Dell Technologies World 2026 | Robert Hormuth, AMD & David Schmidt, Dell Technologies

Clips
More from Dell Technologies World 2026

Robert Hormuth

Corporate Vice President, Architecture and Strategy, Data Center Solutions Group

AMD

David Schmidt

Vice President, PowerEdge Product Management

Dell Technologies

play_circle_outline Agentic AI Puts CPUs Back in Charge: Orchestrating Threads and GPUs for Complex Planning

play_circle_outline Inference GPU:CPU ratio shrinking; estimates moving from 8:1 to near 1:1

play_circle_outline Air Cooling Dominates Enterprise Servers: How 15–20 kW Rack Limits Shape Density and Economics

play_circle_outline AMD offers air-cooled EPYC, Instinct PCIe cards, Helios DLC for larger racks

play_circle_outline Server consolidation (7:1 to 13:1) frees rack space and power for AI infrastructure

Info

Robert Hormuth, AMD & David Schmidt, Dell Technologies

Robert Hormuth

Corporate Vice President, Architecture and Strategy, Data Center Solutions Group AMD

David Schmidt

Vice President, PowerEdge Product Management Dell Technologies

search

Dave Vellante

>> Hi, everybody. We're back at Dell Tech World 2026. This is wrapping up day two. This has been quite a day. John Furrier has had 10 interviews in a row. I took a quick break to go at a round table with Jeff Clarke. Gemma Allen is with me. I'm Dave Vellante, and we're excited to have Robert Hormuth here. He's the corporate vice president for Silicon Design and Engineering at AMD. David Schmidt is the vice president of PowerEdge Product Management at Dell. Folks, good to see you. Thanks for coming on.

David Schmidt

>> Great to be here. Thanks for having us.

Robert Hormuth

>> Great to here. Nice. Yeah.

Gemma Allen

>> Welcome on.

Dave Vellante

>> All right. Why are CPUs all of a sudden so important? We've been talking about GPUs for three years. GPUs are taking over the world, but now all I hear is CPUs. They're outnumbering GPUs potentially at some point. Not quite yet, but what's going on? Explain.

Robert Hormuth

>> Well, I think, one, I don't know that CPUs ever weren't important first. They were always needed in traditional IT, traditional enterprise cloud, IaaS, PaaS, SaaS. That's what fueled the internet, right? CPUs. AI came along, and they were needed to manage the GPUs and training, primarily foundational model training. GPUs alone can't do a whole lot. They need a CPU to orchestrate them, to load them, to instruct them what to do. And there was a high growth of CPU demand for training and inference on the LLMs. And then along comes agentic AI, and inference starts to really take off. Agentic AI is a different kind of workload, where if you think about LLMs, it's, I interact with the computer, I ask it a question, it goes off, gives me an answer, and I might iterate. Agentic AI is very goal-oriented. You ask it to achieve a goal, and it's going to use every tool, every software, everything in the book to go achieve that goal. So in the world of agentic AI, it may kick off hundreds of different, thousands of different software stacks or threads to go achieve that goal. One of those might be querying a GPU to get a complex math equation, but the rest of it is planning, checking, kicking off tools, verification, iterating. So there's just a lot of work to do around the complex math that the foundational model might return. There's just a lot more going on in the world of agentic. That's one of the big kickers right now, I would say, for CPU demand.

Dave Vellante

>> Of course, inference is sort of all in vogue now, and it's a matrix math, and a lot of it. Lisa Su, I think, was on your last earnings call, or maybe it was approximate to that, basically said, "Hey, we were at eight to one, then it went to four to one. We see it heading to two to one." People now saying it's even sort of one to one. This is the ratio of GPU to CPU. So that's cool. We were just, as I said, in a round table with Jeff Clarke, Robert, and he was explaining some of what you guys are doing internally, and basically framing it as, "Hey, we're a big enterprise, but we're enterprise. And much, if not most, of the stuff we're doing inside the enterprise is air cooled." You don't necessarily need the latest and greatest water-cooled GPUs melting down the data center. We can get a lot done. So that is interesting. I think a lot of people don't necessarily understand that. If you're training, you want those large training clusters, obviously they've got to be liquid-cooled, but air-cooled is still the big volume driver in the enterprise.

David Schmidt

>> Absolutely. I mean, as we looked at our next generation, the PowerEdge servers we launched here this week, our customers were loud and clear. Many, many enterprise customers weren't prepared for liquid cooling within their data center. And then really what drives liquid cooling, it's the amount of power that you can get to a single rack. So it's probably better to talk just about rack fundamentals. You're in a rack environment where maybe you have power constraints along the lines of 15 kilowatts a rack, maybe 20 kilowatts per rack. We get that feedback quite a bit. So it was really important to us to have a wide range of air-cooled systems. And then it's a simple economic conversation. Robert went into the technical reasons that you need more CPUs to drive these agentic workflows. You can do it on a GPU, but it doesn't make sense. You can do that on a stack of just core compute CPU-based PowerEdge servers.

Gemma Allen

>> Well, bring this to life for us a little bit. Talk to us about the real-life use cases, examples, realities of where CPUs are having most effect at this moment. What's new? What's changing?

David Schmidt

>> The way I view it, there's two distinct workloads to discuss. There are the agents, and Robert talked about kicking off a process. You maybe want it to crunch a large amount of data you pulled off the internet, or maybe you have some internal documents, and you have some of that data structured, some of that data's unstructured. Well, the first thing you have to do in an artificial intelligence environment is structure that data in some rational way. So that's that data transformation part. That's going to run on a series of CPU-driven tasks or operations. And then you might be asking all of these great tools that we read about. I could name them, but we would be here all day. There's a lot of really awesome tools that can string together a series of events. It can go out and search the web and crawl it and help you research the next vehicle you want to buy for your child, which is something I did here very recently. And that's going to kick off a lot of just different scripts. It's going to open browsers. Now, if you wanted to shop for a car every day, that means you need to run it on something besides your notebook. You need to host that on a virtual machine, on a server somewhere, so that it can be constantly running as something you would call an agent. So that is the kind of real-world example just on a personal basis. But now, let's transform it into the enterprise. The other example would be, now you have these agents working, they're generating data. They're going to have to store that data on your traditional, they're going to interact with your traditional application stacks in your environment. If you have 10,000 more agents, you have 10,000 more users driving traffic to your internal IT-based workloads that you've been running for years, maybe decades, all of a sudden you're going to see increased traffic on that core application stack inside your data center.

Dave Vellante

>> Okay. So Epic and PowerEdge are like peanut butter and gelatine.

Robert Hormuth

>> PB and carrots. Yeah, absolutely.

Dave Vellante

>> Take us through where you're at in the roadmap, what's state of the art, what can you tell us about where you're headed?

Robert Hormuth

>> Yeah. So I'll start a little bit and then let David talk about the Dell portfolio. But from an AMD perspective, the path that we've been on in growing the portfolio has been very pragmatic. It's about framing the market portfolio and the swim lanes that we want to go after, and then planning a roadmap that we can execute on. Because the first thing we have to do with our customers, they have to trust us. They have to trust us that we can execute on what we plan. The last thing we want to do is put a bunch of things on the roadmap we can't execute to. That's a quick way for David to get mad at us. So we spent a lot of time with David and the PowerEdge team on what should we be doing for what swim lanes. I'll just take like the air cooling. David and team shared a lot of their insights that they had from their enterprise customers, and that helped shape our roadmap. So just talking about air cooling for a minute, Epic is entirely air coolable. Everything in the CPU Epic domain is air coolable. If you drive enough density, you might need DLC, but that's a density decision you as an end user made. And then on the instinct side, we started air coolable. As we got into the foundational training models with the bigger GPUs and the group of eight, we added DLC, but maintained air. And then recently, working with Dell, we added a PCI chem card to our air coolable roadmap for GPU instinct, so that it fit in more Dell PowerEdge servers that go into the broader enterprise, because as David said, not everybody is adopting DLC, nor can they, and not everybody needs 72 or 8 or 128 GPUs. They might need a little bit of AI in an application and a powered server. So our goal is to offer that end-to-end spectrum. At AMD, from a roadmap wise now, we've got our little rise in AI Pro that's a notebook processor that runs ROCm. It runs 20-billion parameter models just fine. We've got air-cooled PCI chem, we've got air cooled 8GPU, and then we get into the bigger rack scale with our Helios, which is DLC. And that is our plan in the future, is maintaining both swim lanes of air and liquid for the different swim lanes of the market that the customers are requiring.

David Schmidt

>> Yeah, we're loud and clear. Robert's right. We have a really great partnership. We give feedback directly. I think AMD's designed a great next-generation portfolio. We're going to spend the entire year talking about that. You're going to have a lot of great options in all the ways that Robert described.

Gemma Allen

>> I had a great visit to the AMD booth earlier. I met your colleague and your ex-colleague, Mike Darby.

David Schmidt

>> Okay.

Gemma Allen

>> He talked me through some of the tech, some of those specs. We joked there's everything there from what looks like a John Deere engine to a scooter, right? It is quite the scale of tech happening at the AMD booth. But one thing he spoke about was deployment, actual execution, realized value, fast, effective turnaround times. Talk me through what you guys are both seeing and I guess how important a necessity that is in the decision-making process.

David Schmidt

>> Can I take that one?

Robert Hormuth

>> Yeah, go first. Yeah.

David Schmidt

>> It's like near and dear to my heart because we talked about it on stage this morning. Arthur talked about it. The deployment and standing up these environments, whether it's the largest of the large or something that fits in these air-cooled environments, it's something we've excelled at for years, generations. Delivering what our customers needed on time, fit to the spec of their data center, but we did it one chassis at a time, one platform at a time. Now we're at a rack scale level. We have the power rack infrastructure. We have it for compute, we have it for networking, we have it for storage. Delivering that outcome, thinking about the product as the rack is so critical to that initial deployment into the customer's environment. No two environments so far have proven to be the same. So that's the beginning of it. And then you come along with the software stack and all the other things you have to do to get the applications up and running.

Robert Hormuth

>> Well, let me just add to that. So from an AMD standpoint, first, we work closely to define the products. One of our fundamental goals is to aid in that deployability to make AMD technology adoption easy. And that's why we really embrace the open source mindset with ROCm and all the tools that we put out there. Why we've put so much investment with some of the framework vendors so that we're out of the box, day zero support. We're day zero, new model comes out, we're day zero optimized. Hugging Face runs a CI/CD stream every night, so they're rebuilding every night on the latest instinct and ROCm, so that we have a software stack that they can deploy with confidence to help their customers get up and running. Because at the end of the day, customers have to see value out of the technology and the services and the product that we deliver, or they're going to go somewhere else, and that's what we're trying to solve.

Dave Vellante

>> David, how should we think about segmenting the market? Obviously you have the training clusters that are going on. Okay, that's cool. You get the neoclouds rising up, everybody's spending money, like crazy CapEx. And then if I go back two years, a lot of the experimentation from the enterprises was in the cloud, a lot of POCs. Now, you guys got what? 5,000 AI factory customers? Okay, great. What are you seeing them doing? Where are they on their journey? They're obviously not replumbing for liquid cooling, most of them. So what are they actually doing?

David Schmidt

>> So what's resonated the most is, where we can have a conversation with our customers, show them our next-generation portfolio and really present a spectrum of choices for the environment they have. And that could be anywhere from five kilowatts to 50 kilowatts and up. By that, I mean the amount of power in a single rack. So we've found a lot of value having the conversation that way to say, "Where are you today? Where are you with the rack power that you have? And then let's talk about the type of workload, whether you want compute density, storage density, air cooling, or maybe you are ready for liquid." We have those ranges in the portfolio.

Dave Vellante

>> You hear so much about changes in networking and storage and IO. In the PowerEdge world, what are the requirements that are unique to that world as it relates to outside the server complex, outside the rack and the storage and the networking?

David Schmidt

>> Yeah. So within the server, you see Robert and team give us a great amount of... They give us all these great lanes, all these IO lanes. They give us all these cores, and we get to design a system to take advantage of all of that. So we can put a lot of storage in a platform. We can drive nine petabytes of storage in a single AMD-based platform. How are you going to get that data on and off the box? So now you have an IO component in the back of the system. Now you're trying to get something like 800-gigabit speeds so that you have the right balance of IO, the right balance of storage upfront, and the right balance of cores relative to the memory. Then you think about all those components in a platform, and it circles all the way back around to the thermal power conversation we were just having of, do you want to try to make that as dense as you can? If so, you're probably in a liquid-cooled environment. Or we're going to build chassis that are so well performing from a thermal perspective. We're going to optimize the airflow that you've got in your data center. We're going to optimize your cooling capacity. You're going to see that in all the next-generation designs that come out of us for PowerEdge.

Gemma Allen

>> So when we think about the year out, the next year we're back here, 2027, Dell Tech World again.

David Schmidt

>> Yes, we will.

Gemma Allen

>> What are the priorities for this relationship? It seems though there is ample opportunity here. There is a market for the taking, right? What are you guys really honing in on? Are there long-tail opportunities like an addressable market you guys haven't yet covered together? We talk about sovereign AI a lot. Where are you guys really focusing?

Robert Hormuth

>> I mean, I'll go first, and I would say that I think what we're really aiming for is ensuring that our customers are successful, first and foremost. I think the things that are going on, we started the discussion about why do cores matter more with agentic AI, and I think we kind of missed one of the other tailwinds that's been going on, especially with Epic and the Dell portfolio, we've been driving a consolidation story with a lot of our customers where they can do seven-to-one consolidation of old legacy servers to a modern, high core count, air coolable server. And they're able to free up space, capacity, power to make room for AI. So I think it's going to be actually kind of a duopoly of things that are going on, is customers are going to have to free up space and power to afford the space and power for AI. So we're going to keep driving the consolidation pretty hard with our customers, I think, so that they can make room to be successful in AI. Because the era of AI is not waiting. This is not one of these market trends that you can set back and debate for six months or nine months to enter, because your competition is going to go fast. Just all the disruptions that we're seeing at AMD on our own design side, whether it be silicon design, validation, software design, validation, it's not about doing the same with less, it's about doing more projects. We're able to take the resources we have and do more than we were. So I think that's going to be the mantra in most of the other companies, is how do I make room, space, power, and then invest in AI so I can outrun my competition?

Dave Vellante

>> Doing so, it's important within the same power envelope, right?

Robert Hormuth

>> Yeah.

Dave Vellante

>> So every, I don't know, whatever, 12, 18 months, you're on a steep improvement performance curve. I can now drop that into your point. Same infrastructure, I can, I don't know, increase my work by 30%, 40%, 50%, or sometimes more.

Robert Hormuth

>> Yeah. I don't know if David has a rule of thumb here. I'll ask the question for David. So how many servers do we consolidate down to make room and space for one AI server?

Dave Vellante

>> Yeah. If you're seven to one, let's say. I mean, I don't know, is that-

Robert Hormuth

>> That might be one area to cool a server.

David Schmidt

>> You'll see a 13-to-1 consolidation ratio in our latest generation. That's 13 just core rack servers down to one. So Robert's point, it's going to free up resources, power resources, rack space, et cetera, that you can then deploy for these agentic workloads that we've been talking about. To your question about 2027, I'm super excited about helping these customers manage that journey of, what do I need to do in my data center so I can meet all the demands and all the opportunities that Robert just highlighted? How do they address it? How do they redesign? How do we help them fit within our portfolio? And then, of course, look beyond that, I think the rest of this decade is going to be really exciting from a core compute perspective. So we should be back here next year talking about probably things, completely new things that AI has uncovered for us to go solve.

Dave Vellante

>> Last question is, what's the hard engineering problem your customers are asking? Your customer's right here. So the hard engineering problem that you're working on that if you could solve it's going to really drive customer value. Start with the silicon and then the end customer, if we could.

Robert Hormuth

>> I have to bring it back to the customers. There's a lot of hard engineering and silicon level that we have to go tackle, but tackling that without having a value proposition that helps a business go further, faster, add new business, create new revenue. AI to me is not about just keeping the lights on. It's not about... Traditional IT is running the business. AI is generating new revenue streams, new business opportunities, new customers. So that to me is the problem that we have to go tackle, is make sure that our customers are successful. Yeah, there's a lot of hard engineering on silicon packaging process nodes, but none of that matters if we don't solve a problem for the customer at the end of the day.

Dave Vellante

>> David, that's good framing to lead up to the question. You got to run the business, to grow the business, and to transform the business. Historically, the sort of on-prem business is, run the business, refresh cycle. Now you're driving new innovation. What's the hard problem customers are asking you to solve?

David Schmidt

>> Can I say all of it? They're asking us to help them solve all of it, but if I had to pick one right now, the storage aspect that rides alongside all of this great compute. We could go on another hour, we're not going to do that, but we could go another hour of just the amount of storage that gets dragged alongside this. They're asking us to solve and build data platforms that I know we've talked a lot about it here this week, but they're asking us to solve those data challenges and deliver them exabytes of capacity as soon as possible. To do that, you have to have a well-thought-out storage platform. We have excellent... We talked about exit scale at a rack scale level. We have great things with PowerStore. We talked about lightning and the record-breaking speeds that we provide there. All of that is at our disposal. We have all these tools in our toolbox to help solve those problems, whether they're giant rack scale challenges that we have to help our customers solve or transform my enterprise using the data center space that I have today.

Dave Vellante

>> So you got the infrastructure, you're ready there. It's that data problem that's the hard one. Varun Chhabra called it the hidden blocker or the hidden barrier. So, excellent guys, thanks so much for coming to the community.

David Schmidt

>> Thanks for having us.

Robert Hormuth

>> Yeah. Thanks for having us. Thank you so much.

Dave Vellante

>> Really appreciate your time and the update. Best of luck to you both.

David Schmidt

>> Thanks.

Robert Hormuth

>> All right.

Dave Vellante

>> All right, this is Dave Vellante for Gemma Allen, wrapped up day two here at theCUBE's coverage of Dell Tech World 2026. John Furrier will be back here tomorrow with myself and Gemma. Go to SiliconANGLE.com. We got all the news on DTW, and of course, around the enterprise tech industry. theCUBE.net is where you watch this. Check out theCUBEresearch.com. That's where we do all the deep dives. Go to the CUBE AI, ask it a question, find out what the top takeaways are from Dell Tech World 2026, and tune in tomorrow. Thanks for watching.

Robert Hormuth, AMD & David Schmidt, Dell Technologies

search

Dave Vellante

>> Hi, everybody. We're back at Dell Tech World 2026. This is wrapping up day two. This has been quite a day. John Furrier has had 10 interviews in a row. I took a quick break to go at a round table with Jeff Clarke. Gemma Allen is with me. I'm Dave Vellante, and we're excited to have Robert Hormuth here. He's the corporate vice president for Silicon Design and Engineering at AMD. David Schmidt is the vice president of PowerEdge Product Management at Dell. Folks, good to see you. Thanks for coming on.

David Schmidt

>> Great to be here. Thanks for having us.

Robert Hormuth

>> Great to here. Nice. Yeah.

Gemma Allen

>> Welcome on.

Dave Vellante

>> All right. Why are CPUs all of a sudden so important? We've been talking about GPUs for three years. GPUs are taking over the world, but now all I hear is CPUs. They're outnumbering GPUs potentially at some point. Not quite yet, but what's going on? Explain.

Robert Hormuth

>> Well, I think, one, I don't know that CPUs ever weren't important first. They were always needed in traditional IT, traditional enterprise cloud, IaaS, PaaS, SaaS. That's what fueled the internet, right? CPUs. AI came along, and they were needed to manage the GPUs and training, primarily foundational model training. GPUs alone can't do a whole lot. They need a CPU to orchestrate them, to load them, to instruct them what to do. And there was a high growth of CPU demand for training and inference on the LLMs. And then along comes agentic AI, and inference starts to really take off. Agentic AI is a different kind of workload, where if you think about LLMs, it's, I interact with the computer, I ask it a question, it goes off, gives me an answer, and I might iterate. Agentic AI is very goal-oriented. You ask it to achieve a goal, and it's going to use every tool, every software, everything in the book to go achieve that goal. So in the world of agentic AI, it may kick off hundreds of different, thousands of different software stacks or threads to go achieve that goal. One of those might be querying a GPU to get a complex math equation, but the rest of it is planning, checking, kicking off tools, verification, iterating. So there's just a lot of work to do around the complex math that the foundational model might return. There's just a lot more going on in the world of agentic. That's one of the big kickers right now, I would say, for CPU demand.

Dave Vellante

>> Of course, inference is sort of all in vogue now, and it's a matrix math, and a lot of it. Lisa Su, I think, was on your last earnings call, or maybe it was approximate to that, basically said, "Hey, we were at eight to one, then it went to four to one. We see it heading to two to one." People now saying it's even sort of one to one. This is the ratio of GPU to CPU. So that's cool. We were just, as I said, in a round table with Jeff Clarke, Robert, and he was explaining some of what you guys are doing internally, and basically framing it as, "Hey, we're a big enterprise, but we're enterprise. And much, if not most, of the stuff we're doing inside the enterprise is air cooled." You don't necessarily need the latest and greatest water-cooled GPUs melting down the data center. We can get a lot done. So that is interesting. I think a lot of people don't necessarily understand that. If you're training, you want those large training clusters, obviously they've got to be liquid-cooled, but air-cooled is still the big volume driver in the enterprise.

David Schmidt

>> Absolutely. I mean, as we looked at our next generation, the PowerEdge servers we launched here this week, our customers were loud and clear. Many, many enterprise customers weren't prepared for liquid cooling within their data center. And then really what drives liquid cooling, it's the amount of power that you can get to a single rack. So it's probably better to talk just about rack fundamentals. You're in a rack environment where maybe you have power constraints along the lines of 15 kilowatts a rack, maybe 20 kilowatts per rack. We get that feedback quite a bit. So it was really important to us to have a wide range of air-cooled systems. And then it's a simple economic conversation. Robert went into the technical reasons that you need more CPUs to drive these agentic workflows. You can do it on a GPU, but it doesn't make sense. You can do that on a stack of just core compute CPU-based PowerEdge servers.

Gemma Allen

>> Well, bring this to life for us a little bit. Talk to us about the real-life use cases, examples, realities of where CPUs are having most effect at this moment. What's new? What's changing?

David Schmidt

>> The way I view it, there's two distinct workloads to discuss. There are the agents, and Robert talked about kicking off a process. You maybe want it to crunch a large amount of data you pulled off the internet, or maybe you have some internal documents, and you have some of that data structured, some of that data's unstructured. Well, the first thing you have to do in an artificial intelligence environment is structure that data in some rational way. So that's that data transformation part. That's going to run on a series of CPU-driven tasks or operations. And then you might be asking all of these great tools that we read about. I could name them, but we would be here all day. There's a lot of really awesome tools that can string together a series of events. It can go out and search the web and crawl it and help you research the next vehicle you want to buy for your child, which is something I did here very recently. And that's going to kick off a lot of just different scripts. It's going to open browsers. Now, if you wanted to shop for a car every day, that means you need to run it on something besides your notebook. You need to host that on a virtual machine, on a server somewhere, so that it can be constantly running as something you would call an agent. So that is the kind of real-world example just on a personal basis. But now, let's transform it into the enterprise. The other example would be, now you have these agents working, they're generating data. They're going to have to store that data on your traditional, they're going to interact with your traditional application stacks in your environment. If you have 10,000 more agents, you have 10,000 more users driving traffic to your internal IT-based workloads that you've been running for years, maybe decades, all of a sudden you're going to see increased traffic on that core application stack inside your data center.

Dave Vellante

>> Okay. So Epic and PowerEdge are like peanut butter and gelatine.

Robert Hormuth

>> PB and carrots. Yeah, absolutely.

Dave Vellante

>> Take us through where you're at in the roadmap, what's state of the art, what can you tell us about where you're headed?

Robert Hormuth

>> Yeah. So I'll start a little bit and then let David talk about the Dell portfolio. But from an AMD perspective, the path that we've been on in growing the portfolio has been very pragmatic. It's about framing the market portfolio and the swim lanes that we want to go after, and then planning a roadmap that we can execute on. Because the first thing we have to do with our customers, they have to trust us. They have to trust us that we can execute on what we plan. The last thing we want to do is put a bunch of things on the roadmap we can't execute to. That's a quick way for David to get mad at us. So we spent a lot of time with David and the PowerEdge team on what should we be doing for what swim lanes. I'll just take like the air cooling. David and team shared a lot of their insights that they had from their enterprise customers, and that helped shape our roadmap. So just talking about air cooling for a minute, Epic is entirely air coolable. Everything in the CPU Epic domain is air coolable. If you drive enough density, you might need DLC, but that's a density decision you as an end user made. And then on the instinct side, we started air coolable. As we got into the foundational training models with the bigger GPUs and the group of eight, we added DLC, but maintained air. And then recently, working with Dell, we added a PCI chem card to our air coolable roadmap for GPU instinct, so that it fit in more Dell PowerEdge servers that go into the broader enterprise, because as David said, not everybody is adopting DLC, nor can they, and not everybody needs 72 or 8 or 128 GPUs. They might need a little bit of AI in an application and a powered server. So our goal is to offer that end-to-end spectrum. At AMD, from a roadmap wise now, we've got our little rise in AI Pro that's a notebook processor that runs ROCm. It runs 20-billion parameter models just fine. We've got air-cooled PCI chem, we've got air cooled 8GPU, and then we get into the bigger rack scale with our Helios, which is DLC. And that is our plan in the future, is maintaining both swim lanes of air and liquid for the different swim lanes of the market that the customers are requiring.

David Schmidt

>> Yeah, we're loud and clear. Robert's right. We have a really great partnership. We give feedback directly. I think AMD's designed a great next-generation portfolio. We're going to spend the entire year talking about that. You're going to have a lot of great options in all the ways that Robert described.

Gemma Allen

>> I had a great visit to the AMD booth earlier. I met your colleague and your ex-colleague, Mike Darby.

David Schmidt

>> Okay.

Gemma Allen

>> He talked me through some of the tech, some of those specs. We joked there's everything there from what looks like a John Deere engine to a scooter, right? It is quite the scale of tech happening at the AMD booth. But one thing he spoke about was deployment, actual execution, realized value, fast, effective turnaround times. Talk me through what you guys are both seeing and I guess how important a necessity that is in the decision-making process.

David Schmidt

>> Can I take that one?

Robert Hormuth

>> Yeah, go first. Yeah.

David Schmidt

>> It's like near and dear to my heart because we talked about it on stage this morning. Arthur talked about it. The deployment and standing up these environments, whether it's the largest of the large or something that fits in these air-cooled environments, it's something we've excelled at for years, generations. Delivering what our customers needed on time, fit to the spec of their data center, but we did it one chassis at a time, one platform at a time. Now we're at a rack scale level. We have the power rack infrastructure. We have it for compute, we have it for networking, we have it for storage. Delivering that outcome, thinking about the product as the rack is so critical to that initial deployment into the customer's environment. No two environments so far have proven to be the same. So that's the beginning of it. And then you come along with the software stack and all the other things you have to do to get the applications up and running.

Robert Hormuth

>> Well, let me just add to that. So from an AMD standpoint, first, we work closely to define the products. One of our fundamental goals is to aid in that deployability to make AMD technology adoption easy. And that's why we really embrace the open source mindset with ROCm and all the tools that we put out there. Why we've put so much investment with some of the framework vendors so that we're out of the box, day zero support. We're day zero, new model comes out, we're day zero optimized. Hugging Face runs a CI/CD stream every night, so they're rebuilding every night on the latest instinct and ROCm, so that we have a software stack that they can deploy with confidence to help their customers get up and running. Because at the end of the day, customers have to see value out of the technology and the services and the product that we deliver, or they're going to go somewhere else, and that's what we're trying to solve.

Dave Vellante

>> David, how should we think about segmenting the market? Obviously you have the training clusters that are going on. Okay, that's cool. You get the neoclouds rising up, everybody's spending money, like crazy CapEx. And then if I go back two years, a lot of the experimentation from the enterprises was in the cloud, a lot of POCs. Now, you guys got what? 5,000 AI factory customers? Okay, great. What are you seeing them doing? Where are they on their journey? They're obviously not replumbing for liquid cooling, most of them. So what are they actually doing?

David Schmidt

>> So what's resonated the most is, where we can have a conversation with our customers, show them our next-generation portfolio and really present a spectrum of choices for the environment they have. And that could be anywhere from five kilowatts to 50 kilowatts and up. By that, I mean the amount of power in a single rack. So we've found a lot of value having the conversation that way to say, "Where are you today? Where are you with the rack power that you have? And then let's talk about the type of workload, whether you want compute density, storage density, air cooling, or maybe you are ready for liquid." We have those ranges in the portfolio.

Dave Vellante

>> You hear so much about changes in networking and storage and IO. In the PowerEdge world, what are the requirements that are unique to that world as it relates to outside the server complex, outside the rack and the storage and the networking?

David Schmidt

>> Yeah. So within the server, you see Robert and team give us a great amount of... They give us all these great lanes, all these IO lanes. They give us all these cores, and we get to design a system to take advantage of all of that. So we can put a lot of storage in a platform. We can drive nine petabytes of storage in a single AMD-based platform. How are you going to get that data on and off the box? So now you have an IO component in the back of the system. Now you're trying to get something like 800-gigabit speeds so that you have the right balance of IO, the right balance of storage upfront, and the right balance of cores relative to the memory. Then you think about all those components in a platform, and it circles all the way back around to the thermal power conversation we were just having of, do you want to try to make that as dense as you can? If so, you're probably in a liquid-cooled environment. Or we're going to build chassis that are so well performing from a thermal perspective. We're going to optimize the airflow that you've got in your data center. We're going to optimize your cooling capacity. You're going to see that in all the next-generation designs that come out of us for PowerEdge.

Gemma Allen

>> So when we think about the year out, the next year we're back here, 2027, Dell Tech World again.

David Schmidt

>> Yes, we will.

Gemma Allen

>> What are the priorities for this relationship? It seems though there is ample opportunity here. There is a market for the taking, right? What are you guys really honing in on? Are there long-tail opportunities like an addressable market you guys haven't yet covered together? We talk about sovereign AI a lot. Where are you guys really focusing?

Robert Hormuth

>> I mean, I'll go first, and I would say that I think what we're really aiming for is ensuring that our customers are successful, first and foremost. I think the things that are going on, we started the discussion about why do cores matter more with agentic AI, and I think we kind of missed one of the other tailwinds that's been going on, especially with Epic and the Dell portfolio, we've been driving a consolidation story with a lot of our customers where they can do seven-to-one consolidation of old legacy servers to a modern, high core count, air coolable server. And they're able to free up space, capacity, power to make room for AI. So I think it's going to be actually kind of a duopoly of things that are going on, is customers are going to have to free up space and power to afford the space and power for AI. So we're going to keep driving the consolidation pretty hard with our customers, I think, so that they can make room to be successful in AI. Because the era of AI is not waiting. This is not one of these market trends that you can set back and debate for six months or nine months to enter, because your competition is going to go fast. Just all the disruptions that we're seeing at AMD on our own design side, whether it be silicon design, validation, software design, validation, it's not about doing the same with less, it's about doing more projects. We're able to take the resources we have and do more than we were. So I think that's going to be the mantra in most of the other companies, is how do I make room, space, power, and then invest in AI so I can outrun my competition?

Dave Vellante

>> Doing so, it's important within the same power envelope, right?

Robert Hormuth

>> Yeah.

Dave Vellante

>> So every, I don't know, whatever, 12, 18 months, you're on a steep improvement performance curve. I can now drop that into your point. Same infrastructure, I can, I don't know, increase my work by 30%, 40%, 50%, or sometimes more.

Robert Hormuth

>> Yeah. I don't know if David has a rule of thumb here. I'll ask the question for David. So how many servers do we consolidate down to make room and space for one AI server?

Dave Vellante

>> Yeah. If you're seven to one, let's say. I mean, I don't know, is that-

Robert Hormuth

>> That might be one area to cool a server.

David Schmidt

>> You'll see a 13-to-1 consolidation ratio in our latest generation. That's 13 just core rack servers down to one. So Robert's point, it's going to free up resources, power resources, rack space, et cetera, that you can then deploy for these agentic workloads that we've been talking about. To your question about 2027, I'm super excited about helping these customers manage that journey of, what do I need to do in my data center so I can meet all the demands and all the opportunities that Robert just highlighted? How do they address it? How do they redesign? How do we help them fit within our portfolio? And then, of course, look beyond that, I think the rest of this decade is going to be really exciting from a core compute perspective. So we should be back here next year talking about probably things, completely new things that AI has uncovered for us to go solve.

Dave Vellante

>> Last question is, what's the hard engineering problem your customers are asking? Your customer's right here. So the hard engineering problem that you're working on that if you could solve it's going to really drive customer value. Start with the silicon and then the end customer, if we could.

Robert Hormuth

>> I have to bring it back to the customers. There's a lot of hard engineering and silicon level that we have to go tackle, but tackling that without having a value proposition that helps a business go further, faster, add new business, create new revenue. AI to me is not about just keeping the lights on. It's not about... Traditional IT is running the business. AI is generating new revenue streams, new business opportunities, new customers. So that to me is the problem that we have to go tackle, is make sure that our customers are successful. Yeah, there's a lot of hard engineering on silicon packaging process nodes, but none of that matters if we don't solve a problem for the customer at the end of the day.

Dave Vellante

>> David, that's good framing to lead up to the question. You got to run the business, to grow the business, and to transform the business. Historically, the sort of on-prem business is, run the business, refresh cycle. Now you're driving new innovation. What's the hard problem customers are asking you to solve?

David Schmidt

>> Can I say all of it? They're asking us to help them solve all of it, but if I had to pick one right now, the storage aspect that rides alongside all of this great compute. We could go on another hour, we're not going to do that, but we could go another hour of just the amount of storage that gets dragged alongside this. They're asking us to solve and build data platforms that I know we've talked a lot about it here this week, but they're asking us to solve those data challenges and deliver them exabytes of capacity as soon as possible. To do that, you have to have a well-thought-out storage platform. We have excellent... We talked about exit scale at a rack scale level. We have great things with PowerStore. We talked about lightning and the record-breaking speeds that we provide there. All of that is at our disposal. We have all these tools in our toolbox to help solve those problems, whether they're giant rack scale challenges that we have to help our customers solve or transform my enterprise using the data center space that I have today.

Dave Vellante

>> So you got the infrastructure, you're ready there. It's that data problem that's the hard one. Varun Chhabra called it the hidden blocker or the hidden barrier. So, excellent guys, thanks so much for coming to the community.

David Schmidt

>> Thanks for having us.

Robert Hormuth

>> Yeah. Thanks for having us. Thank you so much.

Dave Vellante

>> Really appreciate your time and the update. Best of luck to you both.

David Schmidt

>> Thanks.

Robert Hormuth

>> All right.

Dave Vellante

>> All right, this is Dave Vellante for Gemma Allen, wrapped up day two here at theCUBE's coverage of Dell Tech World 2026. John Furrier will be back here tomorrow with myself and Gemma. Go to SiliconANGLE.com. We got all the news on DTW, and of course, around the enterprise tech industry. theCUBE.net is where you watch this. Check out theCUBEresearch.com. That's where we do all the deep dives. Go to the CUBE AI, ask it a question, find out what the top takeaways are from Dell Tech World 2026, and tune in tomorrow. Thanks for watching.