Advanced Micro Devices, Inc. (AMD) Earnings Call Transcript & Summary

Bob, maybe to make your life easier, you can go with Ben, and we can move to Simon and Aaron since they're all sitting next to each other.

Ben Reitzes with Melius. It's great to be here. This is probably for you, Vamsi. ROCm AI, the software, how are you looking at that in terms of disrupting or impacting the ability to run apps versus CUDA? Do you think it's revolutionary? How should we think about, is that a game changer? Is it evolutionary? And does that help level the playing field, do you think, with developers?

It's a great question. And obviously, I have enormous passion for this, right? We truly believe it's the biggest leap that we have made maybe since the early days where we laid out our strategy. We've made excellent progress every year. But if you have to point to is there a moment in time where we say, okay, this is actually the biggest leap and spring forward, I would point to this time. Now I don't mean that by August 14, everything is different, right? But the inflection that's happening now, what is likely to happen that we build over the next many months, together with the biggest labs, right, our collaborations with OpenAI where CODEC's better, our collaboration with Anthropic where Claude gets better, it's going to be a significant differentiation in terms of like how easily accessible the platform would be relative to any time in the past. So over the coming months, you can expect the productivity of people to access platforms to be quite different.

Simon Leopold with Raymond James. When we think about the new TAM outlook, I wanted to see how you're thinking about the biggest risk to that, in particular, the ability of your customers to get power to their data centers or your ability to get manufacturing capacity, wafers, et cetera, how is that factored into your view on that? And what's your consideration?

Sure. So when we think about TAM and especially with the accelerator TAM being as large as it is, I think we look at all of those components. So not just broad demand, but we also look at what is the rate and pace that power is coming up, as well as what is the rate and pace that our suppliers are adding capacity. So from that standpoint, I think in the near term, I think we have very much planned sort of capacity for significant growth in '27 as well as '28. And then over the longer term, as you're thinking about '29 and '30, I think it's a rate and pace of growth that would require the entire ecosystem to be building at the same pace and have the same vision. Probably the largest change that we've seen is everyone has been thinking about the Accelerator TAM growing very fast, so that has sort of been in the numbers. The fact that the CPU TAM has accelerated as much as it has, has required some adjustments to overall capacity. But we are very happy with the supply chain relationships that we have, and we do see significantly more capacity coming online to satisfy those larger TAMs.

Aaron, go ahead.

Aaron Rakers at Wells Fargo, and congrats on all the announcements today. I guess I want to build on that question. Maybe it's not the supply chain. It's your ability to actually stand up these massive rack configuration. So Lisa or for anybody, if you were to conceptualize like you've got 6 gigawatts here signed up for Meta, OpenAI, 2 gigawatts of Anthropic, how do we think about the pace of your ability to ramp from a gigawatt per quarter basis? How do you -- how quickly can you stand up that much capacity? Any kind of color would be helpful.

Well, it's a great question. So if we start answering it from sort of the rack level on up, because I think Lisa has already addressed it from the rest of the supply chain. First off, we're working very, very closely with our key OEM and ODM parts. So Sanmina, [ WWAN ], et cetera, as well as the OEMs, to ensure that we've got the manufacturing capacity in place to build the racks, to build, integrate, test and validate the racks at the right pace. And that is an important part of it because the better you can get at that, the easier it is to actually support the deployment in the data centers. Shipping a very high-quality rack is an important part of making sure that you can turn them on very quickly in the data center. Beyond that, we see the next chokepoint is in actually deploying both logical as well as -- physical as well as logical deployment of the racks. And that's something that we're working again very closely with our manufacturing and OEM partners. One of the things that we acquired as part of the [ ZT ] acquisition as we acquired a large services arm which we have retained as part of AMD. And we are actually using that team right now and not just to support some legacy customers, but also do all of our internal deployments within AMD and then to help our customers deploy very rapidly both MI 350s as well as MI-455 systems in their data center. So with that set of capabilities and training our partners, we're pretty confident that we'll be able to stand up to build at the pace required and then to stand up and provision and get turned on the systems in the customers' data centers.

And maybe the only thing I would add to that is now when you talk about overall data center power, we're also very active in that process with our customers so that as they're planning power, we're planning the GPUs and the Helios systems that go along with that. So it's much, much more involved than it was in the past where somebody just places an order. I think there is easily 12 to 18 months [ possibility ].

Maybe we'll go to Srini here. And if I missed people out there, the quality of these spotlights is spectacular. So I'm not doing it essentially...

They are kind of on the bright side.

Srini from RBC. Lisa, I had a question on your road map that you talked about, in particular, the scale of networking. I think you mentioned optical and copper with MI 500. I'm just curious if you think the market and the ecosystem is ready for optical or will be ready for optical next year. And if so, do you have all the pieces of the puzzle to be able to support that? And also, as part of that, I saw [ E.SUN ] highlighted a bit more than ULN. I just want to hear your thoughts on which, I guess, scale-up you will be supporting going forward.

Yes. You want to start that up?

Yes, let me start. So first off, I'll take both of those pieces in order. So we do see the MI 500 generation is the one where we begin transitioning from optical -- I'm sorry, a purely electrical interconnect for scale-up networking to start to see ethnical play a role as well. It's going to be a transition. We don't view this as a light switch. We don't view this as, hey, we're going to hit a generation whatever MI 500 and MI 600 and everything is going to flip to optical. Instead, we see MI 500 starting the transition. We're working very closely with a number of partners across the ecosystem as well as we've been investing in optics for quite some time. And we're highly confident of our ability to begin that transition, that the terminus of that, generations out in the future, is co-packaged optics on all the major components and optical really being the backbone of many connections within the rack as well as between. But that's going to take a little bit of time to get there. And we think, again, doing it in this phased approach allows us, our customers and our suppliers -- our partners in supply chain all to gain experience and to make sure that we're moving at the appropriate pace and not taking any operational disruptions. On the scale-up protocol, look, we -- on MI 450, we support UA Link transported over Ethernet. [ E.SUN ] is a set of extensions to Ethernet, which is helpful with that, and it actually is going to continue to evolve. And so we do expect to see that protocol brought forward and again, UA Link over Ethernet brought forward and being available in MI-500 as well. But that's not the only thing we're doing. So we'll unpack more around scale-up as we get closer to the MI 500 time frame.

Atif, maybe you want to -- Lisa, did you want to expand on that at all? I kind of jumped the gun there. So Atif, go ahead.

Atif Malik, Citigroup. I have a question on the MI 500 ramp. HPM content is a very important part of your performance in token economics. And a couple of your peers have cut their content for HPM memory in the future because of the availability of the memory. And my question is if your thinking has changed or evolved maybe in the last 6 months or so on how you're thinking about the content increase for MI 500.

I think -- so we obviously study the workload characteristics and how capacity impacts, right? The first order, right, we separate our bandwidth and capacity. Bandwidth has a tendency to lift more boats in terms of more workloads directly getting impacted. So that's one order of consideration, and then you look at capacity after that. One advantage is because of the way we have our chiplet architecture, it actually gives us more flexibility and options in terms of how we are able to optimize capacity while preserving [indiscernible] bandwidth constraints. So that's the uniqueness of our architecture. We've leveraged that with our existing products, and we do expect to leverage that with future products as well. We're not sharing the exact configurations of what 500 would have or the road map. But that's one unique piece that actually we believe will play to our advantage.

Maybe if I just add to that. I think the way to think about it is we absolutely think our chiplet architecture gives us the ability to be very flexible in terms of memory bandwidth, as Vamsi mentioned, but memory capacity is useful. I mean our customers have told us, the fact that we have more memory on MI 450 is one of the reasons that we're getting better inferencing performance. I think the key is, as we're going forward, and all of our customers are doing this, I mean this is an ecosystem discussion, that we need to make sure that the memory that is there is really being used because it is such a larger piece of the TCO. And so we are doing some memory optimization along the way. And that's true on both sort of the CPU systems as well as the integrated Helios types. But memory is definitely super important. We will just make sure that every amount of memory that we're using is valued by the customer appropriately.

Blayne, you want to go ahead? I think Liz is on the other side there.

Blayne Curtis, Jefferies. I just want to expand on Ben's question on ROCm AI. So just kind of curious where you are on this AI journey internal use of AI. I know Jensen [indiscernible] like half a person salary, which is a big number. But I was just kind of curious, are you tracking that? And if you could talk about where you are in terms of like day 0 support and automating that with AI and then where else you're using it?

Yes. I'll comment specifically on ROCm AI and then maybe there's also a broader sort of corporate usage comment in here. So as far as ROCm AI goes, there's actually both internal acceleration of existing features and capabilities, but then externally, what we can put in the hands of developers that come with the platform. So what I mean by that is, imagine you have a profiler or a debugger feature that needs to be built, our engineers in the past, you say, okay, this is going to be a team of 20 people, 6 months, right? And now that's actually dramatically cut down because those profilers, debugger features can get out much faster because of the ability of [ AI]. And that all comes part of the ROCm accelerated release. And the piece where it actually helps significantly from an external perspective is the platform now becomes native in terms of AI agents being able to access it, and that's what we are going to start shipping starting August, both from just general out-of-the-box usability, but performance optimization and running these models [indiscernible] much easier. Almost everybody on the ROCm team, they're all AI-native more or less because of the group they're in, are pretty much using AI assist to be able to accelerate that plan. So just give you a sense for like how fast or how extensive that is going within AMD.

And maybe to the broader point, we are seeing AI usage ramp up across AMD extremely quickly. I would say, every single month, we're seeing token usage, the amount of -- it's not just the number of tokens, but it's the quality of what we're able to get from AI. Dan mentioned what we're doing in terms of serving different models across our [indiscernible] stack. So I would say it is very much a part of our development process across hardware and software, and I see it continuing to ramp and these relationships that we are -- these deep relationships with Anthropic and OpenAI, as well as a number of the other model companies, are helping accelerate that rate and pace.

I think there's a question over here, Bob, to your left.

This is Bhavtosh, CLSA. Lisa, you started your presentation with this 35 quadrillion tokens number, which is already all over the media because it's a shockingly high number for today's environment. My question is, how is AMD projecting demand beyond conversations with your partners in the ecosystem? Do you have a fundamental way of thinking about where token use will go given current cost of compute? A lot of investors worry about the cyclicality of this industry, and that's where this question is coming from.

Yes. No, look, I think the way we project demand is really quite holistically. So we start with customers, we look at workloads, we look at adoption rates. We certainly look at the free cash flow of our customers to make sure there is the capital behind it. But when you put that all together, every projection that we've put out seems like it was really high, and then the market has actually gone faster. So we continue to see just very significant demand across virtually every part of the portfolio. And I think that gives us a lot of near-term confidence in these higher market projections. Now that being said, I mean, we have to see how things develop over time. So I wouldn't say that our crystal ball is perfect. But I can say that it's self-consistent. So it's self-consistent that assumes all of the aspects of is power available, is supply available, is capital available and is productivity going to be able to close that loop when we're looking at them.

All right, folks, I think I'm going to try my best here to wrap the session up and keep my executives here on time because they have a lot of other commitments. Thank you very much for coming out. Lisa, it is, and the whole team, it was a great day and a great conference. I think it's really, really exciting to be in a place where there's so much diverse demand for high-performance computing across what's now, what, a $2 trillion TAM. So Lisa, if you have any closing remarks, I think we'll close the session, if you do.

Yes. No, I'll just say thank you for spending the time with us. It's been a really exciting day. It's a culmination of a lot of work from across the company. What I would like to say is we really think about AI as a complete compute picture. So we talk a lot about CPU TAMs, GPU TAMs, Helios systems, all of that. But we really think about AI as every aspect of compute. And this is a place where we can be quite differentiated in the end-to-end story. So hopefully, you heard a little bit of the comments from Germany at AT&T, the work that we're doing with Cisco, the work that we're doing in physical AI. This is like we're on this 5-year super cycle of just tremendous compute demand and having great partners to work on to unlock all that. So thanks again. We will talk to you soon.

Thanks.