Marvell Technology, Inc. (MRVL) Earnings Call Transcript & Summary

All right. Hope everybody is doing well. This is -- we've saved one of the most interesting sessions for last of the day. And for those of you that have been with me most -- I see some familiar faces that have been with me most of the day here in this room that I haven't left all day. So my name is Matt Ramsay. I lead a semis research franchise at TD Cowen. Thank you, everybody, for participating. And I hope it's been a productive couple of days at the conference. This is our 51st Annual Tech Conference. It's the first one we've done since the acquisition of Cowen by TD. So we're -- in addition to rebranding everything green, we're getting familiar with all of our new colleagues, and it's a pretty exciting opportunity for us. So we were just talking. It's been one of the most interesting 4- or 5-day periods in the semis market in terms of the stock. So it's really exciting to have Willem and Ashish from Marvell here to have a conversation. These guys reported earnings last Thursday. As you may have noticed, the stock was up 30-odd percent in a couple of days post that, so it's been pretty lively. But thank you guys for spending a day of your time with our team and our clients. And Willem, maybe you could give a brief recap of the events of the last 4 or 5 days and how the business stands, and then we've got tons of questions.

Yes. Great. Thanks, Matt. Thanks for having us. So maybe to start, just -- I'll just give a little bit of my own background. As you guys know, I took over as CFO about a quarter ago. But I joined Marvell at the same time as Matt about 7 years ago now and have really been part of this whole transformation of sort of old Marvell, really focusing all of our investments in data infrastructure, all the M&A, all the integration that we did. And so we've had this long-term view on that data infrastructure is an amazing opportunity. And I think the sort of events in the last couple of days is just a perfect sort of planning out of that thesis, where, obviously, AI is sort of the killer data infrastructure in application. And so yes, it's just super exciting couple of days for us.

Yes. I think maybe just before we dive into -- and you've checked the box by saying AI.

Yes. Yes. There you go.

So we got that out of the way in case if there's any worries. Just -- I mean we both know Jean well. All of us do. And as you've taken over as CFO, what did you want -- what impression did you want to make? And what changes did you want to make? What's -- I'm going through this with some of my research associates that have been promoted and got to be first author on some research reports and see what that fun is like. And so just what's been -- as you expected, what's maybe been different than you've expected? And what real imprint do you want to make on the role?

Yes. Thanks, Matt. So first of all, I was really part of a team that sort of built all these systems and processes and how we do it and how we do our capital allocation, how we budget. And really, what's changed in the last couple of quarters is that the macro, we've seen a very significant imagery correction, right? And then as we've seen that, we've seen the sort of inflection point on demand for -- and I am going to say AI a couple of times here, but specifically for our optical products, right? And so as we look through this inflection, we're having to move really quickly to make sure that from a capital allocation standpoint that we invest in those areas where we see the highest ROI, and then when there's other opportunities where things maybe have pushed out and take the opportunity to sort of lower our investment there. And then also, when you look back at going through the supply crunch, obviously, the focus there was to secure supply, right? And so we invest a lot in the supply chain, built up inventory, really -- our working capital really increased significantly. Now I don't think we go back to where we were pre-pandemic, but as a strategic focus for me, I'm very focused on free cash flow and how we manage our working capital, and then getting back to -- and we'll talk about our capital allocation, and I'm sure at some point, but really being able to restart our buyback program.

Guys, things have gone so crazy around AI since the results of NVIDIA. I think that was one of the more covered reports that any of us have seen in our time covering tech, but it was a big focus on your earnings call last Thursday. And in the days since, I'm sure you guys have gotten feedback as I've gotten feedback. The stock reaction was super positive to the results. And then there's been, I'd say, some healthy debate and pushback as to the sort of categorizing of revenue associated with AI, and you've outlined some of those. So maybe you could walk really quickly through the things that you -- the categorizations that you did and the growth rates that you announced on the earnings call. But what I really want to dig into is exact revenue is being categorized under AI? And how much is it tied to these AI servers? Is it infrastructure that just goes around general purpose servers that happens to be tied to AI? If you could just walk through, I'm sure there were -- I mean as a big company, there's lots of meetings to go through and categorizing these things. What were the debates internally? And how did you come to land where you landed?

Yes. Good question. I'll kick it off and Ashish, you can jump in.

Sure.

So first of all, when we looked at it, we said, "Hey, we have to have a very robust process." We're not just going to go estimate and sort of -- so what we did is really have a systematic process where we went and for each customer card combination looked at. And our definition was, does it go into a cloud AI data center? So not is there some AI IP on the chip? Or is it somehow tied loosely to it? Is it specifically going into the AI application in the cloud? And so then when we started going through it, there's a couple of things that's very clear. So first of all, on our 800-gig PAM product, when we looked at that for last year, that's really not being rolled out broadly in the fabric. And so when we looked at that and who the end customers were, almost all of that is going into AI clusters. And that's where we're seeing ramping this year, just can talk a little bit about the actual dollars, but that's the main driver. And then in addition to that, we have this data center interconnect business, the 400ZR. And that -- we've been shipping it to the run rate, and it's mainly 1 customer. It's really 1 customer today. And as we sort of got into this year, suddenly, that inflected up very significantly. And so we had this conversation with that end customers, what's driving it. And so it was very clear that sort of that step-up in demand. And just to be clear, we took a portion of that 400ZR. It's not all of it that's being tagged to AI. It's really the portion that's getting driven by inferencing demand at the edge and the additional bandwidth that's needed to connect those edge data centers. Maybe you want to...

Yes. I mean, just at a high level, Matt, the reason we did it was no different than what we did a few years back when we started, "hey, we're going to be big in 5G." And the first question for investors is, "how much do you have and how big can it be, right?" So it was a very similar thought process, which was the biggest question I got last quarter after I got through gross margin explains was how big is AI? How much are you writing today? Where do you think it can go? And it was a great question. And quite frankly, we don't have a -- we were still in the early part of that process. So just like we did for 5G a number of years back, we went on, mapped out every customers in that particular case base station architecture and figure out the attach rates for our products and said, "okay, here's what the math looks like." It's the exact same process we did this time. We went and looked at every single architecture public as well as what we knew from our own sources, mapped out the attach rates, what are they using today? How many GPUs, other custom ASICs? What the attach it looks like? Where is it going from today? And that's the basis of how we came up with not just the $200 million in revenue last year, which we said is going to at least double this year. And we chose those words very carefully because we don't want to, quite frankly, cap the upside. And we said, "just looking at next year, which is still obviously very far away, we see the path for not just our optics products next year, right? That's most of our revenue this year, but we also see our custom silicon kind of layering on top of that." We did not include areas like storage today, even though we know at some point, those will be beneficiaries. There's things like active electrical cables. There's retimers, gearbox, a number of other components, which will benefit. But because those can be a little fuzzy upfront, we said, "Hey, we're just not going to count those right now." So I would say the numbers we provided at a pretty conservative estimate today. This is the way I would look at it.

The interesting bid to me is when Willem and I talked about this the other night right after the earnings, there's 2 coincidental $400 million numbers, which is both good and bad, I guess. So one of them has to do with the AI products this year, but being roughly $400 million at least, and then potentially doubling again next year. The other one was custom compute products going into hyperscale that were going to originally $400 million this year and got pushed out a couple of quarters. The question I asked you, and we only had a few minutes the other night to try to dig into this stuff is, what's the overlap of the AI bucket and the custom compute bucket? Are these 2 -- I mean, what's that Venn diagram look like?

Sure. Yes. I think the best way for us to start just explaining that is we went back to when we sized this originally at our last investor day when we sort of said $400 million going to $800 million in the '25, '26 time frame. And we said, okay, how much -- because all these design wins that's driving the upside today, those are design wins that we had already won back then. And so when we looked at it at that time, and we said, "okay, what's the lifetime revenue of these design wins that's related to AI," and at that time, it was about 20% of the total lifetime revenue that was AI. Fast forward to today and with sort of increased visibility, we're much closer to actually getting those into production and have better visibility to the end demand and obviously, what's happened in the last couple of months here, and it's increased significantly where it's actually over 50% of the total lifetime revenue. And so that's sort of how I would characterize it. So the $400 million cloud optimized going to $800, million, we think that $800 million is significantly higher. We're not putting a number on it today. Frankly, we just -- we're not sure, but it's significantly higher because of the upside in AI.

Yes. And maybe on the kind of Venn diagram, the simple way to think about it is today, when we doubled from last year to this year, it's all optics, okay? So there's no overlap. Next year, when we had another, call it, minimum $400 million additional AI revenue, some of that's going to be optics, some of that's going to be cloud optimized, right? So that's kind of the way you can think about it.

Right. So no overlap this year, some overlap in Venn diagram.

That's correct. That's correct.

That's helpful. It's interesting when folks talk about -- normally, you talk about an end market. So you're talking about cloud optimized or wireless or whatever end market, you're talking about a group of customers in an end market. As we found out last Wednesday night, the things that are being deployed for AI servers are essentially 1 customer or things that go around 1 customer for now. Can you talk a little bit about what -- around an H100 either server blade, server rack and then cluster that the hyperscale companies are building out with NVIDIA, what is -- what components are actually in there? What's the content? If you can just kind of go through that, I think that would be helpful to have people understand what's going to be doing this ramp in the next couple of years.

Sure. Yes. I guess first is, I think we're seeing activity at more than 1 place, right? So there's also customers which are ramping up their internally ASIC-based accelerator programs. And the building block here is not really a server, not the traditional way you think about it, which is core built around basically a dual CPU stack. These are basically -- think of these as kind of AI building blocks, which are typically multiple accelerators, which do the majority of the compute function. They typically connect to each other with some kind of a local switch, typically running on a corporate trace and that's not where we play. But if you want to deploy these in a traditional kind of a large cluster, you think about thousands of these mini building blocks connected to each other. And that connectivity is essentially gone all optical given the massive speeds, given the distances, the latency requirements and power. And we started to see this last year, and we talked about it a couple of quarters back. Our 800-gig products really started to ramp in very heavy volume last year, and it was only because of AI because you did not see the same thing happened on the switch side, right? So the first big advantage as you go more towards the spending is that the attach rate of optics to an AI cluster is significantly higher. As an example, we said on the call for 1 popular application where it's based around basically 8 accelerators in 1 building block, that particular box has 400-, 800-gig connections coming out of it. And each connection has a DSP in each end, by the way. So a simple way to think about it is like a one-to-one attach rate. If you compare that to a server application, right, where you've got 1 server putting out 100 to 200 gigs, there's no optical requirement. It's a direct attached copper cable. You only go optical somewhere up on the stack where the -- so think of that as a fractional amount of optics per server versus in the [ AI applic ] cluster, it's exactly the opposite. It's a much, much higher attach rate. So that's one of the big advantages we see. And the second thing which we also talked about is as you start deploying these clusters across your geographic locations and especially as you drive inferencing across applications, you're connecting these data centers together, typically using some kind of a direct connection, which is called a data center interconnect. It's our 400ZR product, where we've seen, again, a very significant growth in demand starting about 1 or 2 quarters back.

Got it. No, that's helpful. And I guess circling around these topics of the cloud optimized, the 400 and 800 that got pushed a little bit in time, but now the portion that's $800 million is potentially larger, I'm wondering the basis for that being larger. Is this in reaction to what's happened in the last 4, 5 months with generative AI and the builds that have happened? Were customers coming to you for new programs and then you've added those in? That doesn't necessarily seem like enough development time. Is the specific programs and customers coming back to you and saying, "Hey, you need to go talk to TSM and reserve foundry capacity because the orders are going to be bigger than we told you initially," and did they quantify that? Just the nature of -- what were the inputs to saying 800 can be bigger than....

It looks like you sat in one of our production meetings, clearly. So yes, that's what happened essentially. These are the same existing programs because remember, these chips start to develop in 2-plus years back, right? One of them has finished development. We talked about it as multiple programs here. One of them finished the development very soon, tapes out and start sampling. So the exact same programs are upsizing in the revenue forecast starting as early as next year, right? And the lifetime revenue expectations our customers had given us are significantly higher. And yes, all of that has really happened in, call it, the last 90 days.

I mean if you look at the optical upside we're seeing -- we've already basically booked the orders for this year. And a lot of it is actually within lead time. And so again, there's not a capacity constraint, but if you're ordering within lead time, there's only so much that you can do to actually get that capacity out, right?

No, it totally makes sense. I've been listening to myself talk for 2 days here. So if anyone in the audience has any questions, just wave your arms around and...

Just on the topic of custom silicon given the upside that you're seeing come early, and I think what you're seeing in maybe the demand and even Intel's talked about demand for Gaudi, are you still anticipating only $200 million for that [ closed-custom silicon ] program in this year or have those expectations?

Yes. So these AI programs are really going into volume production next year, so our expectations for this year, which is really think of this mostly as non-AI applications. It's still in the same zip code. Those are still ramping. We still expect that to get to 200, perhaps a little bit better as we get towards the end of the year. And then the AI programs, we're going to start ramping into volume production essentially next fiscal year.

So if I could characterize it, some of the stuff that got moved around in order to -- what we see now is a massive [ quality ] for NVIDIA were non-AI and the AI stuff is on track and was always going to be a 2024 [indiscernible]...

That's right. That's right. That's a fair way thinking about it. Yes.

I wanted to jump around to a couple of other segments that until the last 90 days were super topical, but now it seems to not be. One that I get asked about a lot is the wireless infrastructure business. You guys have a couple of new customers with a -- well, not new customers for the company, but new programs that are significant with Nokia and Samsung. And I wonder -- but then the end markets not be good. And the customer -- many of the customers in that space have had challenging financial results and a whole bunch of rocky things with the macro. And so there's a concern there that the new programs are carrying you through but eventually, you'll be correlated with the end market. And folks wonder if those numbers and expectations are realistic for continued growth there. And maybe you could just walk through the building blocks of how you're putting the wireless forecast together from here over the next, say, 18, 24 months?

Yes. Maybe I'll start, and Ashish, you can add. But when we look back, this business has grown phenomenally, right? We were at $100 million. I said we're going to get to $600 million. We got it to $600 million much quicker than we actually thought. It sort of kept growing. And if you look back over the last 8 to 10 quarters, maybe there was 1 or 2 exceptions, but for the most part, it's sort of kept growing sequentially, right? And then, we expect that to continue to grow this quarter and next quarter. And we signaled quite a significant step-down in Q4 is sort of in the outlook. So looking back at that significant period of growth, we know that carrier is lumpy, right? So it's really not a surprise when we're seeing a step-down. And then probably continues to be somewhat soft in the first half of next year. I mean, long term, you expect this business to be sort of GDP growth business. We still have additional program that we've won that should start ramping sometime next year, probably more weighted towards the second half. And so exactly how that plays, it's still early, but we do expect some softness in Q4 into the first half and then that growth to resume sometime in the second half.

Got it. Ashish, is anything?

No, I think Willem captured it. I mean, just not to put a too fine point on it, but we basically grew wireless 25% sequentially from -- in Q1 of already a very solid base. And I think we're going to be guiding essentially, call it, flat to up in Q2 and up again in Q3. So I think this is more really a result of more product cycles, success with our customers, and quite frankly, some of the regional areas they're deploying in, but we do have a call it, little softer outlook in Q4 for a couple of quarters. And then you will still see more capacity deployments in 5G, right? The first wave of deployment was coverage, where all your subscribers get a 5G signal, but as time goes on, you need to add more capacity. So there is still more legs to the 5G growth, for sure. And then as Matt -- as Willem said, we've got a second fairly chunky processor socket, which is completely in front of us at a very large customer.

That makes sense. I wanted to ask -- start to dig into storage a little bit, but there's 2 pieces of the storage business that I wanted to separate out. The first one is what was happening within the data center business through the majority of last calendar year that had the fairly sizable correction for a couple of quarters? And we're all trying to figure out timing and magnitude of recovery there. So if you guys have any sort of thoughts to summarize that situation of where we are and visibility toward any recovery on the data center storage side.

Yes, sure. When we look back, I think a normalized run rate is sort of pre-pandemic, right? And at that time, we said about 200 million storage and data center. And you pointed out, it's weakened very significantly to the point where in Q1, we were actually sub-100 million. Now we do see a very modest recovery this quarter, actually getting back to close to, let's call it, 100 million. And then we see that sort of bold through the rest of the year, but not to get back to our pre-pandemic level exiting the year. And that is really driven by just -- we're under shipping the end market so significantly. And we do expect to just -- to start creeping back up to actually shipping to what the end market demand is. So we're not expecting or forecasting recovery or growth in the end market. It's just simply that we're under shipping so significantly that we do expect to sort of start narrowing that gap towards the end of the year.

No, that makes sense. It's interesting when you see some of the magnitudes of stock moves that we've seen in the last couple of weeks. You might imagine that my inbox is full of very happy people and also some very anxious and not happy people. And so from -- there's been a lot of questions around the guidance that you guys gave for July and particularly around the strength of the consumer SSD business in that guidance. And maybe we could walk through the trends there. Was there stuff pulled in and then things might fall off in the back half of the year? That's a seasonally stronger period, anyway. I'm just trying to understand if there's -- that the up 30 shocks in folks. And so that brings a lot of questions that if you could give some comfort to some people that might be nervous about that, that would be really helpful.

Sure. I think you have to really look at consumer for Q1 and Q2 together. So if you remember, in Q1, we actually missed our forecast about $20 million. So we actually ended up doing Q1 consumer down 20% sequentially, and then we were up 30% is what we guided. But in reality, if you look at 2 quarters together, there was stuff right at the edge of which basically slipped out into Q2. So that's really what happened. So if you kind of look at the average number, it's up, but it's probably not up any much higher than we would have typically expected. In general, the consumer part of -- and a lot of the consumer revenue is storage related. As you know, we have a custom SSD controller going into a very large consumer application, which has been essentially been short of components for almost all of last year, and now it's finally ramping in high volumes. So I think if you look at Q1, Q2 together, the math actually makes a lot more sense. And what we've essentially told, in fact, we answered the question on the call the same way, you should expect it's going to trim down as we get into the back half. This is where you typically see peak seasonality for us on the upward direction in Q2. We also have a legacy printer ASIC business as part of overall consumer, which in pandemic who's doing okay for obvious reasons. But over time, it's going to trail down, right? I mean consumer at the end of the day is not an investment area for us. So that, over time, you should expect it's going to trim down as time goes on. The only really important business here, which we will remain focused on is this custom SSD controller, which tends to be a multi-generation kind of design win. So that continues to grow quite nicely. Everything else is going to trim around the edges.

Makes sense. Anybody? All right. The -- I wanted to kind of switch gears and think about long-term portfolio management and investment for the company. There's -- it's really interesting to watch some of the changes that cloud and in AI and then generative AI on top of cloud have brought to the winners and losers and semis to supply these markets, right? So there's certainly going to be networking switching and routing that's super important to scale with compute. There's custom compute that you guys are participating in along with Broadcom and a few others. There's much more focus on computing companies, whether it be NVIDIA, whether it be NVIDIA, whether it be eventually, AMD, whether as Intel gets its road map together, potentially them selling not just compute components but whole servers. And I wanted to think about the portfolio that Marvell has put together since Matt came to the company. How are you guys feeling about that? And Willem, I'm sure as you took over your new role, there was some thoughts given to the next 2, 3, 5 years strategically and the IP that the company has and the R&D that you're investing. Just against the landscape of compute going forward for the next 5 years, how are you thinking about the portfolio? Do you need to add things? Do you need to change focus on investments? You mentioned some of that to start the conversation.

Yes, sure. It's a great question. So I think, first of all, I'd start by just looking at the market, right? And so when you look at the architecture of what we're seeing in data center, historically, you've had sort of x86 -- well, not historically, but today, right? You've got a x86-driven server architecture with some component of acceleration that's getting attached to that, right? And so when we set up the ASIC business and the custom business, when you look at from a compute standpoint, that acceleration was really where we had an opportunity, right? And so when we look at sort of the architectural changes that everyone is talking about right now, there's a significant shift, right, where acceleration has become the core, and there's some CPU that's supporting that, but it's really inverted, right? And in our prepared remarks, we were talking about this, where because of this inversion and the acceleration becoming the core, our opportunity is actually significantly bigger, right? And so from a capital allocation standpoint, just talking about that process a little bit, we go through an annual exercise in the fall every year, where we go through every single business, look at what the market is, look at what our technology is, what the competitors are and look at our ability to compete. And clearly, the last couple of months here, a couple of quarters, the macro has changed very significantly, right? And so we've gone back and actually evaluate all those opportunities, all of our product lines, all of our products, and there are some areas, clearly, where there's significant upside, and we're investing more in those areas. And then there's other areas where things have clearly pushed out, and we can sort of take our time to invest. And net-net, when you look year-over-year, we're still increasing our R&D spend mid-single digits, but we're just really optimizing our model, still driving sort of towards our long-term growth model. And then seeing a ton of leverage in our model looking at next year, where we can drive to that long-term growth while really controlling our OpEx going into next year.

Yes. Matt, I mean, the beauty of, as I think Willem mentioned earlier is that AI is simply kind of the best killer application today of data infrastructure. But from an IP perspective, we already had all of that. Even before this upswing in demand came out, we were already focused on driving basically higher speed optics of 400ZR products, more custom silicon and of course, the need for them has grown, right? So if anything, I would say what's changed is, yes, I think in some cases, because the cadence of adoption in AI is faster, we're probably pushing us. So if we may have said, "Hey, we're going to go upgrade from 1.6 to 3.2 in 24 to 36 months, maybe that number is more like 18 to 24 months on, right? So we've got to make some changes, as Willem talked about. You going to pull some stuff in, push some things out. But in terms of like IP, I think we have everything we need for being a leader in data infrastructure. So I don't think anything changes from that perspective.

Got it. That's helpful. Last topic for me and is -- and this has been really an area of focus of conversation over the last 9 months is where you see the gross margin of the business going, right? And so there's been how much margin was contributed from storage and maybe that's gotten a little bit softer. What are the cloud optimized margins as they come through? What are the AI margins as they come through? If you can walk through sort of the moving parts and how you're thinking about gross margin over the next -- not just a couple of quarters while the macro is still crazy, but next 3 to 5 years? What are you thinking really drives the margin of the business?

Sure. No, good question. So just looking sort of more near term, a couple of drivers to get us back to sort of our bottom end of our 64% to 66% long-term range. First of all, as we talked about the wireless carrier, we do expect that to step down, right? That, plus storage continues to strengthen through the year. We see these optical products continue to ramp through the end of the year. So all of those are already margin accretive for us, which is great. See, the top line growing, so a little bit better overhead absorption. And then right now, we're driving super hard in our supply chain to optimize pricing, make sure that our operations group is really efficient and sized correctly. But all of that takes a little bit of time as inventory turns to flow through the P&L. And so we have a very good visibility to sort of exiting the year at 64%. Looking ahead, when we just step back, our long-term model always assumed a diversified portfolio. And we've grown extremely strongly over the last couple of years, and we've always sort of maintained our gross margin in that range, right? And so we know that as we see growth across all of our diversified portfolio, it typically -- we have growth in both our strong gross margin and lower gross margin and sort of balances out. Now I will say that if we see -- and we know that the custom is slightly less than our corporate average. And so if that starts growing way more than we expect today, clearly, it would put pressure on gross margin, right? That's fairly obvious. But we do see that being very accretive to operating margin. And so over the long term, very accretive to operating margin and then all falling through to EPS. So it should be very positive, higher growth rate, higher operating margin, higher EPS if suddenly this grows way more than we expect today.

Well, I think we're unfortunately out of time. But I just wanted to say thank you to everyone that's been in an audience for all of these sessions over the last 2 days. Willem, Ashish, thank you so much for your time and a great conversation. We really do appreciate it.

Appreciate it. Thanks, Matt.

Yes, thank you so much.

Thanks, everyone.

Thanks, everyone.

Thank you.