now the models are non-deterministic, meaning you can ask the same question to the model a couple of times, you might get a slightly different answer. And so, it's very hard to trace back where the source of the the prediction was.

Um, and so the pipeline actually becomes a source of model error that will feel like it's a model problem when in reality it's a data problem. Oh wow, that's really interesting. So, everyone says AI needs good data, but that phrase can mean a lot of different things in practice. What separates data that's actually useful for AI from data that just sits in storage? >> This one's not that hard to answer because it's we're I think we're have to go back to basics here and we we did this, you know, maybe a decade or so ago.

The phrase good data, to your point, does get thrown around a lot. uh but I break it down into basically three things roughly. The first I think is relevance. >> Mhm. >> Even if you have a lot of data, does the does the data you have actually represent the problem domain? A lot of enterprises have like huge archives of historical data that is technically clean but it doesn't reflect current operational reality though. >> Yeah. >> So that's one. Number two is completeness and consistency. meaning you might have missing values, you might have schema drift over time, you might have inconsistent labels to the PDFs or the emails. >> Um, so that's problem number two.

Problem number three is accessibility. This might not be the case for relatively smaller organizations, but for large or even mediumsiz organizations, even if you have high quality data that's that's maybe if it's use, you know, useful, if the retrieval path to get that data is really slow or the format isn't compatible, it's useless. And so it's the combination of is it relevant, is it complete, and is it accessible in the right way at the right time is sort of the the thing that differentiates good data from any other data. >> And the one thing I'd add is >> AI specifically, it's all about representational diversity.

I mean, your chat bots will do a better job if you're feeding it not just an email conversation, but also allowing it to go query a database to doublech checkck, you know, what it actually has come up with. So reasoning models are now trained to ask questions, double check their answers, which means every time it does that. It would be really good if it had access to a lot of really good data. >> Yeah, it would. Yeah, it would. So we're talking here essentially about the idea of of bringing data closer to insights, which sounds simple, but technically is quite the challenge.

I understand. Why does data location matter so much for AI performance? And what problems arise when data and compute are too far apart? Yeah, it this is a it goes back to that old adage of uh moving your insights closer to the data rather than the data closer to the insights because frankly data has way more gravity than anything else anything else in the entire organization. So, and this is again where physics maybe reasserts itself. GPUs are very fast, but they're only fast when they're fed fast.

If they're left hungry, that's money left on the table. Y >> and so if your storage can't keep up with the throughput demands of a GPU cluster, which let's be frank, it's becoming double, triple, quadruple every year, uh we're starting to talk about hundreds of gigabits per second, you end up with GPU starvation, and that's the most expensive line item in your entire IT budget right now. So, it's one of the most common costly infrastructure mistakes we see, which is we jump to the latest and greatest GPU and we underinvest in storage and storage throughput.

And this is especially important when you're talking about inferencing for large number of people with bigger and bigger models, right? If if your model has to reach out to a remote store to pick up context or embeddings or even look at old tokens that is already generated, you're adding network round trips back and forth and what should be a net net subund millisecond adds up and starts to become a a you know a 1second latency which is frankly untenable.

Um so this is where we really start talking about data locality meaning if your data if your most closely guarded data sets are on premise it would be much more feasible to bring that compute closer to that storage potentially in the same data center maybe even in the same cluster subnet and that's really where our storage solutions like power scale are designed for that kind of sustained high throughput when it's living side by side with the GPU cluster if that makes sense. >> That makes a lot of sense.

It does. It does. And it sounds like a lot of money left on the table if you're not careful. So enterprises aren't just running one AI workload anymore. They're experimenting with training, fine-tuning, inference, analytics, and even more. How is this growing diversity of workloads changing what companies need from their data infrastructure? Yeah, this this is a really hard one because I mean like maybe even two years ago when we said AI, we really had just one or two use cases in production and they were primarily inferencing because we were just we were just wowed by the fact that an LLM can start to do things in natural language. >> Yeah. >> Now fast forward even a midsize organization is likely running training jobs because the the models have become so good and so small but yet so powerful.

They're running fine-tuning pipelines. They're running batch inferencing. They're doing real time inferencing. And don't forget, they're still running old analytics. That has always been the case. >> And so each of these workloads have a very different IO profile. Um training is sequential and throughput hungry. You really need to push a lot of data very quickly. Realtime inference is random access, right? It could ping any object, any file, anywhere.

And it's very latency sensitive. analytics might need a big scan across really old archive data um to find that needle in the haststack. And so what I'm basically going towards is there's no single storage tier that's actually optimal for all of them. And so the infrastructure question becomes how do you build something that's intelligent enough but capable enough to serve all of these patterns without forcing any one of these teams to compromise for somebody else?

And how do you do this without maintaining five different storage systems? That's the big part of where our our AI data platform is focused because we're trying to give them a semblance of a unified data platform but with the ability to have diversity for workloads uh without exploding the operational complexity of how to stand this up and run this over time. >> Yeah. So a lot of the conversation around AI focuses on models and GPUs. But in in your experience for Shank, is data infrastructure actually becoming the bigger bottleneck in enterprise adoption? >> Honestly, yes.

And if you look at GTC this year where Jensen uh talked about Nvidia's investments in QDF and QVS, I take that as a message from Jensen to his investors which are frankly everybody at this point that data is a big problem and if Nvidia is focused on solving it, it's a pretty good sign that it's it's happening all over the world. You know, for the last few years, you're right, the do the conversation was dominated with models uh GPU procurement. I mean just in the last one year we've had six or seven different models become number one at various points of the year >> and so I think customers are now starting to see that the models are getting good enough but the investments are flowing now to vector databases data orchestration feature stores it's the stuff that has always been on our road map we just never got to it because we never had a killer use case we finally have a killer use case >> and once people have gone beyond the GPU shortage that gets a lot of press uh they're going to start to talk about the bottleneck of data architecture because you just cannot feed these GPUs well enough right now. >> Agreed.

Agreed. And it's not going to slow down. Many IT teams feel overwhelmed by the number of tools involved in modern AI pipelines like data lakes and vector databases and feature stores and orchestration tools. What does simplifying data management for AI actually look like in practice? Oh, it's the million-dollar question. Um, there's a very maybe a trillion. I agree. Um, there's a VC uh investor, his name is Matt Turk, and every year he does a a MAD landscape.

It's called a machine learning, analytics, and data landscape. And he started about maybe five or six years ago. And that was like a, you know, a handful of logos, maybe about 50. you look at the latest one and you have to basically scroll through the page to go through all of the logos at this point. So it is a very fragmented very diversified market. I don't think that's the problem in itself. I don't think it's the fact that we have variety of technologies.

The real problem I think is reducing the number of handoffs between uh a data scientist or an ML engineer who has to navigate this tool landscape. Right? In a lot of enterprises today, you have data sets being getting ready using six or eight different tools that don't necessarily talk to each other. So a data catalog, a Spark engine, a feature store, a vector database, a pipeline orchestration layer, an observability tool, a monitoring system. The problem is all of these things, they sounds really good when they look at them individually.

They sound like a nightmare when you have to manage this at scale. So, in my view, simplification doesn't necessarily remove all of these tools or condense them into one because let's be real, it's never going to happen. >> Um, I think it really looks like a unified metadata lineage layer that's sort of sensing things as they're going from tool to tool so that you as a steward or as a as a data owner, you're not worried about which tool the data is going into.

You're just worried about making sure that you know at all times which tool the data is getting into, what's happening to it, how it's reporting back. So it's so it's really all about not losing track of where the data came from, where it's going, and where it's headed next across the tools that will always be there. >> Good unified logs kind of. >> Yeah, agreed. Some people say we need a unified tool. I wish there's a tool hood that does everything for me.

One, that's a pipe dream and two that is a recipe for a significant vendor lock in. So yeah, >> you've worked with large enterprises and Fortune 500 organizations for years. >> What are some patterns you're seeing among companies that are successfully scaling AI versus those that are still struggling? >> I think there are some good companies and maybe there are some companies that have struggled and I've started to get some patterns, right? So the companies that have usually done this well.

I think they do a few things really well. Number one, they've always treated data as a first class product, not as an afterthought. meaning they have robust organizational guidelines or or structures in place so that there are SLAs on data quality. There are data owners. There are pipelines that are monitored in the same way that you monitor production software. So they have the hygiene right. Second is that they've built for iteration not for perfection.

By that I mean you start with imperfect data, you invest in feedback loops and you basically let it grow to better quality over time rather than not getting off the ground because you don't have the perfect data set which frankly never materializes. And the third thing they do really well is that they make deliberate choices about where AI workloads live. Meaning they're not just lifting and shifting it to the cloud because that is a cloud mandate.

They're thinking about data gravity. They're thinking about regulatory. They're thinking about cost per inference. And so the more intelligence you put into the process and the more decisions, small decisions you make, the more nimble you are. On the other hand, if you look at the ones that are that struggle, I feel like they're either trying to boil the ocean, right? They go into this 2year, threeyear big data transformation journey that never materializes or that they're running too many maybe isolated experiments, right? which are an idea here and an idea there, but they've never really connected the dots to say, are we thinking of a coherent platform here, or are we just trying to solve one thing at at a time?

So, those are the two kinds of patterns that I think I've seen between the the good and the bad. >> That really lines up with what you might expect. So, let's talk a little about AI transformation. It's not solely a technology shift. This is this is a thing that often requires new collaborations between data engineers, ML teams, and IT infrastructure teams. How are you seeing organizations adapt their teams and workflows to handle this new datadriven AI environment we live in? >> It goes back to like the teams that are doing doing this really well.

One thing I I forgot to mention is they've always they've also broken down the walls within which each role lives any longer. If you go back a few years ago, we knew exactly what a data engineer did versus an ML engineer versus a data analyst and a and an AI engineer. Now, there's no such lines anymore. Everybody can build anything. >> Which means the companies that have been able to be nimble and said, "Hey, your job isn't just to do data engineering.

Your job is to think about AI at scale so that you're building the right pipelines for AI in mind." They are forced to think about beyond what they're doing on a day-to-day basis. So at at scale I think it requires much tighter collaboration and so the good organizations are ones that are combining an AI function into like every single skill set that they have rather than still siloing them into you know what they were traditionally doing. This does require a technology answer I will say which means you do need the tools to allow people to go think about other things.

Um but largely I think you're right this is a this is an organizational workflow question rather than a technology question and I'm also seeing machine learning engineers just maturing very quickly and the MLOps tools are becoming much better too right they've all started to call themselves agent ops or or you know AI ops for a good reason because they're seeing that traditional ML is >> is still there it's just being repackaged into a bigger AI project >> of which ML is now a smaller part not not the big Yeah.

Yeah. It's more like the root. >> Exactly. >> If we fast forward, let's say three to five years from now, how do you expect enterprise AI data platforms to evolve as models get larger and workloads expand, real time AI becomes a lot more common? >> I think the few things I'm fairly confident about is I think we've already talked about this, the the line between compute and storage I think is going to blur. uh we're already seeing processing coming to where the data lives.

We're seeing tremendous investments in on-prem compute for GPUs. I mean, you can see Dell's results, but that's an indication of where the market's going. And I think that trend will accelerate because models are going to get larger and better. Um, and so they're going to become much more dependent on real-time data that needs to live right next door. >> I think the second thing that's going to happen is 20% of data that is accessible today, which is unstructured data, really becomes a first class citizen. and it starts dominating.

By that I mean we'll expand our definition of what a system of record is for a company from just a a data warehouse to this well-labeled highquality multimodal data set which is inclusive of text images video and audio and because these things are going to need multiple technologies to go through I think it requires a different management paradigm and a different technology. Maybe the third thing is the tighter integration between data and the orchestration around the data.

Meaning this is already happening in the AI world, right? First we built the models, then we built an agent framework, and now we're automating the agents because the agents are becoming autonomous. >> The same thing is going to happen in data. First, we built the data platforms, then we built the data pipelines, now we're going to be building an orchestration which will automate the data pipelines. uh not just becoming a a passive thing, but a more an active thing as when a reasoning model thinks I need answers, the pipeline should kick in. >> More of a thing you're monitoring than a thing you're doing. >> Exactly.

More more a thing that you're letting a model dictate what it needs at the time. And all your job to your point is making sure that the pipelines are good. They are, you know, they're not slowing down and they're secure. >> Yeah. And then agents to fix those. >> And then agents to fix those as well. Exactly. It's a it's an agent 8 agent world. >> It is. It is. Well, Versank, thank you so much for joining us today and sharing your insights on the the vital role data infrastructure plays in the age of AI. >> Oh, thanks for having me, Corey.

I loved it. >> Excellent. Well, as AI adoption accelerates across industries, one thing is becoming increasingly clear. Success is about more than building smarter models. It's about ensuring those models have access to the right data in the right place at the right time. For many organizations, solving the data challenge may ultimately determine how far their AI strategies can go. Thanks so much for listening to ES Speaks and we'll see you next time.

This transcript was generated automatically from the video's captions and may contain errors.

Sponsorisé par Dell AI Factory with NVIDIA Logo
The following is sponsored content. It may not reflect the views of our editorial staff.
May 29, 2026
17 minute read
eWeek Logo

eWeek has the latest technology news and analysis, buying guides, and product reviews for IT professionals and technology buyers. The site's focus is on innovative solutions and covering in-depth technical content. eWeek stays on the cutting edge of technology news and IT trends through interviews and expert analysis. Gain insight from top innovators and thought leaders in the fields of IT, business, enterprise software, startups, and more.

Propriété de TechnologyAdvice. © 2026 TechnologyAdvice. Tous droits réservés

Divulgation publicitaire : Certains des produits qui apparaissent sur ce site proviennent d'entreprises dont TechnologyAdvice reçoit une compensation. Cette compensation peut influencer la façon dont les produits apparaissent sur ce site, notamment l'ordre dans lequel ils apparaissent. TechnologyAdvice n'inclut pas toutes les entreprises ou tous les types de produits disponibles sur le marché.