AI Compute: Rebuilt for the Agentic Era | Mark Papermaster, AMD | RAISE Summit 2026

Well, first of all, Kareem, thanks for the chat here today. Looking forward to it. But I'll go back before six months because when you think about the AMD story, we really had to reinvent what we were doing. It started with getting a leadership CPU. We got that out in 2017. So we've actually been getting very steady market share gain in CPU and in fact we're probably on a path to have server market share leadership in the next few months. So very exciting path. And with GPU, we've always been in GPUs and the key was first getting those GPUs back to be fully competitive. We did that. We won the top number one supercomputer in 2022, 2024 we beat that and had yet again a number one supercomputer, and then we flipped into AI and that's when we started. We went from 0 to 5 billion revenue in one year. That was in 2023 and we're on an annual cadence of new GPUs. But what happened in the last 6 months is agentic AI. So now those leadership CPUs are needed more than ever because when you're running the agentic applications, it's not just the GPU, you're actually using more and more CPU and that's how we're getting fundamental productivity gains. It's truly game-changing.

Yeah, it's a great question and I am very fortunate to have hit each of those inflection points. I mean, think about PC brought information to all of us that used to have to dig through exhaustive research, but the PC plus the internet brought all that information to us, and mobile with the phone brought it to where it was with us everywhere, as an extension of our body. And so all those were huge changes. But what's so different here is now with AI, with the type of models that we have now that have a reasoning capability that's outstanding, and with the agentic AI where you can really put the pieces together to get work done. We're going from productivity gains that we had prior to the last 6 months of 10% kind of gains to now dramatic improvements in productivity. And just in our own chip design at AMD, we're shaving months off time to market. We're taking new features that we needed that would take many, many months and we're getting it down to weeks and days.

Well, what you talk about is certainly we're all seeing that if you're transforming your business into being an AI-first, AI-driven business, and when you think about complexity, it's actually what's giving you the efficiency I just mentioned a moment ago. Because how are you getting massive efficiency gains? You're actually orchestrating many, many parallel workflows that used to be run serially. One team would get a piece of work done, they pass it over to the next team. It might be in a technology aspect, it might be on a manufacturing floor, it might be in accounting. These are all serial processes. Well, with agentic, we all hear about agentic agents. They're great. I have an agentic agent that speeds up my everyday. It's my right hand. But it's agentic workflows that are adding that complexity because now you're running many, many agents at one time. You're creating sub-agents that represent skill classes, groups that use...

Exactly.

Exactly. And the context window is larger. So think about the role of CPUs. Used to be, hey, I have a CPU as a head node, a controller for my GPU. So that's what everyone says: AI is GPUs. Well, now you have a bigger and bigger context window. We're all asking it, but the prompts are much more complex. We have to orchestrate that and that's done in the CPU. So you have this whole orchestration and reasoning level to do that reasoning. It's much more CPU to orchestrate, and then you've got to get the work done. And where is the work done? That's generally been CPU-based workloads. We have x86 CPUs, everything runs on x86 CPU. So all of that is changing from a ratio of one CPU to eight GPUs, 1 to 4, now it's literally becoming one CPU to one GPU. And also, so that's one way is just get more CPUs in. The other way is can we get GPUs to be more efficient? And you've got now a lot of different companies. Someone's going to have to win at some point, but trying to drive more efficiencies between GPUs. This is getting quite technical, but I had to learn all this stuff. It is GPUs, but honestly, Kareem, it's the whole system. Like we've changed at AMD. We used to think about and we used to be organized with a CPU division, a GPU division, and we have all those teams have to work together. Gone is the day of creating one little piece. You have to do holistic design. You have to design for the system all the way up through the application stack, and that's how we get the efficiencies. So it's feeding the beast. The beast is that huge CPU, GPU, your compute engines. So you need networking to be able to scale them together efficiently. You need to be able to analyze from the software. You have to look at where are all the bottlenecks and you have to eliminate those bottlenecks. And then you just have to really truly optimize one against the other, not one at the expense of any one component at the expense of the other components.

Well, two things. One, there's going to be efficiency of getting the job done like I said earlier. So really thinking about truly the inflection we're in, where you're optimizing for agentic workflows, you're running much more efficiently because you've got these racks of computers that used to run at a fraction of capacity. With agentic workflows and the parallelization you can do, you can actually run at a much higher efficiency. So one, you're utilizing the compute infrastructure you have in a much better way. And then what we're all doing, we're power limited across the world. We're energy gated in terms of building out. So that's a huge impetus to all of us in industry to collaborate, to work together. We do it in every aspect of our design, driving more energy efficiency into the chip. You're seeing new types of chips that are running inference much more efficiently, running computing more efficiently. So it's economically and frankly from just a pure ability to meet the demand, we have to drive more.

Constant process. I mean you literally every generation you're going to see significant efficiency improvements.

Yeah, it is a different business model and I'm actually going to tie it back to the efficiency question you had just prior. So everyone is struggling as they apply AI to make sure they can realize the benefit. It's expensive. And so the way to do that is for us to optimize solutions, and it is at the rack level. So we've acquired ZT Systems. We optimize an entire cluster of CPU, GPU, networking. It's all there and we optimize that. But that's not enough. You want to run AI where it runs best. You want it to run very efficiently on the factory floor. If you can analyze all that factory data right there, get answers right there, and compress what you need that has to go back to the big cluster. Personal AI is huge. I mean, you look at the new PC we put out called Strix Halo, and you can run hundreds of billion parameter models on that very efficiently. And you also have the sovereignty of owning all of that, and you can run that standalone from the internet. So it has changed how we talk about our products, how we go to market. It's all about what problem is it solving? How are we making you more efficient with our products?

Well, you're absolutely right. So again, AI is just not an option for any of us because it's affecting every single one of our personal lives every day, our work lives. If not, I think it's pretty well accepted. I certainly and our company certainly accepts it. You'll be left behind if you're not adopting. So our entire company including every person in sales has been trained in AI. They're writing their own agents. They're using it every day just as part of the training. We've hit 100%. We have a phase two training them much more in-depth based on the solutions that they're trying to sell. So the world has changed. We are adapting very quickly. Our goal is to think like a startup. If there is an AI native startup, how would they go about solving that problem? That's what we want to do.

It's actually our secret weapon. It's our secret sauce is our culture. We have that culture of being a scrappy underdog, a fighter. It's always been part of AMD. I joined with Lisa Su almost 15 years ago to AMD and we didn't lose that scrappiness, that underdog. But what we've added is a culture of execution, getting listening to our customers, understanding the problem we're solving and getting those products out to market at very high quality, at top performance. People run their business. They need to know you're delivering it when you said it would be there. And so I think it's that aspect of not losing the scrappiness and equally our culture of really listening to our customers and really collaborating with them. It's a huge differentiator for us.

We're 33,000 wide in our company now.

Yes, we have been committed to open systems, open ecosystems and open software stack. And it's been a huge differentiator for us. It's part of what I said earlier. We want to win on the merit of our products. We want to collaborate and create a partnership that is a win-win. And that we are able to do with this open ecosystem approach. I told you we bought a system company, ZT Systems. Well, we didn't just go compete with all our customers. We created a blueprint, a design of systems that we give to our system manufacturing partners. So we speed their time to market. We've done a holistic optimization with our ZT skills and with our customers. And that's just a great example. Those racks, this huge 72-node Helios AI rack, CPU, GPU, networking coming out this fall, is an open standard OCP. The networking of how we connect all this is an open standard, and our AI stack is open, fully open. But you could argue that openness could actually stifle speed. The fact that if you're closed, you know exactly what you're doing, you can move much faster in the market. If you're open, you're learning, you're adapting, more people coming in. Do you see a downside to the openness or things that you're being challenged at thinking through on that aspect? We play the long game. So what we do in any new feature we get out, we do drive that feature. We work with a set of partners rather than think about a new capability you need.

It absolutely has. And you think about our presence with running the Lumi supercomputer in Finland, with Alice, with KO here in Paris. We have ENI in Italy, HLRS in Germany. So we're not new to Europe, and creating really understanding their needs. And all of those systems are running our open stack. You take what we did in Lumi, that open stack allowed them to then tailor their models, natural language, native language Finnish optimization, and they control their destiny. And I have to really applaud the EU AI Gigafactory. When you look at early drafts, it really wasn't open to the industry, it was really tailored to one vendor. When you look at the revisions and where it's going now, it's not only accepting diversity and choice, it's encouraging it. And so we believe that the messages about our open ecosystem, not just us, open ecosystems means we're bringing many partners with us. And in Europe, we've targeted those customers that can help us build open and sovereign systems. Will you double down on Europe even more now or? Yes, we will.

Now, we're really excited. So what we do at Advancing AI is we bring in developers, we have workshops, we show the best practices. I talked about those AI agent workflows. We'll be showing examples where people are just creating amazing productivity gains. But we will be really talking about pulling the covers back on that Helios rack. I'm super excited about the information we're going to show and we'll put it out there. We'll show why we're so excited about it. And likewise, our new 2 nanometer Venice CPUs, they're killer. So we've got that as well as a lot more coming out in two weeks.

Thank you.