Dell Technologies World 2026 | Suresh Andani, AMD & Melissa Crichton, Dell Technologies

Clips
More from Dell Technologies World 2026

Suresh Andani

CVP, Compute & Enterprise AI

AMD

Melissa Crichton

VP, Server & AI Solutions

Dell Technologies

play_circle_outline Cost-Efficient Hybrid AI: Agent Placement, Model Sizing, and Token-Driven On‑Prem Inference

play_circle_outline AMD MI350P PCIe GPUs for Enterprise Inference: Solving ≤30 kW Rack Limits with Memory and Storage Pipelining

play_circle_outline Preference for starting small: one or two servers with validated scaling path

play_circle_outline Dell AI Factory Launches ISV Self-Certification, Deep Dell–AMD Engineering and ROCm/CUDA Abstraction for Enterprise Red Hat/VMware Stacks

Info
Transcript

Suresh Andani, AMD & Melissa Crichton, Dell Technologies

Suresh Andani

CVP, Compute & Enterprise AI AMD

Melissa Crichton

VP, Server & AI Solutions Dell Technologies

In this interview from Dell Technologies World 2026 in Las Vegas, Melissa Crichton, vice president of server and AI solutions at Dell Technologies, joins Suresh Andani, vice president of business development for enterprise AI at AMD, to talk with theCUBE's John Furrier and Dave Vellante about the shift from AI experimentation to agentic production and the hybrid infrastructure architectures enterprises need to get there. Crichton explains how Dell's AI factory strategy is built to let organizations start with one or two servers and scale incrementally — wheth... Read more

explore Keep Exploring

What are the salient aspects (or main drivers) of a hybrid AI architecture for enterprises? add

How does the AMD MI350P (PCIe) fit enterprise AI inferencing workloads and address typical data-center power, cooling, and memory/storage constraints compared with large-scale GPU systems? add

How is the AI factory designed to accommodate enterprises that want to start small and scale up their AI workloads, and what engineering, integration, and partner support does it include? add

How do you validate the hardware and software stack, and what engineering work do you and your partners perform to deliver a fully validated, enterprise-ready AI platform? add

bolt Powered by CUBE AI

Suresh Andani, AMD & Melissa Crichton, Dell Technologies

search

John Furrier

>> Welcome back around day two of Dell Tech Worlds. This is theCUBE's live coverage here in Las Vegas. I'm John Furrier. Dave Vellante. We're just plowing along, getting all the actions. Power-packed, full of product announcements, but building on the momentum of the AI factory as a gravity point for rearchitecting the infrastructure and resetting the enterprise and what large scale competing is all about. We got two great guests, Suresh Andani, VP of business development, Enterprise AI at AMD, Melissa Creighton, VP of server and AI solutions at Dell Technologies. Thanks for coming on.

Suresh Andani

>> Thank you.

Melissa Crichton

>> Thanks for having us.

John Furrier

>> We was talking even before camera came on. You guys are in the hottest area, the AI infrastructure's center stage servers keynoting everything. It's on stage the whole time. It's a set piece.

Melissa Crichton

>> Yes.

John Furrier

>> Congratulations.

Melissa Crichton

>> Thank you. Thank you. It's definitely a good time to be in the AI business. And certainly Dell has got a major opportunity to play in the space for sure.

John Furrier

>> Guys, talk about the innovation. We were right next to the AMD booth. I'm hearing some of the talks. Suresh, the architecture is the story this year, execution, because the ROIs on the table. People are seeing revenue come out of this, but they're seeing a new architecture around how to think about existing enterprise or new neo clouds are popping up for some specialty workflow. Talk about this architecture wave we're in.

Suresh Andani

>> Yeah, no absolutely. If you look at the enterprise, it's like more adopting the hybrid AI architecture. You probably heard Michael in his keynote yesterday talk about the hybrid AI since the ChatGPT moment happened in 2022, those chatbots, a lot of it was just like, "Hey, it's generative AI chatbots." Again, the last year, year and a half, as we are moving from a generative chatbot era to going more towards agentic AI, one size does not fit all. You basically want to go to your frontier models for certain class of workloads, but then the architecture is evolving into what you can also bring on-prem or run on-prem. There are reasons such as serenity, control, tokenomics, you'll hear that a lot. So the architecture is evolving into a hybrid AI architecture for enterprises to make sure the agent runtime is running at the right place. The AI/ML calls are happening at the right place, what you can go run locally, you run locally and privately, but you got to go into-

John Furrier

>> Yeah. And we just heard from the co-founder of OpenCloud. Everyone's using AI. So there's on prem, so therefore that's where the action is. It's not like you can mix and match.

Suresh Andani

>> You got to bring AI to the data many times than bringing data to AI.

Dave Vellante

>> So what are the salient aspects of that hybrid AI architecture? Can we unpack that a little bit?

Suresh Andani

>> Yeah. There are three main things that are driving it, right? So one, I think I mentioned this earlier. One is the token costs. Just when you were doing POCs and there were like 10 developers, 100 engineers using your AI, it was fine kind of just doing everything through the APIs, to the frontier models. But now as like within AMD, there are 20,000 engineers who are using Corex or CloudCore almost on a daily basis. And it's required pretty much I think once you go scale at that level, the token costs reach a threshold that you really start thinking about, how do I do this more cost effectively and more securely? So some of the aspects is like, can I run some of the planning aspects of the coding agent on prem? And then really I go to a Opus 4.7 or CloudCode with the APIs. That's kind of what-

Dave Vellante

>> Okay. So the physical location is part of that hybrid calculus. And then also presumably the model that you're running, you don't always need the big model, you might need a small model. And then also GPUs and CPUs.

Melissa Crichton

>> Right.

Dave Vellante

>> CPUs where you want to do some work, but please pick it up from there, Melissa.

Melissa Crichton

>> Yeah. I mean, I think Michael talked a lot about it. We also heard Arthur talk about it. Where your data is is also going to be where you're going to do the AI, right? And now roughly we think 80% of data is created at the edge. And so no enterprise is going to be able to run in one place. So that's why we believe we have to build out a hybrid platform for our customers, whether it runs at the edge, the core, or in a hyperscaler and being basically creating that platform. And so that's how we think about the AI factory really is creating that enterprise scalable model for our customers to run, start small, scale to the largest models that they may and that can run on prem or at the edge or in the cloud. And again, creating that orchestration and abstraction layer to move workloads where it's best suited.

John Furrier

>> I'd love to get you both to weigh in on this next question because it comes up a lot around, "Should I do this or that?" People are looking at their peers trying to figure out what the best use case is. I was talking to an engineer and a system architect the other day and they're like, "We optimize around the constraints." So constraints are obviously energy and money, put those aside for a second, but footprint matters. Not everyone can have a monster data center with all that power, but also the demand curve is so high that you're starting to see this distributed nodes of compute and horsepower. And the constraint around engineering, around memory comes in. So there's a lot of discussions around it. So it's very nerdy, but it has to do with getting the costs right, leveraging the footprint opportunities, but also using the right compute at the right workloads and small language models are back in the conversation again, Dave. So this is a discussion. It's out in the open.

Melissa Crichton

>> And I think the announcement that we're making with AMD fits right into that, right? So using the right tech for the right workload. So specific to the MI350P that we're announcing with AMD here in short order, these are going to be for inferencing workloads, enterprises getting started in AI. We're seeing a heavy opportunity within enterprise for PCI-based GPU workloads, right? When we started this journey a couple years ago, everything was on our 9680 with 8-way GPUs and rack scale systems, but we're recognizing in the enterprise space, you may need one or two servers to run the workload that you're looking to. And so we're super excited about what we're doing with you guys on the MI350P.

Suresh Andani

>> Yeah. If I could just add to what Melissa said, right?

John Furrier

>> Yeah.

Suresh Andani

>> Very critical point. About 70% of enterprise data centers are 30 kilowatt rack power density or lower and about 50% of them are like lower than like about 15 kilowatts. Plus a lot of them are not wired up to be liquid cool. So now if you are a traditional data center and you have ambitions to be AI-sophisticated enterprise, what do you do? Do you really need to go... That's where 350P comes in, where you can take your existing servers like a 7725 or a 7745 and plug in these 350P cards and still get 150, 170 billion parameter inference models run very efficiently on those. So we are very, very excited about our joint partnership with Dell on the 350P launch.

John Furrier

>> And just to put a finer point on that, is that the memory and storage equation on the large scale systems, you got to be highly engineered, very acute, pipelining that data into the GPUs or large scale systems with the power and cooling. I mean, that's a very expensive value proposition to sit there and be not utilized. Otherwise, it has nothing to do with the processor itself. It's the memory around. So memory constraints and footprint factor in.

Melissa Crichton

>> Yeah.

Dave Vellante

>> So Lisa Su, your CEO the other day, said something that caught my attention. She was saying that maybe a year ago the ratio of GPU to CPU was 8:1 in accelerated computing. It's now 4:1. She said, "We think it could go to 2:1."

Suresh Andani

>> It's 1:1 now.

Dave Vellante

>> Okay. So she underestimated that.

Melissa Crichton

>> Yeah.

Dave Vellante

>> Explain why, when is the CPU needed, when is the GPU needed? Share that with our audience.

Suresh Andani

>> Yeah, that's a great question. If you look at the agentic flow now, previously prior to the agentic AI, what was happening is there was a prompt that came in that was converted to tokens. The GPU did most of the work and it gave back an answer that was converted back from prompt into a response and that was it. It was just on step like CPU was just orchestrating the GPU. Now in the agentic flow where you're running like multi-system agents, the first step you do when an agent request comes in, you need to start planning. What is the whole plan for performing this agentic task? And that's a combination of CPU and GPU. Then you got to go execute to that plan, which involves a lot of like orchestration, which is a serial job. It's not like a parallel job that GPUs do well. So orchestrating the agentic runtime. And finally there's a lot of tool calling and execution. So whenever an agent request comes in, the CPU goes and talks to the database to get some data, it goes and talks to the web. So all of that tool execution is optimized on a serial architecture like a CPU versus a massively parallel architecture like a GPU. So what that does is now for every agentic task, you need more and more high thread count CPUs to parallelly perform all this orchestration and tool calling because if you don't do that, guess what? Your very expensive GPUs are sitting idle, and that is a waste of money. So CPUs now, there was 1:8, 1:4, now almost touching 1:1 ratio, but you'll need one CPU for one. And we believe in future it could be higher than 1:1.

Melissa Crichton

>> Yeah. The other way around.

John Furrier

>> I mean, Melissa, this is like old school product positioning. You got a flagship, high price performance, and then you got the entry level, but it's almost more sophisticated because the AI factories are being tooled or built and deployed based upon what you just said, which is use case.

Melissa Crichton

>> Right. Yeah.

John Furrier

>> So it's similar server kind of game.

Melissa Crichton

>> Yeah. Exactly. And I think, again, when we go back a couple years ago, we thought GPUs and large scale 8-way systems is what everybody needed, right? And as now we have agentic and we've seen all these different use cases, most enterprises want to start small and they want to start and they want to have the opportunity to scale up. And that's really how we've designed the AI factory, one or two servers with a couple of GPUs. And then as workloads scale, we can add on. And we've done all the certification, all the engineering work, we've integrated it with our AI data platform and all the storage and data protection requirements for our customers, networking. And then the last piece, which is super important, is that ecosystem of partners that we need to help our customers with to run and operate these AI factories.

Dave Vellante

>> That other part you're saying about validation is very important because customers don't want to be your field testers.

Melissa Crichton

>> Correct.

Dave Vellante

>> So yes, they want to start small, but they want to know what's validated and then when they scale, it's validated. So can you explain sort of how you guys do the validation, what's validated-

Melissa Crichton

>> Certainly.

Dave Vellante

>> ... and the engineering work that you guys do together in that work?

Melissa Crichton

>> Certainly. I can start and then I'll let Suresh. I mean, what we do is we've been working with AMD for years on the CPU side, certainly the last few years on the GPU side, deep engineering relationships. And we basically will create and build the runbook, right? We've done all the testing, certifications on the hardware. And then what I think is even additionally valuable for our customers is the software stack, right? Working closely with AMD and their software that I would love Suresh to highlight here, and helping it make it an easy button for our customers because to your point, they don't want to be messing with the bits. They want it to be working, running, and have time-to-first token as soon as possible. So that's really, we take all that work out and let you talk about the software maybe. Yeah.

Suresh Andani

>> Yeah. I think Melissa covered the validation piece. I'll just touch upon the software piece. The one main thing, what do enterprises really care about, right? TCU, obviously we all understand that, but they want solutions. They want something that on day zero when they receive the gear from Dell powered by AMD, they're able to kind of start running agents the same day. So what we are doing is we are partnering with all of the key OSVs with Red Hat, OpenShift AI, with Nutanix, with VMware. We're building a full enterprise AI stack. The vision over there and the goal is to really make the ROCm or CUDA level invisible to the enterprises. They don't want to program or code at that level. Now there are some unique use cases that you got to get to that level, but the mainstream enterprise use case where you are building agents and using GPUs to serve the models needs to be abstracted at a level that they are just pretty much use that as a fully validated platform and they start building their use cases on top of that. So we work closely with Dell, but not only Dell, with all the ISVs like Red Hat, like Nutanix and VMware to get to that point.

John Furrier

>> Well, Suresh and Melissa, it's fast and furious, and that relationship probably comes in handy with the speed game. My final question is, what's next? What do you guys focus on? Because we appreciate what you do, but we want more of it. Everyone. The demand is so high, the requirements are changing, you start to see more distributed, disaggregated infrastructure. What's your vision? What are you guys doing? What's the focus on the relationship?

Melissa Crichton

>> Yeah. I mean, I'll start. I think you're going to see us just as Dell has been iterating on the AI factory for the last couple years, we're going to continue to iterate, right? What was new news two to three weeks ago is now old news, right? And so we got to continue to be innovative and help take a lot of this out of the customer's way. So trust that we're going to keep doing that in partnership with AMD and that entire open ecosystem so we can give our customers choice. So that's what we'll be working on.

John Furrier

>> Suresh, close us out. What's next? What's the focus?

Suresh Andani

>> Yeah. From a partnership standpoint, as Melissa said, we're going multidimensional. We don't have just spot products, 350P, 355, but we are working on a full roadmap that goes three years out at least. That's one. The second is we are working closely with Melissa on all the sales and go-to market. We are working closely with, you mentioned, Gil earlier on all the solutions and AI factory. We're working with Arthur's team on all the ISG stuff with Arun and team. So this is like bringing fully validated baked AI solutions, not products, not platforms, but solutions to our enterprise customers starting from infrastructure to AI data factory to ISVs and third party solutions.

John Furrier

>> Yeah. And was under reported, in my opinion, is the partner impact because more than ever the customers aren't really in a great position to build all this stuff fast. So you got the competencies on the partners. It's more, I won't say turnkey, but that's the only word I can think of, but more modular.

Melissa Crichton

>> Right.

John Furrier

>> Scalable. They come in small, build out. I mean, it's a dream scenario. Partners must be excited.

Melissa Crichton

>> Yeah, they are. And we just also announced here this week, a new ISV certification program, right? So allowing our ISV partners to self-certify on the AI factory. And then that'll give our customers more choice, but know that those solutions have been certified on our portfolio and hopefully taking, again, more of the decision making out of their way. And we've done a lot of the work.

John Furrier

>> Well, congratulations. We thank what you do. Appreciate it.

Melissa Crichton

>> Yeah.

John Furrier

>> I'm John Furrier, Dave Vellante. Day two, fast and furious. Michael Dell's coming into the queue. We can see him walking up now. He's got his name on all the machines that need to sign the boxes.

Dave Vellante

>> Don't need to sign. Their name's already on it.

John Furrier

>> the new factory, so we'll be right back with Michael Dell after the short break.

Dave Vellante

>> That was great.