AMD Advancing AI 2026: Lisa Su Full Keynote

Morning.

Good morning.

We did. Yeah.

Yeah. Yeah. And first, thank you so much. A hard act to follow with Helios. Amazing. So with Anthropic, as you mentioned, the scale of growth in the industry is enormous. We've been working to make sure we have the best chips and can use the best chips for the best workloads.

Yeah. I think the first thing was just Helios is an amazing machine. It's absolutely fantastic.

And then the thing we were thinking about originally was that bringing up a new hardware platform is a big effort. It's a huge thing. As part of that, we started our own evaluation of MI355. You guys generously got us a rack to start working. We expected a big process. Our actual experience was we had one engineer start doing it. They spun it up, connected it to Claude, asked it to bring up the machine, left it going over the weekend, and we ended up with a graph of the performance of our leading model on it just going up and up and up over the weekend. And so yeah, I think that's a testament to the open platform you've made, where anyone, human or AI, can now build real models on your platform.

Yeah. So I think this is a great place for collaboration. The bread and butter of Claude is doing software engineering, and now more and more we're seeing it help with workloads like design layout, which is adjacent to software engineering. I think that's where we can work together to make the next generation of chips even better.

Yeah. That's a great question. I think working together for the scale-up is probably the biggest thing I'm most excited about. We see consistently that using more and better compute results in better models that add more value to downstream users. That's the biggest thing. Also, now that we're investing more with you, security is a place where we can work together. As machines get more complex, we can secure things at the chip layer, server level, rack, and whole network. That will help not just us but the entire industry stay safe.

Thank you so much.

No, thanks. Great to be here. I was thinking about this event over the last two years. It feels like attendance is following scaling laws, doubling every year. But more seriously to your question, we've always predicted that models will evolve from chatbots to reasoners to agents to interns. That trajectory has kept up. We are seeing models become more capable, more agentic, able to do long-running tasks. All that means we need to keep scaling compute. The more compute we scale for training, the more capabilities emerge. The more compute we scale for distributing intelligence, the more people want to do with it. That's what we see with exploding token budgets. Internally at OpenAI, not just engineering but every aspect of the enterprise is beginning to use agents for every part of work. This is why we need platforms like yours that can scale with how quickly capabilities are scaling and how quickly the world is embracing these things.

It's been a phenomenal partnership. As you mentioned, we were first to bet on AMD, and we are thrilled with how that bet is turning out. We started on MI300 and expanded very quickly to MI355. It's been fun for our team to see how we could optimize the software stack, networking, and everything to deploy our models on AMD infrastructure. As Lisa mentioned, we just got hands on the Helios racks three months ago, and the collaboration has gone to the next level. Our engineers are working side by side with AMD engineers to optimize the software stack and run GPT-class workloads on Helios already. Phil will probably be on stage later talking about how productive and fast that collaboration is. We're really excited about the capabilities Helios is already showing us, and we expect to be deploying Helios at massive scale starting towards the end of this year and accelerating throughout 2027.

We need it earlier.

Yeah. I think one of our key bets is about recursion: AI recursively improving the systems it needs to run on, and eventually AI doing the research itself. Today, expert engineers spend enormous time optimizing kernels, translating new models to run efficiently on different hardware, dealing with compilers, communications libraries, system configurations. What we've been super excited and encouraged by is the work we are doing with you, where AI is showing the potential to automate a significant portion of the work when anyone programs to an AMD GPU. It dramatically shortens the path from a new model coming up to efficient production deployment, and something we can much more easily tune to changing workload requirements. That's a big deal. Also, because AMD is built around an open software ecosystem, all these things we're inventing internally for our own use, we believe we can bring to the whole world and allow the whole world to use them for all AI models, not just ours.

I think the thing the whole industry is realizing is that it's a systems problem at a data center scale. We quickly internalized that AI was a rack-level problem, but now it's clear it's not just a rack-level problem; it's a rack in a data center. The whole data center is a system. We have to think about how we design the future based on what we know about our workloads and what's coming down the line, but co-design it with you from CPUs to GPUs to memory, networking, storage, power distribution, and cooling systems. These are all intertwined and have to be co-designed. One of the fun parts about working with you and your team is how easy and productive it is to give you insight into how we expect the world to change, and how quickly your team responds, turning those insights into better chips, systems, and software. That worked really well with MI400 and left-shifted a lot of engineering work. That's why we are very confident we can productionize those systems quickly. We're really looking forward to accelerating that with MI500 and beyond, and increasingly using AI to do that work for us rather than being limited by humans.

Thank you so much.

Good.

And you lifted it up. I have to say they won't care.

Sure. First of all, thank you. It's awesome. This is like a hardware geek's paradise. You come in, you're talking about transistors, you're talking about hardware, you're talking about cooling. This is sort of the jam. I like it. I think this is an audience that's receptive to it. I usually talk to audiences that have either not been to a data center or not really seen a chip ever. So this is refreshing. The thing I'll say is we are seeing demand go through the roof. When you look at inference, training, recommendation systems, that's how a newsfeed works. Content creation, the demand for that is just exponential right now. And when I think about the overall, Mark has this vision of delivering personal super intelligence to billions of people. So if you go to Facebook, Instagram, WhatsApp, or whatever surface, the idea is that we'll meet you there and deliver super personalized intelligence wherever you are. That's why we established this Meta Compute initiative, and because it's a top level thing, Mark himself oversees it. Now what that means is that we are now moving away from an area where we used to take servers and optimize them. This is what the conversations we were having many years ago: 'Hey, I'm deploying a CPU, I really need to maximize performance out of it.' It's different now. You talked about how we should be thinking about the whole system end to end. We are at a point where we have to look at the whole data center and think about it as one integrated system. Servers, hardware, networking, cooling, power, all of that is not optional. All of that has to work together. And this is where I think there's a huge opportunity for collaboration with partners like AMD, because you have to now co-design, co-create systems, not just deploy something that comes off the shelf. The other thing is flexibility. The thing I really liked about what you said is that this is such a big opportunity in the industry right now. All of us need to lean in and work on this together. It needs to be open. It needs to be heterogeneous. There's no one company, one partner that works for any one of us. We need to work on this together. And that's why I think you're one of our most important strategic partners.

Sure. It's been many years. The story that Lisa was saying was many years ago. I think we went from Milan to Bergamo to Turin now to Venice. So it's at least the fourth generation. Long time partnership. Obviously, I actually think the world is changing in the sense that every year is a new world these days.

Months exactly. So what ends up happening now is that it's not just a GPU game anymore. It's CPUs and GPUs. If anything, I think CPUs are becoming at least as important if not more. You're looking at a world where the workload is changing because while you have your workhorses in GPUs, at the end of the day, for agentic workloads you have to run your tools, you have to run your systems, and all the code that's being generated still needs to run somewhere. So I'm super excited to work on Venice. I think there's a lot to come there. And you have to think again, this is a theme that I assume will go throughout your presentation, that you have to start thinking about CPUs and GPUs as conjoined things. You hand off workloads depending on the workload, you employ the right hardware.

Again, multi-year collaboration. I think we started with MI300, we did a bunch. It had a really nice memory capacity. I remember talking to you about that. And then there's a bunch of pipelining we did. It was important to deploy that at production scale, so get hands-on experience on both sides. 350 is the first time we deployed this on ranking and recommendation systems, so that's now getting into some good numbers. And 450 is, I think I'm super excited about it. We've got some of the racks and we are going to deploy it across the board because it gives us the opportunity to really collaborate. I think the big difference between the 300, 350, and 450 is that we have had an engineer sitting in the same room talking to each other, co-designing, co-creating, like power cooling. As I was saying, it's not just a single thing. And we are not shy in feedback.

But you're very deceptive. So this is true partnership. This is the point about co-designing and co-creating that I think, and as I said, that's a huge deployment coming. The more we deploy, the earlier we co-design, the better we are.

Listen, all of you know about the bottlenecks, you know about the power, the data centers, the silicon. All of these are choke points that the industry is sort of waiting on. But those are things I'm pretty confident we'll work our way through. At the end of the day, I think that demand is immense, people are responding. But the thing that I really want to make sure all of us realize is that we should not think about systems in isolation. As I was saying, CPUs and GPUs are one conjoint system. You should start thinking about it that way. We really need to co-design things early. Data centers take years to build. One of my favorite stories is Mark comes to me and says, 'Hey, I want a gigawatt worth of data centers.' Well, you should have talked to me two years ago. It takes time. It's the same with silicon. It just takes time. So we need to be starting and sitting down in a room co-designing today for what we need to deploy in 27 and 28. That I think is when we truly unlock the powers of the system. The systems are amazing, the specs are amazing, but think about how much more you can get out of it if you sit in a room and just co-design it today, and then in 28 we'll be having much better graphs out there.

Thank you. Thank you.

Sure. Great to be here and great to be among people who love hardware. It's nice. Look, I'm thrilled to be here with you and to announce our cool new partnership at Cerebras. We build the world's largest and fastest chip. It's a full wafer and we package this into a system and systems into racks and then racks into clusters, and we deliver them both on premise and via the cloud. And you've had some of our customers up here already to date on the frontier labs. We serve customers like OpenAI and in the hyperscalers like AWS, in the small agentic and the coding space leaders like Cognition, and they do this because we're blisteringly fast.

Sure. Right. I think what's happened and you described it previously is that AI has moved from being a novelty to being useful and then in some domains a necessity. And when something's a necessity, people want to use it and they want to use it quickly. And to serve this market, this segment of ultra low latency, we sought a partnership that could extend our capabilities and our footprint. And there was no better answer than AMD. We were already an AMD customer as we use AMD CPUs to surround our systems, and we were so excited at the opportunity to build a disaggregated solution that combined AMD CPUs, the cool new Helios rack, and the Cerebras wafer scale engine.

Sure. So as Lisa described, what you have with Instinct in the Helios rack is you have the leader in performance and memory capacity, and you marry that with our wafer scale engine which is the leader in SRAM and in memory bandwidth, and that combination allows us to deliver a solution that is unmatched in the industry.

Later this year and it will be first in the Cerebras cloud and it will be later at a store near you.

I think that's right. I think until recently customers could have high throughput or they could have extraordinary speed. And by bringing together the Helios rack with the Cerebras wafer scale engine, we give you five times the throughput while continuing to deliver this extraordinary speed. It's really something amazing.

Soon.

Good to see you. Thank you so much.

Thank you Lisa. Good morning everyone. It's great to be back to talk about software. Just a few years ago, programming AMD GPUs required deep expertise and significant engineering effort. We've come a long way in a remarkably short period of time. Through our open source community collaboration and sustained investment in Rockom, our software stack, AMD Instinct GPUs are powering some of the most important and consequential workloads on the planet. But the biggest change is still ahead of us. AI is transforming how software is built and we are bringing that transformation to Rockom. Our software teams have been moving fast with relentless focus on developers. We've invested at every level of the stack in build and test infrastructure so we ship faster. Rockom releases now go out every six weeks, not every four months. We've expanded the work we do with our AI ecosystem partners and the result is more features, faster performance, and a significantly better out of the box experience. Our strategy is built on two core principles: partner deeply with the open-source ecosystem and build with the right layers of abstraction to enable developer productivity. Open source gives us velocity and scale, and abstraction makes developers more productive. Our deep commitment to open source has resonated with the community. From frameworks and compiler stacks to inference engines and model hubs, AMD is now becoming part of default enablement for the most important AI communities such as Hugging Face, PyTorch, Jax, VLM, and SG Lang. This is why new models get day zero support and why the world's most important AI workloads run on AMD today. That is a big shift. Inspired by engineers seeking productivity, abstractions have evolved to enable work at the right level of detail, hiding complexity when they want speed and exposing control when they need performance. From low-level programming in assembly and C to block programming abstractions like Triton to Pythonic frameworks, we've invested in key abstractions. Aer, our kernel library, and Atom, our serving engine, get you peak performance without writing the kernels yourself. Fly DSL is a brand new Pythonic domain specific language that gives you low-level control with the performance of hand-tuned C++. and Mori, our communications library, helps deliver leadership performance on the latest models like Mixtral.

But look, something very exciting is happening. Over the past year, I've seen something remarkable inside AMD. Our engineers are using AI models to generate GPU kernels, optimize code, debug issues, and improve performance. In some cases, these AI generated kernels are shockingly good, better than what we expected, sometimes better than the most manually tuned versions. The first time you see this, you're skeptical. You run more tests, you try to break it, you look for what went wrong, and then you realize this is real. What has happened in general software development is coming to significantly transform GPU programming. It will reduce the time to bring up workloads, make optimization automated, and remove barriers to broad adoption. We want to put that capability in the hands of every developer. So today I am excited to introduce Rockom.ai. It's an agentic AI platform that brings AI assisted GPU programming to developers. It starts with giving developers access to the power of the AMD ecosystem through the AI coding agents they already use. Whether you're using Cursor, Claude, Codex, Rockom.AI helps those agents understand AMD platforms, understand Rockom, and helps you build and optimize your workloads. We are making those popular coding agents into Rockom super users. You describe the workload you want to run, the performance target, and let the agents help you get there. For years, we have been building the open software foundation. Now we are adding an AI assisted layer that helps developers use that foundation faster. Let me show you what is inside Rockom.AI. It's built on the foundational capabilities of Rockom: core software stack, runtime, libraries, compilers, tools, framework integration. On top of that, we built a layer of AI assisted optimization called Hyperloom. Using Hyperloom, the system can analyze a workload, tune configurations, select and tune kernels, adjust parallelism strategies, and iterate towards performance goals. To give an example, I was talking to the lead developer last week. He told me they just pushed a suite of 14,000 models through Hyperloom, optimizing them and creating valuable insights. This would have been impossible even with a large team of engineers before. That's the power of Hyperloom. And finally, we provide an AI native interface for developers: AI skills that allow coding agents to understand Rockom natively. I am incredibly excited by what we have built. It will profoundly transform how developers interact with our platforms. Let me make this concrete with an example. Imagine you want to optimize Mixtral M3 with VLM on MI355s. Today you need the right VLM optimizations, model recipes, environment variables, and you may need to write new kernels. With Rockom.ai, you say 'optimize Mixtral M3 with Hyperloom.' The agent pulls a known good recipe, runs and profiles the workload, creates and tests many configurations, and iterates towards the target. You can watch it write a GPU kernel. It sees an opportunity for a more optimized group gem. The result is a 38% improvement in tokens per second. This is the experience we want. Simple to start.

That's pretty cool, huh? Simple to start and powerful to optimize. We've continued our relentless focus on performance. With every release, we've delivered significant gains on leading models like DeepSeek. Rockom.ai delivers 3.3 times speed up over Rockom 7. Innovations at every level: block scaled fused kernels, KV cache quantization, expert parallelism. But the bigger story is how AI agents profiled the workload, proposed new kernels, tested configurations, and validated results. That same focus is on training. On the same hardware, Rockom.ai improves training performance by an average of 2.4 times through optimized kernels like fused flash attention, efficient checkpointing, and advanced parallelism strategies. Our engineers have done amazing work to unlock the capabilities of MI455 through Rockom.AI. So I'm delighted to share some incredible results on hardware running real workloads. We are able to see the system deliver 20 terabytes of memory bandwidth and 20 petaflops of delivered compute. These are the highest demonstrated compute plus memory capabilities of any accelerator platform in the industry today. It's measured, not theoretical. Super exciting to share this on day one. Also thrilled to share that leading AI ecosystem partners who have had their hands on MI455 have had a delightful experience. From end to end PyTorch testing to Hugging Face model validation to VLM and SG Lang running well on MI455. We are incredibly encouraged by what the community is about to unlock with MI455. This matters because day zero readiness is one of our most important objectives, and that's what Rockom.ai enables.

So now let's look at Rockom.ai in action on Helios. This time we are using a Codex front end. Let's deploy DeepSeek V4 Pro, a 1.6 trillion parameter frontier model on Helios. Watch what happens. Rockom.AI pulls the right recipes, configures the runtime, implements the model graph onto hardware, and serves up the model. The model is up and running. Let's see it write a simple poem about advancing AI. Isn't that a pretty cool poem? Now there are few organizations shaping the frontier of AI as much as OpenAI. You heard Lisa and Sachin talk about our partnership. It's been truly wonderful collaborating closely with the OpenAI technical team to make rapid progress with Rockom. To talk about how they are advancing the frontier of software development and our collaboration, I'm delighted to welcome to stage Philip Tle from OpenAI.

Thank you for joining us here, Philip. So, just so you all know, Philip is the creator of Triton, one of the most important software innovations in modern AI infrastructure. It's a huge privilege having you here.

Thank you.

OpenAI has always pushed the limits of infrastructure. So, can you tell the team a little bit how your team develops GPU software and how technologies like Triton have evolved?

Yeah, for sure. Well, first of all, thanks so much for having me. I was here 2 years ago. It's my great pleasure to be back now. Look, our way of looking at things is there's no one way to fit all applications. We try to meet users where they are. So we've developed multiple solutions. For workflows where speed of iteration is more important, we've developed Triton as you mentioned. This is typically what researchers use. But more and more we've been in cases where we've had to hyperoptimize performance and Triton didn't give us the level of control we wanted. So we've developed Gluon for that, which is a lower level language. This way we've been able to optimize our kernels for AMD among other things. Our agents have become extremely good at both of them, and we've worked with AMD on both as well.

Yeah, and more and more what we're seeing is agents slowly taking over. They're getting very good at writing this kind of code. Amazing. So look, I think the work you have done with Triton and now Gluon has been really moving the ball forward for the entire industry. Now we began our work together with 300 and extended at 350, but now we're collaborating at a different scale with 450 and Helios. So would love to hear your thoughts on how that collaboration has been going and how has that experience been.

Yeah, super well. It feels like ages. I think we started years ago on hardware co-design when MI450 was in the very early prototyping stages. I think we were able to work really well together and it's so nice to see the outcome of that and the chip. More recently, we've been working very deeply on software enablement for MI450 both for our models and for the broader ecosystem. A lot of this development has been done open source. And yeah, we're collaborating on a full stack now, down to the very low-level LLVM code generation, so that all the nice hardware advances that came with MI450 can be properly leveraged in end-to-end applications.

Yeah, and you got your Helios racks. So tell us a little bit about that.

Yeah, it worked like we got the chip and very quickly a single engineer in a few days was able to confirm that everything worked. Obviously we have to optimize performance more, but we're very confident we can get in a very good place.

Awesome. That's awesome to hear. So look, one of the more amazing things we've done recently is teams have collaborated even more closely in terms of using AI to program AMD GPUs. So would love to hear your thoughts and anything you can share with the audience.

Yeah, no. I mean everyone in this room knows how much agents are taking over our jobs as kernel engineers. It's been amazing to see how much better they've gotten over just the past 6 months. They're extremely good at generating high quality GPU kernels in a way that were not before. It's been great collaborating with AMD on making sure that agents not only are capable enough, but also have all the right context they need to make good decisions and good optimizations.

That's great to hear. Yeah, I think the quality of code produced by AI assisted kernel generation has been truly remarkable through our collaboration. Very excited. Now as you look ahead, what crystal ball do you have for us for the future of AI software ecosystem and collaborations like ours?

Yeah, so there's two things that come to mind. One is just how good agents are getting. The other is how much the open approach to software of AMD is enabling not only for OpenAI but for the industry as a whole. I talked a little bit earlier about how we've been collaborating on LLVM code generation. LLVM is an open source project. The entire compiler stack from AMD is open source, and that has allowed us to make our agents really good at very low-level code generation down to scheduling instructions in assembly. This has led to very significant performance gains that I don't think we would have been able to achieve in a fully closed source stack.

That's one of the real big value propositions we have to offer, because everything we do is in the open and the fact that AI can actually pick that up is a big one. Thank you, Philip. We truly appreciate the partnership and we're really excited about what we're going to build together.

Yeah, my pleasure. Thank you so much for having me.

What you just heard is extremely important. When we combine better abstractions with AI assisted development and close hardware software co-design, we can create incredibly powerful capabilities. That's exactly aligned with our strategy. Now everything I have shown you so far has been about AI, but Rockom is also a great stack for developing high performance and scientific computing applications from frontier models to climate simulation. We use the same Rockom, same library, same tools, same AI assisted development, one open stack for every workload. When you look at our road map, we've retained that strong focus on HPC and scientific computing. Just like MI455 leads in AI compute with FP4 and memory performance, MI430X leads HPC compute with FP64 and memory performance. Two platforms purpose-built for very different classes of workload. The MI430X is built by leveraging our modular chiplet architecture to deliver a purpose-built variant with native FP64 hardware, giving customers leadership performance across AI and HPC in a single platform. This is hardware double precision, not emulation. For the scientific community, that distinction is everything. This product delivers 288 teraflops of FP64 compute, nearly nine times better than competition, and the same leadership memory capacity and bandwidth of MI455. The MI430X ships in the first half of 2027. This is why leaders in sovereign AI love this product. The first sovereign exascale AI factories in both US and Europe are being built on the MI430X at Oak Ridge National Labs and with Cineca in Europe. So look, as we look ahead, let me leave you with this thought. Every so often our industry goes through inflection points. We've felt it when neural networks first started recognizing cats and dogs, when transformers changed what was possible with AI. I believe we are experiencing another one right now, a real inflection in the ability of AI to program complex hardware systems. Building on the surface area from open-source software and leveraging abstracted interfaces, AI is making it dramatically easier to program AI systems. We cannot wait what you will build with it. Thank you.

Now AI is getting pervasively infused into all forms of computing. Enterprises are experiencing the same shift with their own unique requirements. So to tell us more about that transformation, please welcome to stage senior vice president and general manager of compute and enterprise AI Dan McNamara.

Thank you, Vomi. And good morning everyone. It is a great pleasure for me to bring to you the third pillar of our strategy, which is powering AI everywhere. And the vision behind this has remained the same for several years: to deliver the right compute engine to the right workload and then continuously drive optimization points across our entire portfolio from CPUs, GPUs, networking, and FPGAs. So, building on what we've already shared, I'll cover how our strategy comes together across our enterprise portfolio and how the foundation we built with the CPU franchise has truly positioned us for the next era of AI across multiple deployment models. Lisa talked about the frontier model and hyperscaler landscape and the rapid and massive advances we've seen over the last 6 to 12 months. Those advances are driving the adoption of AI well beyond the cloud. As AI continues to grow, it will demand compute beyond the massive purpose-built data centers, extending into enterprise, personal, and physical AI. And each of these brings a new or unique set of requirements.

Enterprises are usually constrained by power and cooling. Personal AI usually must operate within a fairly limited compute and memory footprint, and physical AI must perform reliably in real-world demanding environments. AMD is the only company delivering a complete portfolio that spans all of these model development models. So I just mentioned that our journey began in the data center with our CPU franchise, and I want to touch on it a bit. With each generation of EPYC, our strategy has been to listen to our customers, address their evolving needs, and deliver the highest performance, lowest total cost of ownership, and the fastest time to value. As Lisa mentioned earlier, we've introduced many industry firsts across our generations that our enterprise customers have actually asked for. That customer focus has brought us to where we are today with EPYC as the leader in enterprise computing. Over the last three years, there has been very strong adoption across the enterprise, with the world's largest companies moving more and more workloads to AMD. It's reflected in a number of places: first, public cloud where our VM consumption has grown at 76% CAGR over the last three years, and on deployments with our OEM and ODM partners, it is growing quite aggressively across key industries. But just as important, this position allows us to understand truly the diversity of the workloads that our enterprise customers run every day. Each workload places new demands on the CPU. Some require maximum thread count, some depend on per-core performance while others depend on memory bandwidth, cache efficiency, or IO performance. The really important point is that there is no single skew that solves every enterprise application. That is why our EPYC portfolio with Venice spans 8 to 256 cores with a range of power, frequencies, and IO capabilities, giving customers the flexibility to choose the right solution for their workload, and most importantly at the most optimized cost. So Lisa mentioned this earlier, and now we all know we covered Venice, and she showed how we have strong leadership across AI host nodes and agentic workflows. But I wanted to show you how Venice actually leads across the general-purpose workloads of the enterprise, these workloads that our customers rely on every day. As you can see, we have a minimum of 2.5 times the performance across our competition on these very important enterprise workloads. So now as we all know, agentic AI is really the next major transition for the enterprise, and it truly is a force multiplier for enterprises, and adoption is accelerating. In fact, the latest data shows that nearly every enterprise next year will have some form of agentic deployments. As enterprises move from pilots to production deployments, we talk to CIOs all the time, and they are looking at three major things: first, predictable infrastructure cost and token costs; second, making sure the data is secure; and lastly, and sometimes this is overlooked, integrating AI into their existing infrastructure. It's not all greenfield for the enterprise customer. So to solve these problems, we truly believe this is going to be a distributed deployment model that leverages frontier models, cloud services, on-prem infrastructure, and of course AI clients. As AI has moved from chatbots responding to prompts to agents doing real work, writing code, summarizing documents, and running complex flows, the requirements have evolved, and those are exactly the requirements we have been designing for for multiple years. So while AI will be deployed across all these environments, on-prem deployments in our enterprise customers have fairly unique constraints. First and foremost, power and cooling is a big challenge in space, and of course costs remain top of the list for CIOs trying to drive things forward with AI. And lastly, the full solutions need to be easy to deploy. That has been common for many years for the enterprise, but it is something we are very focused on. So to address these constraints, thank you, Laura. I am very happy to announce the launch of the Instinct MI 350P.

So today, every enterprise wants AI in the data center, but the challenge is how to add those capabilities without starting over. That is exactly what the 350P was designed to solve. It is an air-cooled GPU designed to fit within the power and cooling of enterprise servers today without a facilities upgrade, making it easy to bring LLM-scale inference into today's enterprise data center. At the same time, there is no compromise in capability. A single MI350P can support up to 260 billion parameters, allowing customers to run the majority of enterprise AI workloads on a single GPU. The economics are even more compelling: the 350P delivers more than four times the tokens per second per dollar than the competition. It is pretty cool. It actually turns the customer's existing data center into an AI data center. Now, we all know that benchmarks are one part of the equation. The other part is evaluating whether the advantage holds up across the actual workloads you run and how your business runs. So we tested a set of enterprise use cases using production models, and the 350P delivers much higher productivity than the competition, delivering up to two to five times faster tokens per second. So if you think about that, more tokens per second, more work completed, more users supported, all at a lower cost for the business.

So another area I wanted to talk about is our earliest customer, our own AMD IT team. They are a tough group to work with, even for us, but they put us through our paces. This has been our model from day one: we want to put our technology into our own data center first. So like many companies, we are looking at how to deliver AI services to our employee base while managing costs, protecting our data, and controlling the data. We focused on two use cases: first, autonomous threat detection, which is probably one of the faster-growing agentic applications we are seeing today; and second, a personalized AI assistant running on Open Co-pilot. Both of these route requests through the router and gateway, testing the complexity, sensitivity of the information, and service-level requirements. Some requests are sent to the frontier model, while others are routed to open-weights models running on EPYC and an MI350P. With intelligent routing, we reduced our token cost by 43% while delivering up to 3x faster response times for the workloads running locally. So we believe this is how enterprise AI will be deployed.

But we also know from our history with EPYC that enabling enterprise takes a lot more than great silicon. It requires a complete ecosystem of software and solutions. Over the last several years, we have truly grown and expanded our enterprise AI ecosystem. Today, we support numerous AI ISVs, provide zero-day support for all of the industry's leading open models, and deliver robust frameworks that help customers bring AI into production as fast as possible. It is all built on trusted platform, software, and OEM partnerships that enterprises rely on every single day.

Now, everything I just shared comes down to one goal: helping customers deploy AI where it creates the most value for them, whether in the cloud, on-premise, or at the edge. So to bring this to life, I would like to welcome Jeremy Le, Chief Technology Officer at AT&T. Hey crew, good to see you.

Yeah, thanks for having me.

Yeah, this is great. It is kind of a homecoming. I grew up in the East Bay, so it is good to be back.

Yeah, fantastic. We really appreciate you joining us today. So look, I think everyone knows AT&T is the largest communications company, one of the largest in the world, and I am sure half of the people out there probably have an AT&T network right now. But you know...

Hopefully more than half, but yeah.

But AI is touching way more than the network, and you and your team are doing a fair amount of work driving it. So can you share with us the opportunities you are working on and your experience to date?

Sure, happy to. At AT&T, we are now burning about a trillion-plus tokens per month, and that number has gone up pretty dramatically, moving up double digits. So this is becoming very pervasive inside the enterprise. We have over a hundred gen AI models in production across our enterprise, stretching from things like customer care and fraud to things you would not necessarily think about, like where to place a cell tower or RAN for the most effective use. All of these things are happening across our enterprise at a scale that is not necessarily normal. We get over 300,000 calls a day; the transcription of those calls is an enormous workload, and driving insights out of those calls is another set of insights. It is a really interesting thing. Now we are beginning to rebuild entire workflows inside the company beyond just point use cases. So as we think about HR, finance, and different kinds of things, we are now identifying those entire workloads.

Yeah, so it is you know we work closely with your team, and we have had many conversations. It is truly amazing the number of opportunities you are addressing with AI, a super broad range of use. So can you explain the key learnings in moving these to AI at enterprise scale and maybe a bit on what we have done together?

Sure. I mean, I will start at a place that I do not think everybody starts from, which is the human. You need world-class teams to do this, and you need people that think in workflows. There is an enormous amount of what I call racing from stoplight to stoplight – folks going from zero to 100 and then hitting the next stage of the workflow and a red light because they haven't worked their way through. So I would start with people and the quality of the people you have across your enterprise. The second point is we are big believers in data sovereignty. Data sovereignty for us means not being tied to a specific chipset, model, or set of development tools. You need as an enterprise to manage your data; your data is your fuel for implementing AI. What we have done with AMD, which has been a great partner, is leveraging open-source models, leveraging different chipset sets, building our own models, and retraining other models. We are one of the few companies that has actually post-trained a model on AMD and then been able to drive equivalency in performance and accuracy with other more expensive chipsets. As we have done that, we have been able to drive down token cost through moneyballing across all these different models, leveraging that human capital in a way that has made a big difference. So increasingly, even though token consumption is going up, we are able to manage that underlying token cost at enterprise scale, which is not a small thing.

Yeah, it is just a bunch of great work. Another key area that is groundbreaking is that you are the first telecom company to actually train a model on AMD, and you are driving fully open-source models in telecom. So tell us about that experience and why you felt it was needed.

We felt the industry was moving down a path where folks were only going to use closed-source models to solve solutions. While closed-source models are an important ingredient, we did not want to lose sight of the fact that open-source models would be a very significant part of the solution. So we announced what we call 1.0, the open telco AI model, a bit over a year ago, and we have had over 18 million downloads since then. We are excited to announce today the launch of Telco 2.0, a more advanced set of models that we have trained on AMD, and we are making it available via open source across the broader industry. I think it will give folks an opportunity to see that you need to think about not only leveraging models but also training those models and how they work across your enterprise.

So that is just great. We certainly appreciate the partnership. I really feel like this is germane to the conversation; it is exactly what we have been talking about today. I just want to thank you again for joining us and for the partnership.

Thank you. You are a big part of it.

Thank you.

Okay. So let me close with one last point. We have established ourselves as a leader in enterprise computing by listening to our customers and delivering the right solutions for their workloads. Today, we are extending that leadership to help customers accelerate their AI journey. This is our vision for powering AI everywhere. To expand a little bit beyond the data center into physical and client AI, it is my pleasure to introduce Mr. Jack Yun, who leads our Computing and Graphics group. Thank you.

Thank you, Dan. It is great to see so many friends, partners, and developers here in San Francisco. Today, we have seen how AMD is building the compute foundation for AI in the data center. But the full AI transformation requires two more ingredients: intelligence that becomes personal and intelligence that can understand and act in a physical world. Those are the two frontiers I want to explore with you here today. Let us begin with personal AI. Enterprise AI is already reshaping the data center, but agentic AI will extend far beyond it. Agents are the next great productivity multiplier, the jump from software that responds to software that completes work. A personal agent can double what one person can do; a team of agents can multiply that by 10. As agents gain more capabilities, they can grow exponentially. But that promise has a compute consequence. Unlike a single AI query, as an agent reasons, calls tools, interacts with other agents, and runs non-stop, agents become more capable and pervasive, so will token consumption. That means demand will require every layer of the computing ecosystem to scale together. Data centers remain the foundation for frontier AI in the most demanding workloads, but edge and client systems can extend that capacity, putting more intelligence closer to where data is created and where work gets done. The PC is already one of the most powerful compute engines ever placed in the hands of an individual, across billions of devices. Enormous computing capability is already deployed every day, yet its potential remains largely untapped. Putting that capacity to work changes the economics of computing. More workloads can run on systems that are already in place, reducing overall infrastructure costs and allowing data centers to focus on the largest, most demanding tasks. Local computing brings inherent advantages: your data stays with you, your applications remain responsive, and your system keeps working even when your network does not. Thanks to a new generation of smaller, more capable AI models, we are beginning to unlock that potential. The unlock is happening faster than anyone expected. Last August, GPQA needed 120 billion parameters to score 80 on GPQA. Just seven months later, Qwen 3.5 scored even higher with only 9 billion parameters, 13 times fewer parameters. This is not an incremental improvement; it is a dramatic shift in the compute required to deliver advanced intelligence. This is a trend, not an outlier. Smaller open models are closing the frontier gap at extraordinary speed. Qwen 3.6, a 27 billion parameter model, now outperforms leading frontier models from the previous generation on graduate-level reasoning at a fraction of the size and cost. That changes what is possible. Workloads once reserved for the cloud can now run on PC compute engines designed for personal AI. Let us map these model performance breakthroughs to AMD platforms. Ryzen AI 400 supports models up to 24 billion parameters, more than enough for Qwen 3.5 with substantial headroom. Ryzen AI Max extends that capacity to models as large as 200 billion parameters running natively on a personal system. This is not reduced AI compressed to fit on a PC; it is a new class of personal computing built to run powerful models where the work actually happens. This is exactly what led us to build Ryzen AI Halo. We wanted to give developers a platform powerful enough to run these models locally yet simple enough to use every day. With 120 gigabytes of unified memory and support for models up to 200 billion parameters, developers can build, test, and iterate locally on their desk. Ryzen AI Halo embodies the future of personal AI for everyone, shortening the distance between an idea and a working model. For developers, that means a dramatically more efficient and optimized path from experimentation to deployment. As Vont showed earlier, ROCm is making hardware programming radically more accessible. Developers can understand, modify, and optimize code for ROCm even without deep ROCm expertise. Better yet, the same open software foundation extends across all of AMD's hardware. Developers can use the frameworks and tools they already know, start locally on Ryzen AI Halo, and scale across the full AMD compute platform without rebuilding their work. We are not stopping there. Today, we are expanding our partnership with Hugging Face to bring the open AI ecosystem directly on Ryzen AI Halo. Powerful compute only creates value when developers can quickly access the right models, optimize them for the platform, and turn them into real applications. Together, AMD and Hugging Face will deliver Halo-optimized performance for open models and agentic workflows. How cool is that? Along with co-engineered libraries and toolkits designed to help developers spend less time configuring infrastructure and more time building, later this year every Ryzen AI Halo box will include a full year of Hugging Face Pro, putting the power of partnership in developers' hands from day one. Freedom and more power to the user. This is how we accelerate innovation in the AI era. We are already pushing that vision further with Gorgon Halo. Seeing is believing. Look at how small and beautiful this box is. We have increased unified memory from 128 to 192 gigabytes, the largest unified memory pool in its class, and increased model support from 200 billion to 300 billion parameters. Thank you, Laura. This is no longer limited to a single platform form factor. Together with the world's leading OEMs, AMD now powers the broadest portfolio of personal AI systems, from laptops and workstations to many PCs and developer platforms. That means developers, enterprises, and creators can choose the right system for the way they want to build and work. Personal AI is not a concept; it is a category, and it is available right now on AMD. We now have all the pieces. Models are efficient enough, hardware is powerful enough, the software stack is open and ready. Open harnesses like LM Studio can route workloads to the right model and right compute. The technology is here. The next challenge is taking it from one developer's desk to thousands of users across an enterprise. Cisco sees and believes in that future the same way we do. To take us deeper into how we turn that vision into enterprise-scale reality, please join me in welcoming to the stage G2 Patel, President and Chief Product Officer at Cisco.

G2, how you doing?

How are you?

You look great.

Thank you for joining me here today.

Congratulations. Thank you. Now G2, you and I talk about how we are at an inflection point. For the past few years, the industry has spent time just building intelligence. The next phase is how we deploy this at scale in enterprises. You spend so much of your time talking to CIOs and technology leaders. What are they telling you? What is changing?

Well, firstly, it is a really exciting time to be alive in tech. If you take a step back and see what is happening, there is a fundamental shift in the patterns of inferencing. In the chatbot era, inferencing was human-led, very spiky. You ask a question, you get an answer. What you are starting to see with agents is they are working 7 by 24, very consumptive on bandwidth and infrastructure. So you are starting to see a much more persistent pattern of demand signal for infrastructure. We usually joke internally: humans click, but agents swarm. So as you start to see this, you will have tremendous demand. There are two big concerns: every customer is concerned about security, and every customer is concerned about cost.

So what you are starting to see is inferencing not just limited to the data center. Inferencing is going to be distributed everywhere. There is a whole new class of computing emerging with this deskside computing that you just announced, where there will be inferencing on laptops or desktops where humans work, and deskside computers where agents go out and run their jobs on behalf of humans. That is what we are seeing.

And we are fully aligned, G2. That is exactly the shift we are seeing. If AI is moving closer to employees, compute has to move closer as well. The PC is no longer just a productivity device; it becomes an intelligence node in the enterprise. But deploying intelligence everywhere creates a new challenge for CIOs that you and I talk about so much. How do you operate thousands of these systems with the security, governance, and control enterprises require? How is AMD and Cisco solving this, and what are we building together?

Yeah, this is an exciting time. There are a few things that need to be done. You cannot just deploy deskside computers and hope everything works out. You need to make sure it is governed and managed. Number one, you have to have appropriate network bandwidth to satiate the needs of the agents you will run locally. So modernizing network infrastructure to keep up with the demand is very important. Number two, every customer is worried about token costs, so we need to contain the costs of an agent's token consumption when it starts to go up and down. Number three, we need to monitor agent behavior for safety and security so that if it does start to do things we do not want, we can provide runtime enforcement guardrails. And fourthly, overall safety and security from an apparatus perspective to get the whole thing humming.

Yes. G2, you and I think very alike. The opportunity is not just to bring AI to every device, but to enable enterprises to deploy, manage, and scale with confidence. Why don't we take a look at what we have been building the past few weeks and months?

Yeah, it is exciting. Basically, what we have done is a full stack. You folks have done with Halo: an isolated secure agent sandbox, intelligent routing, MCP core integrations. Above that, we have made sure you have the right level of security policy enforcement, observability for an end-to-end resilient infrastructure stack. Observability of how the apparatus and infrastructure are working, whether agent behavior is done correctly, and whether the token economics are within guidelines. Then we have a single unified management plane, a control plane that can look at every dimension and be managed within one environment. That is the full stack, and we are really excited about the partnership.

Us too. Our teams love working with your team, G2. Now, how is this all managed? How does the CIO deploy this at scale?

So let's take a look at what this looks like. We have Cisco Cloud Control, the overall management plane. If you have Halo devices on every desktop, you want to manage your entire estate, not just your Halo fleet but also what you have running in data centers and the cloud. You can provide full visibility for the whole estate. Then we also have the notion of monitoring token economics: is the agent behaving the way it should, or is it being overly consumptive? At which point you can isolate any individual Halo device to quarantine it.

I love this. Then we actually see the return on invested capital. It could pay for itself in three or six months, even quicker.

Exactly. So what you see here is you can tell how every single agent is performing and whether they are consuming too much token cost, and how to contain the cost.

No, I love this, G2. Now, everyone else is wondering when they can buy this? When is this open for everyone?

Well, the good news is you can buy the Halo already today. And what we are doing is providing this entire management apparatus that is currently in early availability for a select set of customers, but we will have it in general availability in the US in early fall. So we are excited to provide not just the compute but all of the apparatus needed to manage this, so you can have inference running in the cloud, in your data center privately, or on your desk side.

G2, we are so grateful for the partnership with you and Cisco. We are so excited.

Thank you so much. Congratulations. Take care.

What you just saw brings the vision to scale AI seamlessly and securely from personal devices all the way to the cloud. But in the age of AI, intelligence is where AI leaves the screen, enters the world, and acts. That frontier is physical AI, the ultimate application of agentic frameworks. In a digital world, an agent plans and completes a task. In a physical world, it must also sense, motion, navigate uncertainty, and protect people. Here, intelligence does not just produce an answer; it produces an action. When answers become actions, latency becomes safety, and reliability becomes trust. That is why the promise of physical AI is so incredibly powerful. It is not about replacing human potential; it is about amplifying it. Giving surgeons greater precision, keeping people out of harm's way, making farms more productive, supply chains more resilient, and factories more efficient. The best technology does not make people smaller; it gives them much greater reach. For more than 20 years, AMD has helped power the core capabilities that robotics depends on, from sensing and functional safety to deterministic control and precision motion. That foundation is already trusted by many of the companies defining the future of robotics. Now that foundation is coming alive in extraordinary ways. These are not machines following a script; they are beginning to perceive the world, reason through uncertainty, and act with purpose. The promise of physical AI is already in motion. The next era of robotics is not programmed; it is autonomous. While traditional robots execute instructions, autonomous robots are given an objective and determine how to achieve it, all at once and in real time. The robot must orchestrate intelligence, motion, and safety continuously, and every part must respond at the speed of the world around it. That demands an entirely new kind of robotics brain. To solve that, we rethought its architecture from the ground up. Today, I am so excited to introduce the AMD Korea AI system on module, powered by Ryzen AI Embedded X100. It brings CPU, GPU, NPU, and unified memory together in one compact open-standard architecture. This allows robots to process perception, AI reasoning, and do it all simultaneously in real time. The architecture is different, but the results are decisive. In third-party testing against Nvidia Jenson Thor, Korea delivers 3.4 times better real-time results, 2.3 times more concurrent agents, and 1.6 times more CPU capacity. That means faster reactions and greater headroom for the robot to take on more complex work, all without compromising control or safety. But to define the next era of robotics takes more than a powerful brain. Developers need a complete path from idea to a machine operating in the real world. Today, we are introducing AMD's Korea AI robot developer platform, the industry's first turnkey, open, fully integrated robotic system. Thank you, Laura. At the center is the Korea AI system on module paired with a robotics carrier card designed for sensing and connectivity. Built on ROCm and ROS 2, this platform gives developers everything they need to move from concept to prototype in days. The same module carries forward into production with no migration or redesign needed at the finish line. You can build faster and move from imagination to deployment with absolute confidence. This is the breadth only AMD can bring to robotics. From the first signal to the final motion, AMD Korea for the brain, Versal for the spine, Zynq for the joints, and Spartan for the sensors. Each layer purpose-built for its role, working together in one intelligent system, connecting perception and high-level reasoning to real-time control. The future will not be defined by one model, one machine, or one company. It will be built in the open across silicon, software, systems, and an ecosystem moving forward together from the cloud to the PC and now into the physical world. AMD continues to build a compute foundation for intelligence everywhere, with laser focus on the future, so our solutions can augment and expand what people are capable of achieving. The next era of AI will not just unfold behind a screen; it will unfold in the world around us. Now, to bring it all together and take us to what comes next, please welcome back Lisa to the stage.

Hi Lisa.