Welcome, everyone, to day one of the BFA Tech Conference. I'm Vivek Arya from the Semiconductor Semicap Equipment team. Really great to see you all here and really delighted and honored to have the team from AMD join us this morning, Jean Hu, CFO, and Matt Ramsay, the Head of Investor Relations, and prior from our tribe of proud sell-side semiconductor analysts. Really glad to have both of you here. What I'll do is go through my list of questions, but please feel free to raise your hand if you have any question that you would like to bring up. A really warm welcome to you, Jean, and really glad that you're here with us.
Thank you.
Jean, give us a state of the union, the way you see it, a lot of macro cross-currents, and then when AMD itself has kind of been on this very interesting journey in the road towards AI. What do you find exciting right now? We can go through some of the nitty-gritty of each of the segments.
Yeah, yeah. Vivek, thank you so much for having us. It's great to be here. Thank you all for joining us today. Maybe I'll start to just take a step back to set the stage about AMD. When you think about 2024 for AMD, it was a really transformation year. During the year, we really made significant progress over building our high-performance computing platform, which includes CPUs, GPUs, and embedded processing. Specifically, if you think about the AI market, what we have done is since the launch of MI300, which is literally December 2023, the first year of the launch, we ramped the revenue to exceed $5 billion. We made very significant progress with our hardware roadmap, which is annual cadence, introduced MI325 in December 2024, and also accelerated the MI350. We are actually launching next week.
Hopefully all of you can join in to listen to our MI355 launch on June 12th. Most importantly, we also made significant progress in software with ROCm, really maturing AMD Instinct and AMD software are now powering some of the most complex and the most compelling AI models with our top customers at scale. That is really exciting. On the CPU side, we actually have the best product portfolio with both our server CPU and the client CPU. It is very exciting to build that foundation. Now, when you think about we're looking into 2025, we started 2025 with very strong Q1 financial performance. Revenue was up 36% year-over-year, driven by data center was up 57%, and client on the gaming was up 28%. That is really exciting. The earnings per share, which I really focus on, was up 55%. Exciting time.
When we look at the Q2, that's Vivek, what you're saying is there are a lot of noises and a lot of uncertainties. One of the things that really impact AMD is the export license requirement for MI308. When you think about this, we were expecting the first half of data center revenue to be flattish with the second half of 2024 as we were going through the product transition from MI325 to MI350. Because of the export license requirement, we got impacted by $700 million revenue in Q2 alone with the data center GPU business. It was quite significant. Of course, in China, because of DeepSeek, you did see the surge of the demand for GPUs during that period of time. What we feel really good about is we actually guided Q2 at $7.4 billion.
That's the middle point, which will be like 27% year-over-year increase despite this $700 million revenue impact. It is really because of our core business strengths. What we're saying is continue the momentum with our CPU server business. On the client on the business side, it has been performing really well. The sell-through is really strong. We have a much higher, richer product mix versus traditionally because of the product portfolio. The other thing I'm actually pleased with is our gaming business. We went through this really deep correction. Now the inventory normalized, and we actually see customers on the gaming console side, they actually start to build for the holiday season in Q2 and Q3. Client on the gaming business performed really well.
In Q2, we do expect client on the gaming business to up high teens sequentially, which of course offset the high teens decline on the data center side due to the impact. Overall, when you think about the first half for AMD, we actually expect revenue to be up 30% despite of this export control and the tariff uncertainties year-over-year. Of course, we're driving earnings expansion much faster than revenue growth. What's most exciting is actually second half is we're launching MI350. We're on track with our execution for MI400 generation launch next year. Overall, not only we're adding new customers, but existing customers, we're covering more application models from both inferencing side. You will also see or expect us to run more training models with our customers. Overall, very exciting. We're confident about our CPU platforms.
We do believe the product portfolio innovation and the technology strengths will help us continue to gain share. Overall, I would say 2024 was a year to build the foundation. 2025 really is an inflection for us to drive top-line revenue growth, not only for 2025, but beyond. Again, that earnings expansion. It is an exciting time for AMD.
Excellent. No, that's a great introduction. You took away half of my question, Jean, so I have to find.
Oh, okay. No, you asked more interesting questions.
Yes. No, thank you for doing that. Let's start with the data center business. Help us understand why you thought first half would be flattish and what gives you the confidence in this second half upside and with the launch of the MI350. Do you think it is just a product transition issue? Was there something lacking? What changes in the product? How's just your confidence around getting new customers or expanding your potential with existing customers with the MI series?
Yeah, I'll start at the high level that Matt added. When you think about it, since we launched MI300 in December 2023, the annual cadence really accelerated our roadmap. When you think about each generation of our product, we are making significant improvement. We have the competitive advantage on the inferencing side because of memory capacity, bandwidth. We are continuing to drive that advantage with each generation. MI300 is really the first one. MI325, which is very competitive, but MI350 is where we see the inference performance jump by 35 times. Of course, we can also support the training models. When you look at that progression of the product roadmap, typically you go through a transition. The significant ramp in 2024 with MI300 over $5 billion.
Of course, in the second half of this year, that's where we launched MI350, which we see tremendous customer attractions with, not only adding new customers and existing customers. From overall, when you have a business or market growing so fast and you're pushing out a different generation of product, you do go through that natural transition.
Yeah, Vivek, I think the only thing I would add, Jean said it well, but the things that I would add are we were really excited as a company to be able to pull in the MI350, 335 by about a quarter. Given the capabilities jump, both on a scale-out networking perspective, new data types for FP4 and FP6, new sort of memory configs and bandwidth, compute capabilities of the GPUs themselves. You pull a product in in the roadmap and it does impact some of the purchases of the prior gen as you lead up there and have a transition. As Jean mentioned in her opening comments, Q2 was planned to be a bit more of a China-heavy quarter for us anyway as the Western customers went through that transition. We got some additional DeepSeek demand pull in China.
Due to the export restrictions, we've had to exclude that revenue now. It does make the shape of the year look different and the back half look like a really steep ramp. Essentially, we were executing to what we had laid out to this community on our call in February, maybe a tiny bit better than that. Obviously, we had to react to the export restrictions. The confidence levels, and now that we're getting closer and closer to launching the products, we've sampled MI355 systems to a number of customers and gotten their feedback on it. I think we feel pretty good about where we are.
Got it. What is the most frequent workloads that AMD product is used for? Because there is always the question that you have the incumbent, they have all their software and developers and whatnot. There is a natural case for them in a lot of public clouds. Then you have the custom chips on the other side. There is a case to be made for them in a lot of internal workloads. What is AMD's workload where you say, this is where we lead today or this is where we hope to lead going forward?
Yeah, I would say a very, very first thing is inferencing. If you look at AMD from MI300, we do have the advantage on the inferencing side. If you look at the first years of ramp, we are powering some of the most complicated models like ChatGPT 4 and Microsoft Copilot. On the Llama side, the Meta side, the recommendation engine, and the different things. Priority-wise for 2024, we have been really focusing on those very complex models to support the customer into production and cover more and more models and applications. That is really for us from inferencing side, it's the key advantage. Of course, we are increasingly covering the training side, covering different things. I think when you look at the ASIC versus merchant GPU, our view has always been different models need different compute engine.
You definitely can see if the model is very much fixed and not changing, ASIC probably is the most cost-effective solution. When you look at the gen AI market, the model innovation, just the pace is incredible. Everything continues to change. We do think general purpose GPU will continue to be the majority of the market. That's basically how we think about it. What we're improving each generation is not only keep the competitive advantage on the inferencing side, but like Matt said, we're adding more networking capability. We acquired ZT, which will add a system solution and support the rack level build-up clusters in 2026.
I see.
Yeah, Vivek, the only thing I would add to that is we all have to think about that this is a transformative computer science that's, what, 30 months old. We see DeepSeek was one example, and it got a lot of attention because it was in China and some claims that they made on cost. What we see is a sort of a proliferation of folks that are like DeepSeek as the install base of GPUs. Some from our competitors, some from us get larger and larger and larger. You're going to see more folks doing software innovation on the install base. It is certainly where algorithms that settle down, it certainly makes tons of sense to do an ASIC. We continue to think ASICs will be an important part of the market. We've actually done some in certain instances for customers.
In a time when the algorithms are moving so quickly, another way to generate leading TCO for your customer is to take advantage of the industry software innovations over the whole depreciable life of the product. I think that we and our large GPU competitor probably agree on that philosophically around the programmable nature of these systems and how early we are in the software innovation of the models.
Yeah, and I would say we also made tremendous progress on the software side. You should see or expect to see more of our solutions with the Tier 2 CSPs because we have been prioritizing our top customers. In 2025, our major focus is to broaden the customer engagement so enterprise customers can actually go to different cloud and Tier 2 CSP, get the AMD MI300, 325, and 350 to run their application models. You will see more and more of that.
Got it. Just one last one on kind of a near to medium term. Just because of the cycle time of getting a lot of these inputs done, how would you characterize the visibility, Jean? Do you think a lot of the decisions about the second half have already been made? There should not be a, what are the upside drivers or downside risks as you look towards the second half for MI?
Yeah, thanks for the question. I think when we engage our customers, it's always multi-generational. You're absolutely right. The lead time, the cycle time for complicated GPU solutions are really long. Typically, you really need to plan ahead with your supply chain and the capacity continues to be tied with the CoWoS and HBM memory. You absolutely have to get your allocation for the year. From customer engagement side, it's the same because not only you need to resample the hardware, but you also need to make sure your software really works and to get hardware to the performance level customers need, the TCO they will need. We have been working with our customers for a long time. It's not just recently. We do feel quite good about the customer feedback and the customer engagement and the customer orders for the second half.
In the end, we think the market opportunity actually is tremendous. It's really about how we can help a customer ramp on the software side to support their workload.
Got it. How would you gauge AMD's progress so far, Jean, when it comes to things like software, the majority of the ROCm stack, and then also networking? Because some of your peers have a lot of their internal solutions when it comes to scale-up proprietary or Ethernet-based, or they have Ethernet switches and so forth. How would you kind of assess where AMD is from that perspective?
Matt, you want to software is your area of focus.
Yeah, Vivek, I think there's a couple of things. You mentioned and you've seen this as AI has proliferated, our need to go to system-level solutions. That's inclusive of CPU, GPU, networking, software, system-level design, cooling, power supplies. There's a lot to it. We needed to continue to invest more in the ROCm software stack. I think we've done that. There was a focus in 2024 on the largest customers. Now you've seen us start to broaden out the teams doing biweekly ROCm releases. The goal is over time to be the software stack of choice for open-source deployment of AI. That's not going to happen in a week or a month.
You saw with what AMD did from 2016 to 2019 in the server space, where it was a focus on the largest of customers first and then a broadening out of the strategy as we gain scale. I think you'll see that be something that we continue to focus on in the software side with the ROCm stack. The team, the goal is really how many models can you support that are performant on day zero? The ones that aren't performant on day zero, how much can you shrink the time and the friction to go from a competitor solution to an AMD solution and get that model performant in not months, not weeks, but days? I think for a lot of things, we're making really good progress on that.
We have also spent a little bit more time and resource maybe in the long tail, not yet a full support for the long tail, but on educating some of the folks in that community that might be influential in this and in the news flow as to what our progress is, how they should be measuring our progress in software, and have them know what the strategy is versus them just maybe picking up a GPU and trying to run stuff without any engagement or support from AMD and then that experience being a little different than they expected. I think we have made a lot of progress, but there is a lot of work to go. We have done things like Silo AI, where we brought in an acquihire that has some stuff, vertical-specific software for certain industries. I think you will hear more about that.
It's a bit of a weird time to have this conversation with our launch event next Thursday. Anyway, stay tuned.
Okay. As you move to rack scale next year, what is the content lift that AMD can get? Or is it that a lot of the incremental content comes from partners and others? Just is there a simple way, Jean, to look at what kind of content lift you can get as you migrate from more chips and boards towards more rack scale systems?
Yeah, I think AMD's approach has always been embracing the ecosystem. When you think about it, we do have our GPUs. Also, on the networking side, we acquired a company called Pensando. They have the programmable DPUs, which is a great asset for us from the networking side. With the ZT acquisition, we have 1,200 very talented engineers. They are like power management, like thermal system solution engineers. They will help us to design, reference design, the cluster and system-level design. We work with our partners. On the networking side, we work with Arista, Cisco, Broadcom, and also we work with other suppliers. For us, it is about to provide our customers the system solution, but we get benefit by selling GPUs, CPUs, and DPUs. Our customers also get benefit by selling their networking gears.
I would say each generation, the content is increasing, especially for us. If you look at the MI350, we're still just selling GPUs. When we get to MI400, not only GPUs and DPUs and also the higher node of CPUs. Our content will increase very, very significantly. At the cluster level, if we can support really very, very large clusters, that helps us tremendously. Our partners will get benefit by selling Ethernet switches or other things. Yeah.
Significant as in 30%, 40%, 50%?
No, we have not quantified any of that yet.
Yeah, Vivek, the only other thing I would add is the ZT deal that we did both ends of it. We brought the ZT design team in-house and then recently announced the planned investiture of the manufacturing co. That whole ZT design team under consultant contracts was working alongside the AMD design teams on the MI400 series the whole time. The deal was pending. Now that they're part of the company, a unique focus for them, having been part of a manufacturing company, is design for manufacturability. When they were part of ZT as an integrated company, if they did not design systems that could then be manufactured and sold, they had no revenue.
We were really, I think, pleased with some expertise that they brought into the company, not just on system-level design, but system-level design with an emphasis on manufacturability and time to market. I think you'll hear the team talk a bit more about that next week.
Got it. One of the questions, we'll come back to the data center side, but one question that has come up, Jean, as you well know on the client side is just given all the trade and tariff issues, the risk of pull-ins and so forth. And AMD has had such significant strength in your PC business. How are you assessing the risk? I know you've already guided to kind of subseasonal second half, but was that more out of conservatism, or is that something that you're truly seeing?
Yeah, thanks for the question. We absolutely have seen our client business performing extremely well, but we have not seen any meaningful pull-ins and pull-forward from our customers. When we really look at our client performance, for instance, in Q1, client actually PC business is up 68%. But 43% is because of ASP increase. And that is not like to like ASP. It's more AMD has been introducing the latest generation product. So the mix is at the high end of the stack we play. That has been the key driver of our Q1 performance. Frankly, the client discussion has been ongoing for several quarters because performance has been really good. If I look at the trend, what happened in Q1 actually was quite consistent with the prior quarters. It is the ASP increase. It is the richer product mix.
Q2, we actually continue to see the momentum, especially right now where in the middle of a quarter, we have not seen a slowdown of the sell-throughs. It is very good in Q2. We do think sequentially it will be better than seasonality. For the second half, the micro uncertainties just are a lot. We cannot predict what's going to happen. Every day there's some news. From that perspective, we're trying to be really mindful of the micro environment and make sure we are conservative thinking about the second half. We did mention it's sub-seasonal, but it's really driven by the uncertainties in the micro environment and the tariffs.
Right.
Yeah.
Yeah, Vivek, I think if the client market from a unit perspective performed "normally," I think we would have a really good back half of the year in our client business. Given the uncertainties that are out there, I think it's just prudent at this time, given how strong the business was in the first half of the year, to take that approach. If we do better than that, great. I think for now, the business, we have seen a few pull-forwards, as Jean said, but we've been working really hard to manage around that. Maybe a data point that will be helpful, in the first quarter, our units were actually down more than seasonal. The sell-through was more than selling across our client business.
We have been trying to manage the business as best we can to sort of protect ourselves from the trends that are happening. There is a lot of uncertainty. I think being prudent about it right now is probably the best approach.
Right. You would not be surprised to see some semblance of seasonality if macro does not change from where we are today?
Yeah, we'll be happy to see that.
Apply your large grains of salt, Vivek.
Okay, understood. One other thing that also came up in your rivals' call on the x86 CPU side, they said that some customers are going down the stack to kind of address some of the tariff issues. I know the tariff situation keeps on changing and that the mix of AI PCs has come down. Did you also see that? Or how do you see that field of AI PCs? Like you mentioned ASP strength. What is driving that richer mix than ASP strength for AMD?
We have not seen that. We ask our team, sales team, we have not seen that. I think the strength of our ASPs has been because the latest generation of product we have, not only we have the best GPUs for the gaming players, we also have best CPU and APU on the AI capability side. Overall, if you look at this, we're addressing the high end of each stack versus in the past, AMD tend to stay at the lower side, consumer side. More importantly, on the commercial PC side, we were very, very much underrepresented. We are still underrepresented, but with Dell engagement, that really helped us with more OEMs and more platforms in the market. We're focusing on enterprise go-to-market. With enterprise customer adoption, that also helps. The commercial PC side, it tends to have a higher ASP.
I think our strategy in PC market is first is to lead with the technology and the product portfolio. That has been the key driver Lisa and the team is focusing on that really helped us to not from market share side, we're still a small player, but for us, we really want to be profitable growth, not only just growing revenue, but get profitability there.
Yes. The other piece, Vivek, is in the enthusiast desktop market, both for developers and for gamers. I think for this audience, go find whatever Amazon or whatever e-tailer site you'd like to go find and figure out where AMD SKUs are in the top 10 on those. I think you'll be pleased with the results. It's been, over the last six, seven quarters, a significant share shift in the highest end of the desktop market. The sell-through has been very strong relative to the sell-in as well. That's a new ASP-rich part of the TAM that AMD had not traditionally won that much business in. The market share sort of flipped on its head in that piece. We feel pretty strongly that we can sustain that if we continue to innovate on those products.
Got it. In the last minute or so, I just wanted to quickly touch on gross margins since we have our CFO here. Gross margins, the one pushback we hear about is that, well, AMD's growth will be driven by a lot of these GPUs. And so far, they have been kind of below corporate average. As that mix shifts more to these GPUs, and then within that, the mix shifts more to rack scale systems, right? How do you look at the trajectory of gross margins from here, Jean? What are the upside and downside risks?
Yeah, I think AMD's business, there are always a few puts and takes because the mix actually really drives the gross margin. Overall, when you think about the second half, it is true we will see data center to be the number one revenue driver for the second half. Of course, GPU is an important element of that, which is diluted to corporate gross margin. We do have our server business enterprise, both on the server side and the commercial side, will actually help to drive the gross margin up. We do think we will see modest gross margin improvement in second half. Of course, it depends on mix. Going forward into next year and beyond, I'm actually quite confident about continued gross margin expansion. Despite of data center GPU side, it's slightly dilutive or diluted to gross margin.
The overall enterprise play, both on the server side and the commercial side, more importantly, we did not touch embedded business. Actually, we do see really important demand signals that the cyclical upturn is coming. We do think next year, the embedded business will also be much stronger versus this year. That is a really high gross margin business for us that will help us with the mix also.
Right. Terrific. With that, Jean, Matt, thank you so much. Really appreciate the discussion. Thank you all.
Thank you, everybody.
Thank you so much.