Varun Chhabra, Dell & Kari Briski, NVIDIA | Dell Technologies World 2025
>> Welcome back, everyone, to theCUBE's live coverage here in Las Vegas at Dell Tech World 2025. I'm John Furrier. Dave Vellante, this next segment, we're going to talk about redefining the enterprise with AI and Dell AI Factories with NVIDIA.
We've got a great segment here. Varun Chhabra, SVP of ISG Marketing for Dell. Kari Briski, vice president of generative AI software for enterprise with NVIDIA. We have the stars of the show here on theCUBE.
>> This is the big news. >> This is the big news. I am so excited. I loved AI Factory. I've said it from Day 1. When you guys said
that, we're like, "I just love it. " Factories, we just love that. Thanks for coming on. >> Yeah. Thanks for having us. - Thanks for having us- >> All right. So it's the centerpiece of the keynote, the centerpiece of the theme.
It's driving all the innovation we're living in a time, Varun, that we've never seen before. Give us the update. This is, I heard, 2.0. What does that mean? Give us the quick news. >> It's hard to believe. It was just a year ago at GTC,
we announced the AI Factory with NVIDIA. We're already on 2.0. And in the middle, we've had, I think, 120 or 130 new releases since last year. With 2.0, there is a big focus on obviously refreshing the latest infrastructure based on all the innovation that NVIDIA is bringing in.
We also have a greater and expanded relationship with NVIDIA on the networking side where we're now... NVIDIA networking is now available to Dell for customers. Which means now, compute, storage, networking, they need support for their AI Factory with NVIDIA, they are able to get that through us. And then, a lot of work being done at the software layer and the data layer as well, which happy to go into more a little bit. Kari can touch upon that.
>> Yeah. Kari, share the relationship with NVIDIA and Dell. Very special relationship, integrated in. Give the update on your side.
>> Yeah. I think that our relationship has transformed from a traditional hardware vendor relationship to a deep strategic partner that's focused on offering a full stack, comprehensive solution that's going to deliver, and usher in the era of digital workers for the enterprise IT ops. >> And Jeff Clarke was in Taipei with the recent news there. Now, he's back here. Supersonic plane back.
What's the big learnings on the deployment experience? A lot of people hear generative AI and the enterprise and the hyperscalers are all using it. Where are we on the progress? We hear about stuffs in production, in POCs. Where is the AI Factory build out and deployment status progress? Can you guys scope where we're at? >> Yeah. We have, I think, over 3,000 customers
as we just said at the keynote. 3,000 customers in less than a year. They are at various stages of deployment, John.
A lot of people are in POCs. But as Michael talked about yesterday, a lot of overwhelming majority of those POCs are transitioning into production environments as well and what we find... It's always great to talk about the latest gear, the latest solutions, but the biggest learning we've had working with thousands of these customers is you've got to do a lot of work upfront.
As Jeff was talking about today on the keynote, you got to think through what your use cases are because if you don't sort through that and get alignment across all your stakeholders, there's going to be roadblocks down the line. Next step is data, getting your data strategy right and then fixing your processes before you start applying technology to it. And what we're doing with the AI Factory at NVIDIA really is trying to help customers with all parts of that journey. >> So the data center market has exploded a year ago.
It was just incredible. I mean, it was a perpetual $200 billion market that grew 50% in one year. >> Yep. - Mind boggling.
And at GTC, Jensen mapped AI for the hyperscalers, AI for enterprises, and then AI for robots. Robots are going to take a while and there's maybe different types of robots. Robots that do one thing, maybe those will kick up faster.
We know the hyperscalers are sucking up all the CapEx. Saudi is taking half a billion GPUs. Wow. What about enterprises? That's kind of the wheelhouse for Dell. I mean, you guys are partnering up.
What are they telling you? What are those 3, 000 customers telling you? What are their big challenges and how are you helping them get through them? >> Well, I think enterprises are moving from CPU- powered enterprise operations to GPU-enterprise operations. Why is that? It's because of LLMs and Agentic AI. And so, what are agents, right? They're autonomous digital workers that collaborate across disparate enterprise systems.
And so, when you're working with these enterprise systems, you need agents that can perceive and understand the tools that they have access to. They need to reason about a task. They need to access these systems, maybe do deep research, maybe do a chatbot.
And even when you're connecting to these knowledge, these storage servers that we're talking about, we're redefining the enterprise. It's not just about computers, it's about storage and networking as well. So these storage servers are going to have to be semantic storage servers. So serving up knowledge rather than data.
And what does that mean? You have to embed and find meaning in all these unstructured data to be able to serve it up. And so, that's where enterprises are right now. Serving up knowledge to these agents so that it can make manual tasks a lot easier.
>> And I think just to add on to what Kari said, that's the reason why you had Arthur today talking about the AI data platform. And when he was talking about our storage platforms, it's not just about a bigger, better, faster drive. Yes, of course, that's important. It's also about embedding semantic search capabilities for metadata search in the storage platform itself. Basically, making it faster than ever to basically be able to get information and insight that will help agents do more action.
>> So I had to leave in the keynote just as Arthur was talking about the semantic capabilities. I was like, "Ugh-" >> The good part- - ... because it's such an important part of the new stack real estate. So when you talk about semantic search, what exactly do we mean? I mean, we infer harmonizing the data so it can be served up to agents and governed and trusted.
You think about the Benioff and Nadella Wars, Clippy versus CRUD databases, agents talking to CRUD databases. There is some truth in both, but it's like what's real? What can you actually do today with the capabilities that you guys are bringing and how will that evolve? >> Well, I think today it is possible to do the semantic search. So basically, you find a meeting. You embed it in multimodal ways.
So you have to ingest the data, you have to look at the data, or you can even leave the data where it is, and you have pointers and access controls. But you're indexing, you're embedding it in a vector database. And so, when you're looking for that data, you understand that the agent will actually rewrite a query for you. So if you ask a question, it actually reforms the query. And so, semantically says, "This is the type of data I'm looking for," and it goes and finds it in multimodal ways and serves up the knowledge rather than just the data. Does that make sense? >> Yep. Mm-hmm. Okay- - So that's unstructured data,
but it's also for structured data- >> Yes. <inaudible>. Yeah. - ... agents are also able to write SQL, look at the schematics, look at the SQL, rewrite the SQL, say, "Is this what the user really wanted? " And go do it again. So it's reflecting on the answers
that it's getting after it writes the code and retrieves the information. >> Because that's the premise that we have. When we talk to enterprises, they say, "Well, we have all these SaaS systems.
We have all these transaction systems, the data, the business logic, the metadata, the operational and business, but they're all locked inside of those apps. We want to surface them, harmonize them so that we're all talking about the same thing. " And that's a big challenge for those guys. If I understand it correctly, your software stack, the integration, you're attacking that problem directly. It just takes some time to. >> Absolutely. I mean, one of the things Arthur talked about
today was the concept of a Dell data lakehouse, right? And that's a key part of the AI data platform. It's not enough to just have amazing storage platforms and be able to give performant access to data. To Kari's point, you actually need to be able to look at data across all your sources. So things like what we're doing in the Dell data lakehouse, whether it's structured data or unstructured data, whether it's sitting on Dell storage or out of the edge or in the cloud, what apps and developers want is access to all of that data in a seamless way.
And they don't want to have to build a lot of capability around semantic search. They want that inbuilt into the platform. They want federated access. They want the latest querying technology really built into the data platform, so that they can get all of that easily. >> I don't know if my mind is going to want us to jump down the rabbit hole and start talking speeds and feeds, KV cache, the operating system for the clusters. So I'm kind of tempted, but the agents need to run on the factory.
So one of the things we're tracking right now from a customer and tech perspective is most of the enterprise we talk to, they're doing a lot of stuff, this production stuff, but the real kind of stress is, "I don't want to under-provision. " They don't say that, but that's what they're basically saying, because if they come short on the token demand... So we're seeing tokens demand skyrocketing because reasoning is going to change the token equation. So what's happening now is RAG's out there.
We're seeing some use cases. Now, the big discussion is what do I do with the tokens? How do I figure out the scope of the factory? >> That's right. - Okay. So that's one problem. And the second problem is, "Okay. How do I get my innovation in the stack? " So there's kind of two threads there. Kari, how do you guys see this? Because agentic has to get kind of that requirements done.
You got to get to the data platform. You got to understand what you've got for data. You got to wire it up. But then once you start running it,
the tokens take over. What's your thoughts? >> Well, I have a couple of thoughts there because you mentioned RAG and I think RAG is just one use case that agents are utilizing right now. They could be... Again, I mentioned deep research. They're accessing different sort of tools.
They're writing code. They're doing RAG. So all these things are producing tokens in some way because they're reasoning about it. And even writing code to do a task is producing tokens, right? And so, we have things like dynamic resource allocation, intelligent request routing, because where is the data located? So you can go run that agent on the compute near where the data is for this really efficient. Because imagine, it's not just about the request time to you, it's the request time and the inter-agents talking to each other because agents talking to agents. And we know that for even at least one query, it's at least 50 LLM calls.
And so, if you imagine that 50 LLM calls amongst agents and how that's scaling, so you need to be able to do the dynamic resource allocation. So that's why we came up with products like Dynamo, NVIDIA Dynamo. >> I will say though, John. Yes, it is tempting to sometimes think you need massive, massive racks of...
Because that's what you see in the news today. But even for the use cases that we've rolled out at Dell, we've got thousands. We're one of the largest companies in the world.
You would be surprised at how you can start pretty small and scale pretty easily. It doesn't take that much . if you're not training models and doing inferencing at massive scale, you can actually get a lot done with a surprising small amount.
>> Which is important because not everybody can just rip and replace your air-cooled system- >> Yes. - ... plumb with liquid-cooled and there's- >> Yeah. And the latest... I mean, all the co-engineering work that Dell and NVIDIA are doing are putting more and more GPUs in a server options for air-cooling options, for liquid-cooling, of course, wherever customers are.
We've got systems to meet customer needs, whatever they are. So I think it's important to remember that. And of course, it will change with reasoning models and there will be more of a reliance on that. But in our opinion, and as Jeff was stressing today, you can't let analysis paralysis stop you from getting started- >> Yeah, yeah- - You've got to get started. You got to get started already- >> Speed wins. I love that term speed-
>> But you're providing a bridge. I think that's the key. It's not like, "Okay. You've got to go from Point A to Point B and make a huge leap and spend all this money.
You can iterate your way there." >> I mean, the thing... It's always interesting. When Jensen was talking to Michael Dell, I liked... There was a part where they were riffing about the PC days going on a throwback, kind of the OGs talking about the old days.
Glory days as they say, but not more glorious than now with AI as they got to. But the PC, I never heard anyone say that, "I'm done with the 286 processor. " So in that way, which is a smaller scale to AI, the market on the software side always rises up to the capabilities. So I'm kind of stuck on the tokens, because I think token demand will never stop for a long time.
So if that is true, the customers have to grok this and figure it out like, "Okay. " So what are you guys seeing with customers in the enterprise because they have the data to unlock your... I love that word. So this seems to be the top conversation. What's the strategy? What's the thinking? That's the number one question I get.
What's the size of the cluster? What's the size of the system? >> Well, I think since you've seen Jensen's keynotes before and he talks about the Pareto curve of what is your throughput versus latency and you can be great at throughput, you can be great at latency. So you have to find the Pareto curve of exactly what your use case is going to be really good at and then provision your system to that. And I love the idea of a factory because what does a factory do? It mines information, raw materials, and it makes a product and that product is tokens.
And so, tokens are output. And so, tokens equal your product and the money that you're making. So what is your output? And I think we were talking before we came on air about verticalization, specific domain. And so, how are you mining that data for your vertical, for healthcare, for retail, for finance? And how are you mining those tokens and then producing tokens that are of value to the applications that they're going into? >> That was one of the most powerful segments of the Computex talk, when Jensen walked around the stage talking about all the different sort of tools that were very specialized, whether it was for math or biosciences or...
There were dozens and I'm sure there are hundreds in the portfolio. >> Yeah. - Talking about just being able to scale and manage all of this, one of the things that we're also very excited about, we're entering a new phase of our relationship. Dell is now offering, working with NVIDIA, fully managed capabilities. So fully managed services for the AI Factory with NVIDIA.
So for customers, John, you're talking about, "How do I scale? How do I make sure we're getting the right SLA? How do I think about planning around token counts, et cetera? " For our customers that want basically Dell and NVIDIA to work on their behalf and manage that infrastructure for them, that capability is available as well now for the first time. >> Yeah. I think the services angle is really important because that is a frontend to the operating leverage of the factory and it's not like it's just to prime the pump. It's going to always be on. And then, maybe agents take over, which is why I love the data platform narrative you guys are putting together with the products. Because you get that right, a lot of good things happen.
I think you guys talk about the data flywheel with Nemo. So this gets back to the agents. I see agents connecting the models together, model management or whatever you call it. I don't know the word for it yet, but models need to have the intelligence from the data. What is this data flywheel concept with Nemo? How does that scale up? >> Yeah. I think if I were to put it in simpler terms,
we talk about how the IT department has to become the HR department of the digital workforce. And so, in order to do that, how are we making sure that these agents are improving over time? And so, you put them out in the wild and they turn data. And so, once they have output of data, that you learn from that data, you curate that data, you fine tune them, they get better. You evaluate these digital workers. You guardrail these digital workers.
And then, you're going to put them back out. The data flywheel is meant to do that. So as an agent is out working from those logs, from implicit and explicit feedback, from humans and logs, you're able to curate that data, customize, evaluate, guardrail, and put them back out.
The better you get, the better data you get back, the better feedback you get. And that is the flywheel to keep getting better and better. >> And the human input comes from the line of business, for example, right? >> Yeah. That's right. - And then, IT is responsible for implementing it. >> That's right. And remember, also instead of reinforcement
learning through human feedback, we call it reinforcement learning through expert feedback. Because you really want the experts giving the feedback at these systems, because it's really the knowledge workers are experts at their domain of these business systems. >> Yeah. And I love the mixture of experts. You brought up multimodal before. I want to come back before we leave the data platform, because it's not just language. Computer vision is going to be the biggest data ingest.
Okay? Certainly, digital twins. We love that concept as well. How does that factor in? Because there's a lot of diversity in the use cases. These factories have to support, Varun, all these things. >> We already have so many customers that are doing multimodal.
Sam Bird today talked about Northwest Medicine. They've got multimodal LLMs deployed and a perfect example where you have not just text for assisting healthcare workers. But also, how do you ingest... The example he brought up. How do you ingest x-rays and get diagnostics right away? It's absolutely part of the equation.
And I think what Kari is talking about in terms of the flywheel and the amount of data that is coming in that is not just text, that's going to be a lot more storage intensive. It basically, I think, means that we need a different primitive when we think about storage. And that's why the announcements that NVIDIA made at GTC with the AI data platform and the work we're doing with them on our data platform to enable that, I think, is going to be a game changer. Because you can't just take your old storage and apply that here. So all the innovation you saw today coming on stage and more and what we talked about at GTC together. Bringing together this turnkey AI data platform solution that allows customers to take advantage of this flywheel and the influx of that vast amount of data from these multimodal systems is going to be really, really important.
>> I think we should do a deep dive podcast for an hour on this with you guys, if you don't mind, because I think this is a really big, important conversation. I know you guys got to go, but I want to ask one final question. With all the buzz going on, Dave and I have been commenting, Kari, that everyone's at the speed of NVIDIA. So the speed of NVIDIA, whether it's the marketing, which is also the momentum...
I mean, we haven't gotten to robots yet, which is a tell- sign to what's coming with physical. So again, physical and digital merging. Certainly right now, digital twins is like the beginning phases of that. What's the speed of NVIDIA? If you could share your thoughts on moving at the speed of NVIDIA, because it's a real conversation we're having.
Okay. We're keeping up and- >> The speed of NVIDIA. Well, actually, Jensen has the philosophy called speed of light that we must all try to achieve and it's, what are the physical limits of how fast you could possibly go? And that's what we... What is the physical limits? So how do we work at that fast? And the physical limits of what's happening in the world today is evolving that fast and so we must keep up with the pace of innovation. And so, how are we providing the right compute, the storage and networking and infrastructure and software and orchestration and tooling for researchers, for data scientists, for computer scientists, for enterprise knowledge workers? So we want to make everyone's life's work more productive and that's Jensen's goal.
>> And you're also challenging conventional wisdoms that are out there. Previous things and making them better. >> I hope so. - I think that's one of the reasons why our partnership works so well is internally, we've seen how fast NVIDIA works.
And as all great partners do, they've pushed us and it's unbelievable. The energy, the pace at which we are working together, updating things. I mean, even between GTC and Dell Tech World, it's been a few weeks. Already, we have 20, 30 new announcements and updates on things we announced at GTC.
It's so invigorating getting up and coming to work every day because this is kind of what we're all working through and delivering joint success with customers for. >> Well, Varun, we've been watching you guys for 15 years. I can say you're moving super fast, moving the needle. Jeff Clarke just said on stage, "Fast wins. Speed wins. " I'll give you the final word. What's the easy button? How do you make it easy? It's not like a server back in the old day- >> No, it is- - You rack a server, put some cables in.
It's hard. How do you make it simple for customers? >> I think everybody wants to hear that the AI Factory is just one thing. I just put it in my data center and it works. The reality is that's not what happens. It's not a one size fits all. Our experience and 3,000 plus customer experience is the same, tells us the same.
But what really, really helps is having flexibility built into the solution and that's what we do, right? Working with NVIDIA. If somebody wants to deploy their AI Factory on their desk with their AI PCs, they can do that. If they want to build massive racks to train new models and provide GPUs as a service to companies, we can do that. And everything in between with the data, with the professional services. I mean, our goal really jointly is to help customers wherever they are and there really isn't...
We have 3,000 plus customers and 3,000 plus unique stories and unique entry points, but it's a platform that's really built to scale to be able to accommodate all of those needs and deliver value quickly. Well- >> And I think in order to transform enterprise IT, we have to also work with a broad ecosystem of partners. So we have to make sure that they are accelerated and they can run on this AI Factory as well.
>> Absolutely. That's why you see what we saw with Cohere, Glean, Red Hat, Google, et cetera. All built on top of the AI Factory with NVIDIA.
>> It's a great partnership. You guys are moving fast. Congratulations. The world needs it faster, smaller, and more intelligent- >> We'll get right on it- >> We're moving fast here in theCUBE. We're trying to keep up. We love covering it.
We love analyzing the NVIDIA Dell relationship. Again, the needle is moving, the game is changing, society is changing, the whole world is changing super fast. I'm John Furrier with Dave Vellante. Thanks for watching.
2025-05-27 02:08
