Welcome to day two of Citi's Global TMT Conference and the lunch keynote. It is my pleasure to welcome Matt Murphy, Chairman and CEO, and Dan Durn, Chief Financial Officer of Marvell Technology. Marvell is the top-performing AI semis stock year to date, and according to the maker of the five-layer cake, Jensen Huang, the next trillion-dollar company.
That quieted the room.
Right. Exactly. That got the attention.
Got your attention.
Matt, before I get started, I want to give you an opportunity to update us a little bit about the state of AI from a company that has evolved from ground zero over the last few years, from ChatGPT to the Astra moment. If you can just walk us through the evolution and where we are in that journey.
Sure. Can you guys hear me? How about now? Working? Okay. All right. Well, first of all, it is great to see everybody. Enjoy your lunch. Atif, thanks for having me. I would be remiss if I did not reflect, first of all, that this was 10 years ago. My first conference as a CEO was here at Citi. I became CEO in July 2016, came into Citi. It was not the friendliest environment. He had a sell rating on us. I just have to give you one more time.
Thanks for rubbing it in.
It is the 10-year anniversary. Why not? Marvell was in a very different place. I am bringing this up to your question, which is our pivot to the data center began in 2016. Basically, back then, we were a consumer company. 9% of our revenue in our fiscal 2017 was from data center, and most of that was just enterprise. It was really enterprise servers and storage.
That is the pivot we made, was to focus on the data center and data center technologies and data infrastructure technologies. That has been our guiding principle, actually, for the last decade. We went through, I think, great cycles in growth in enterprise, great cycles in carrier and 5G, and then we have had a very strong data center story for the last five or six years. Just to put it in perspective, in calendar 2023, we did about $2.2 billion in data center revenue, and next year we are looking at $15 billion-$16 billion, just to give you a sense. The business has really inflected. While ChatGPT and the rise of AI has been just an enormous driver and the driver of semi TAM, we were already positioned there because this was already our strategy going back from almost a decade ago.
I know now it is the place to be, and everybody has figured out who is in semis, that, wow, I better be in data center and I better be in AI. If you just figured that out over the last couple of years, it is going to take you a long time, is my view. This has been a company we built brick by brick, year- by- year, M&As, divestitures, huge organic investment.
Now we are sitting here, I think with one of the best, my view is, positionings relative to the breadth of the products that we offer, the suite of solutions that we have, and then really what we think is a best-in-class financial model in growth rate for the next four or five years off of the back of a decade of investment already.
Wonderful. To be clear, I do have a buy rating on Marvell now. That was 10 years ago. All right. Matt, we at Citi look at you guys like the Swiss Army knife of AI. There is a lot of focus on your compute business and XPU sockets and all that, but fundamentally, we think your networking portfolio is what outshines the peers. I was at the Hot Chips Conference a couple of weeks ago, and the message was networking is the next compute.
NVIDIA rolled out two more scales, scale-in and scale context, on top of scale-out, scale-up, and scale-across. Clearly, networking companies benefit from more connectivity. So my first question is on the scale-up opportunity. Beyond optics, as you look at the switching layer where different standards and proprietary approaches like ESUN, UALink, NVLink are all in play, how does Marvell think about its silicon switching opportunity across that landscape?
It is a great question. We think we see the scale-up opportunity, which we view very much as a combination of the switching platforms coupled with best-in-class optics, both NPO and then fully integrated CPO, which we offer both of. That is going to be an integrated sale in my view, and it is sort of greenfield at the moment. Now the backdrop, and I think why we are well-positioned is we have spent the last five years building out our scale-out switching platform, which we call Teralynx, and we basically acquired a company, Innovium, five years ago.
It had no revenue, but was projected to do about $150 million of revenue. We now said that that business is on track to do $1 billion +, and over the next years, we will probably grow even faster. That fundamental Ethernet architecture we have lends itself very well to Ethernet scale-up networking, which is ESUN, which is the scale-up standard. Even our existing product, our 100T part, which I talked about at COMPUTEX.
That product already is suitable and applicable to scale up as is. So we have got a whole suite of products coming on ESUN and a roadmap there. In addition, we see the market still not completely decided on a standard, so we have also got a robust investment as well in UALink. I think what is nice and common between the two is both leverage our high performance, best in class electrical SerDes technology for copper. Then we also are going to offer SKUs with the optics as well. So kind of a very full suite.
Then finally, with the investment from NVIDIA and our partnership there, we are also marketing and promoting NVLink with Marvell as a technology partner to enable that as a protocol to enable a lot of network fungibility as well. So there is a lot of shots on goal for us on scale up networking. It is brand new and today it is really dominated just by NVLink. But in the next few years, the need to connect more and more GPUs and XPUs together and enable the scale up opportunity is going to be, I think, a new market and one where Marvell can really make a meaningful impact out the chute.
Great. Matt, there has been a lot of discussion at our conference around the near package optics roadmap leading to the co-packaged optics. Can you help us understand how you're positioned to benefit from this near term wave around near packaged optics before we enter the co-package optics volume?
Sure. A couple of things there. The first is that we've been investing in NPO, it also used to be called OBO, onboard optics, and these types of solutions for at least five years since we acquired Inphi. If you go back to the Inphi acquisition, it's been obviously very, very successful. It was a high multiple at the time, but it worked out really well. One of the key capabilities they brought to Marvell was silicon photonics. The way that that's been used in Marvell to date has been primarily in our DCI products. In our modules that we sell, both for 100 gig, 400 gig, 800 gig, now to 1.6T, that all uses proprietary Marvell silicon photonics inside these modules. We've been in high volume manufacturing for 10 years.
That same team, as a parallel effort, has been developing NPO, and we've demoed that at OFC for the last three years, actually. This isn't something new that we just came up with. This has been demoed and proven, and now its time has come. What's happened is, along with that, we acquired a company called Celestial AI back in December 2025, which brought us even more integration in terms of CPO technology. We've integrated those two teams together. It's one very, very strong combined engineering team, and together we're able to show a suite of solutions from NPO to CPO.
On top of that, which is going to layer on even more revenue opportunity for us, is that in third party NPO modules, where a traditional module manufacturer would be involved, and these are some of the people we already do business with today on scale out modules, like at 800 gig and 1.6T, they're building NPO modules for the hyperscalers. We participate there with our TIAs and drivers, which is our silicon germanium analog product line. This is significant business for us. I mean, we talked about this being a billion-dollar business for us already just on analog components. That's really in the near term. Then Atif, as we ramp into NPO, our content is very high in the NPO module ecosystem with TIAs and drivers. That's another leg of growth.
When you add all that together, the broadband analog, NPO, CPO, and you look at where we were this year, which is effectively zero revenue from those. You look into next year, a quarter ago, we said it was going to be $0 - $300 million. This past quarter, I said that was going to be much higher, and a big part of our guide up for next year was going to be driven by networking and scale-up optics as being one key part of it. Imagine that number is even greater. From a velocity standpoint, unbelievable growth, and that's just at the outset. Calendar 2028, 2029, we're going to see a big ramp in NPO with CPO also layering in.
This is not an or, this is going to be an and you're going to see simultaneously NPO solutions in the market for some time. You're going to see CPO ramping in, and I encourage all the investors here to get away from treating this as sometimes we tend to do as a binary thing. If there's NPO, then it must mean that CPO is now gone. It was like saying, "Well, I don't know, copper's just going to keep going, so optical will never exist." They're all going to coexist. Copper's got a great runway as well. You're going to see all of these coexist in the market for several years, and it's a really exciting time because this is fundamentally new TAM for Marvell that we weren't in before.
Now we're very well positioned when you hear about scale-up optics. Of course, now we have the switching as well coming with it. A lot of our competitors don't have that. They may have one piece. Our expertise on the XPU side, deep design knowledge there, customer partnerships, and our ability to up integrate CPO and NPO with the XPU is also a skill set we bring. When we go to our customers, we really have the end to end, and it's a compelling story. At the end of the day, Atif, these are going to be bookended solutions in a lot of cases, where you really have to have an understanding of the digital silicon on both sides and then the optical link. We're uniquely positioned to enable our customers in this market.
You're already seeing real revenue next year from this. This isn't just a PowerPoint slide of what we might do someday.
Excellent. On the data center interconnect or scale-across, that market has been more mature and steady in the past, and you've had the leadership position on the DSP with the pluggables. You've been riding the 800G wave. As we look into the 1.6T transition, competition seems to be coming in as well. How do you feel about your competitive situation at the 1.6T?
Yeah, we're really excited about the DCI business. Marvell and Inphi pioneered this business at 100G. It was a very publicized product called COLORZ with Microsoft. We've maintained a very strong position at 400 gig and 800 gig, and I think we're as well-positioned as we've ever been at 1.6T. We were first to market on announcing our 2nm DSPs that go inside the module. The bring-up on those has been excellent. Our silicon photonics platform, tracking very well. Customer interest is through the roof. Because of our track record in 10 years of doing this reliably, we have 15 billion device hours of performance and reliability data we can show with best-in-class performance. So I think given our leadership position on the DSP side, that's going to really help us. On top of that, the market is going from traditional DCI to what you mentioned, which is scale-across.
There's market estimates that could 2x, 3x, 4x, 5x the size of the DCI market as scale-across really layers in. So for us, it's the same fundamental set of components, but the use cases when you get to 1.6T enable the scale-across architecture to happen. So we have really outsized growth plans right now in DCI. We've got an excellent leader for the business. We're investing not only on the component side, but in manufacturing capacity because we actually make the modules here, and we do sell our DSPs to third party as well. So there's an ecosystem that Marvell DSPs have created with our own products and then our module partners. But I think the exciting part is the inflection with the scale-across use case coming in.
Even if you didn't have that, just the DCI business, because you get more ASP per generation of DCI module, the number of ports are going up. The DCI use case alone would justify a lot of growth, but I think scale-across when it layers in is going to really upsize the market. So more to come on that, but very pumped on that business, yeah.
Great. Going to you, Dan, you are no stranger to semiconductors, GlobalFoundries, Applied Materials. Welcome to Marvell. I have a question around your R&D prioritization. When you look at the portfolio of products that Marvell has, and recently, when you look at your operating margin profile, helps understand how you are prioritizing your R&D focus areas and driving towards that 38%-40% target model.
Yeah, I think first and foremost, we see this incredible opportunity that is multifaceted. The diversification embedded in the portfolio has been a uniqueness in terms of the way we are driving the business. We are going to orient towards growth, and we are going to take this company structurally larger in size, materially larger in size over time. We will have more to say about this in a few weeks at the Analyst Day. Embedded in that is a philosophical approach that Matt takes to how we run the business, the culture that exists. There is a transparent aspect of how we synthesize the most meaningful opportunities that are out there. In a shared decision-making process, Matt drives discipline from a capital allocation standpoint across those opportunities.
We focus the investments on the most material drivers affecting the company, growth drivers affecting the company, the opportunities that we see that can also yield the best profitability. There is an entire rhythm and culture that Matt instituted when he came into the company. Seeing it on full display over the last month, you see a bit of the magic as to why we have been able to tap into those incredible growth drivers. Again, we will have more to say about taking the company structurally larger. We also want to grow the company in a very disciplined way.
At a moment in time where lots of things are possible, how you narrow the aperture, focus on the most critical items and do it in a disciplined way, and then drive execution with crispness, with precision, with rigor, with discipline makes all the difference in capturing that growth. But doing it in a way where your OpEx will grow roughly half the rate of revenue over this time so that we embed significant operating leverage in the model that as we take operating margin standpoint over time.
We are going to exit this year entering the long-term target model for operating margin at 38%-40%. We will exit next year touching the top end of that range. In a few weeks, we will reset what we think the new long-term target is. It is all about discipline, orient towards growth, execute well, and then grow the company in a profitable way.
Let's talk about supply. It is a never-ending debate around how much supply there is available to you guys and the peers. You have raised your 2027, 2028 outlook last two quarters. How much of that is demand upsiding versus supply unlocking? Can you just walk us through the process for the audience's sake? What is the process to get more supply these days?
Well, I think on the first part of it, if I just contextualize it for a moment. If I go back a year ago this month, post-earnings, I did a fireside chat after this conference to try to clarify for investors how to think about fiscal 2027, which is this year, and fiscal 2028. At that time, I said, basically, "Look, I think we can do $9.5 billion this year in fiscal 2027. There is not going to be a revenue hole. We have not lost any business, and things are looking good." Then in fiscal 2028, which would be calendar 2027, I signaled I thought we could do $11 billion+, let's call it. So that is $20 billion over that two-year period. You fast-forward to today, and we have now said we can do $12 billion this year and $18 billion next year, so that is $30 billion.
In one year, we have basically gone from $20 billion over a two-year period to $30 billion. So our supply chain has been able to flex up 50% over that time on a graduated basis. So it is all in conjunction, Atif, but certainly the demand environment has gotten better. CapEx floated up a lot versus what we thought a year ago. That has helped. I think the take-up rate and our success rate on our new products has been very strong. Our execution success has been great. We are getting to production 90%+ of the time on first-pass silicon on advanced node. So that has really helped us on time to market. I think we also had some pretty conservative plugs in for market share and our growth, and probably called the ball way too low on connectivity.
Looking back a year later, haven't missed a beat, have executed through our product cycles, both on 800 gig, which remains very strong, and 1.6T. Any of those concerns that were out there, I think were overblown. That's gone very well. We've said that that business this year is growing like 70% +, right. That growth rate's going to roll over into next year as well. High growth. We haven't specified exactly, but clearly the raise up to $18 billion, you could assume connectivity and interconnect keeps ripping higher. All that's been great. Supply has continued to perform. I think how you do it, and we had to learn this lesson a little bit the hard way five years ago is, the first five years at Marvell, the company was set up to do smaller packages. It was more tactical in nature.
We would have three or four suppliers for everything. We would beat people up against each other. We used a very traditional model. Once we acquired Avera and then Inphi and Innovium, we got these very complex reticle-sized products that required deep relationships in the supply chain in terms of substrates and back-end capacity that wasn't that fungible. You basically had to pick your partners, and during that time, we promoted Chris Koopmans to Chief Operating Officer, and he and I basically went off and struck a bunch of deals, long-term deals, with all of our supply chain, not just TSMC, our sole source vendor on the fab side, but also on the substrate side, back-end capacity, et cetera. That worked really well. That's basically what's gotten us from, call it, $4.3 billion run rate when we bought Inphi.
That was the run rate the quarter we closed in 2021. Now here we are in 2026, and we're looking at $12 billion in revenue. We've almost tripled the size of the company. If you roll into next year, then obviously it's up 3.5x, getting close to 4x. Those investments we made helped. We put in CapEx at the time. We did prepays. But we told our story. I think that's the biggest thing, actually, is suppliers are a lot like you guys. At the end of the day, they want to figure out who's the management, where's this company actually going, what end markets are they in? Are they credible? Do they have a good plan?
You can go to a lot of meetings in the Asian supply chain, and they'll be very polite, and you'll get a head nod, and they'll say, "Thank you very much." If they don't believe your story, you're not going to get what you need. In our case, we have a very compelling story that we've been telling consistently for five years, so they trust us. We've gone to fewer suppliers, more deep relationships. That's appreciated. We've upped the CapEx and the prepays to represent now the scale of the future opportunity five years later, so we're re-upping some of those. I just got back from an entire week in Asia visiting my supply chain, new vendors, existing suppliers, and just reaffirming where we're going. That has really helped.
That combination of credibility, trust, compelling roadmap, and then capital you put in, which is really putting skin in the game. At the end of the day, they just want to see you're committed and you have conviction. I think we feel very good about our prospects. We got an investor day coming up on October 6th. That is one of the reasons I wanted to do this deep dive with my suppliers is get them on board. I gave them a preview, "Hey, here is where we are going. Are you all in with me and Marvell or not?" All I got last week was a resounding yes, and we are going to be there with you with whatever you need, and let us plan it out. Let us plan it out. Yes, there is going to be challenges. Things are tight. It has been constrained since the pandemic.
It is very acute right now in some areas. I am not making any light of this. I am not saying this is easy. You have to have had a system and a mechanism and a set of relationships that was five years ago if you are going to get what you need. If you just popped out of nowhere and you said, "Hey, I got a data center part now, and I think I need billions of dollars of revenue next year," I think people are going to look at you like, "Where were you five years ago?" You cannot get some of the substrate capacity till the end of the decade if you have not started. How are you going to get it if you did not start five years ago?
That is the advantage we have of having a very single-threaded strategy in data center and data infrastructure, is this is all we do. We do not have some consumer side bet that is distracting us and keeping us defocused. We are fortunate in a way that our concentrated effort on data center and now AI has led us to an ability to tell a very clean story to our suppliers.
Wonderful. Looks like you are getting your fair allocation on the wafer side. Matt, you guys have a seat with the hyperscalers. You engage with three big ones, and then you are working with others on XPU-attach sockets. This topic around LPS, land, power, and shell, is increasingly coming up with clients, and your peers are highlighting that this could be some sort of a bottleneck beyond silicon supply chain. When you are talking to your hyperscaler customers, what are you hearing from them around concerns around whether it is regulatory approval or the timing of availability around land, power, and shell?
Yeah. I'm not sure I have too much new to offer there. I think this has been an ongoing issue and concern as we've ramped through this outsized CapEx cycle. To date, you can obviously see the demand side hasn't been curbed by that. I think the other one that's hanging out there too that actually our suppliers seem worried about is the societal pushback, right? Like, "Hey, are people going to allow some of these data centers to happen?" Because they're worried about it being in their backyard. So there's a bunch of constraints in the system which in a way I actually think is not the worst thing. Because I think in a way, look at the growth that companies like Marvell are experiencing with all these constraints.
I've been around the block long enough in semis to know if you get unfettered access at some point to infinite supply, the industry always overdoes it. I actually think there's some degree of healthiness on the AI infrastructure build, which is very durable in nature, very long-term in terms of its planning cycle, and seems to have a lot of legs and is delivering a lot of growth despite the LPS constraints, the pushback, regulatory, society-wise, and then just memory and advanced node wafers, and you can go through all the constraints. But look, companies that are able to execute are managing through a lot of chaos. The same thing happened in the pandemic, different framework, but some companies just were completely upside down on their supply chains, and they were extended, and they didn't know what to do.
It was like all these semi cycles, you emerge either a lot stronger or a lot weaker. There's no in between. I think we are proving to be one of the resilient companies through these turbulent times. But so far, we're getting what we need, I think, to deliver to the commitments we made.
Dan, coming back to you, this topic of AI financing comes up a lot with clients. You've committed $1 billion in fiscal 2027 supply prepayments, which kind of makes sense. But when you're looking at your peers, some of your peers that are forming partnerships and consortiums with financiers on funding some of these data centers, would love to get your thoughts on the risk profile of that as well as what is Marvell positioned. Are you also looking at those kind of engagements?
Yeah. Every company will do what they feel is right to facilitate the growth for their company. I would say our approach is going to be focused on bringing industry-leading technology to market, differentiating ourselves with that technology, and enabling customer success. That's what we've been doing for a very long time. That's what we're going to continue to do. Solve your customers' highest value problems, drive their success, drive our success. We're going to stay focused on being a technology leader and a key enabler of our customers. I won't comment on what other companies are doing and risk profile. I just know that we're focused on being the industry's best enabler of our customers in those highest value sockets.
Yeah. Can I just make one broad statement that I do think it does point just broadly that the amount of the breadth of investment that's now happening has gone from being very concentrated to much broader in nature because at the end of the day, a lot of smart people are looking at the economics and saying, "This makes sense." I think it's going to be good to see this outside capital at the end of the day come in to really make sure that the financing was there and the capitalization is there to really enable the AI returns to happen. As Dan said, we're really not in that. We're just heads down focused on doing what we do really well as a best-in-class semi company.
But I think the positive for me is that there's a lot of legs behind the capital commitments coming in for the AI build, and they've broadened.
All right. On the NVIDIA NVLink Fusion relationship with NVIDIA as both an investor and a partner via NVLink Fusion, how do you manage the tension between supporting NVIDIA's ecosystem and winning your own custom XPU sockets and the networking?
Yeah. I would say it is the opposite of tension. First of all, the partnership that we struck with them is, as you know, is far-reaching, NVLink Fusion being one of the core pillars which we can utilize as a partner of theirs to really enable network fungibility ultimately, and drive that standard more broadly because it is the most widely adopted scale-up interface in the world. That being said, we are going to go where our customers want to go, and so we have multiple shots on goal right now. We have a ground-up UAL investment. We talked about that earlier on scale-up switching. We have ESUN, which gets a lot of leverage from our scale-out, and then we have NVLink. From an NVIDIA point of view, they know we are working together to promote their solutions.
And we are unique in that we can offer it at multiple layers of the stack. At the end of the day, everyone understands customers are going to want to do what they are going to want to do. We just need to make a compelling case for each of them and then let them choose. I think it is money well spent right now to have multiple shots on goal here, and that team ultimately is pretty common. There is a lot of common investment, especially on the SerDes. So at some point, if one standard wins or we need to move it harder one way or the other, we have got the scale of the engineering team that can really go execute. But right now, we are very comfortable with the shots on goal we have and the partnership. We really appreciate all their support.
But as you saw from the announcement, it is broad reaching. There are DSPs, there is a collaboration on optics, there is a project on AI- RAN. There are a number of things that we are doing with them that were a continuation of what we were already working on, but this definitely memorialized the types of opportunities we have. So it is very strategic in nature and long-term, and so we will see how that goes. But we are actively promoting NVLink Fusion as a key standard in the industry, and trying to help enable their vision.
Awesome. Let us talk about the custom ASIC or the XPU business. Another theme that came out of the Hot Chips conference was this theme around disaggregation that the industry is moving towards prefill, decode, and many different types of chips around SRAM, different memory architectures to improve the flow of data, particularly in the agentic AI era. So you guys, the knock on your stock a couple of years ago was that you had too much concentration risk with one hyperscaler that you were ramping on this XPU socket and waiting for the other one to layer on. But most excitingly, you guys have announced another partnership with Google recently, in the $120 billion warrant-like revenue structure. My question is really around that engagement, the timing of the engagement. When should we expect revenues to flow from that deal that you have signed with Google?
Yeah. No, great question. A couple of thoughts. The first is, you are right. When we started on this custom journey, you got to start from somewhere. We did not have much. We basically got the core of the custom capability from Avera, which was a spin-out of GlobalFoundries. It was the custom ASIC team that ultimately had come from IBM. It was IBM's custom ASIC team. They have been doing this for 30 years, taped out over 1,000 products, a very experienced, capable team. But they were in GlobalFoundries where the roadmap on advanced node had been curtailed, so they needed a home, and so they ended up at Marvell. The whole run rate business of all of Avera was like $300 million when we bought it, and very little to no data center revenue.
But I knew, and we knew that that market was coming because we were getting asked by all these customers to do custom for them, and we did not have a team that could do it. So we started from nowhere. We pretty quickly got it to $1.5 billion, but it was really concentrated in one big customer, to your point, with another one coming. But we said also we got more design wins. But that was not really making investors very happy. It was like, "Hey, this is just too much." And it is the curse of the mega socket. You want to win it. It is so exciting. It is so great. And then you win it and you realize you are over your skis, you are concentrated, and everybody runs for the hills on the multiple.
But in 2025, and I will come to the relationship with Google in a moment, but in June 2025, we tried to assuage some of this concern, and we did a full-blown custom AI silicon event where we showed like, "Look, we got 18, it is now 20+ design wins across all the four U.S. hyperscalers of both XPU and what we called XPU-attach." And then at that time, I think there was a view that, and we called that as about a $15 billion market in 2028, which again, people sort of said, "Well, that looks kind of small," which I think that looks pretty big. When I started Marvell, the whole TAM of all of Marvell was like $8 billion. So this XPU-attach actually is very meaningful. And we talked about all the technologies we have for it in terms of our die-to-die, high-speed I/O, custom SRAM.
We had a whole presentation on that. Mark Kuemerle stood up, detailed it, because that was a whole IBM thing that came over with that acquisition that we still kept the team around on, that is used in high-performance networking chips plus high-performance AI chips. So laid out all the technology, and I think there is still a fair degree of skepticism. But if you look at what we just filed on with the SEC, the 8-K with the warrant with Google, that is really an XPU-attach relationship, and it covers all the same product categories actually we talked about a year ago. NICs, CXL, storage controllers, near memory compute, AI acceleration, a suite of these. And if you look at the total scale of the potential of the warrant at max performance, it is pretty stunning, what it could do.
Now, there's a lot of questions about it and investors trying to model, where does it layer in, how much do I put in. We have a great opportunity at our Investor Day coming up to contextualize that in terms of where our custom business is going. But the way you guys should think about it is, there's no one-trick pony in there. This is going to be a very diversified business within Marvell. I'm talking about the custom business. That's a number of sockets across all the big hyperscale customers that'll be very diversified and very rich in Marvell IP, especially the XPU-attach side, because in many cases on those products, in general, we get asked to do a chip for them, but it's a build to spec.
A lot of cases, it's actually Marvell doing the whole chip, but it's got the customer name on it because they don't either have the team size to go complete the project on their own, or they just don't have the people to do it that have the skill set. So it'll be a mixture of XPU, XPU-attach. So we'll contextualize, and we'll contextualize a range of outcomes actually because I think no one knows exactly how to call the ball with precision in 2030 on anything. So I think some ranges are important. But the key takeaway should be that XPU-attach was not a, hey, look over here strategy from the Investor Day. It was, we think this is a meaningful new market and a market that Marvell is actually pioneering and has a leadership position in.
Somebody asked us earlier today, well, can other people do these XPU-attaches. And how does the competitive landscape work. And we talked about that, which we feel very good about because this, again, is Marvell homegrown IP. A lot of it going all the way back to the acquisitions we did, like Cavium, like Inphi, like Innovium. But I did jokingly say, at some point, I think a whole bunch of companies are going to come out with their XPU-attach strategy, right. And all of a sudden, there's this market now, and here's what we're doing, and here's how we're going to go take Marvell's business. But I think we were ahead of the curve on that one, which has served us well because while other people thought it was a distraction, look at the magnitude with one account. With one account.
And we're working across the whole spectrum of customers, of which we don't have warrants with 8-Ks on them, so you don't know how big the other ones are. But you should assume that this is a very big category for us. So between XPU, XPU-attach, it's a meaningful opportunity. But I don't want anyone here also to go back to the future where we were in 2024 and 2025 and get hung up that there's one big socket we got to go obsess over, and also think that custom is the only game in town. Marvell as a company is driven by its connectivity technology. That's our core. That's what we're the market leader in. We are the category leader in interconnect, in connectivity. That's the bulk of our revenue. I showed this at COMPUTEX. That that's a differentiator of Marvell.
We have compute and we have custom, and that's a great base piece of our business. It's going to grow really fast. That's all going to be led on the back of our leadership in I/O, in connectivity, and that's the delivery mechanism for our large digital silicon chips. The other day, you were joking around about, hey, it was so crazy, Marvell hired an analog guy to go run Marvell. Well, it turns out back to the future, all the key proprietary technologies we have and why we win is in the analog mixed signal domain and DSP. That's where the value is, and that's where Marvell is driving. Now it's moving, by the way, last comment, from electrical SerDes and electrical I/O and copper interconnect to optical.
When you go, I'm now merging us over to this other discussion of connectivity, but when you go to the optical domain, and you then look at XPUs of the future, XPU-attach of the future, switching in the future, having that core capability in CPO and NPO and the ability to integrate, I think is going to be the key to the kingdom for the next 20 years. So we can do both simultaneously, but again, I think there's very few companies that can lead in the electrical domain on connectivity and I/O and lead in the optical domain, or at least do both of those very well simultaneously. Again, that is going to be the key to the kingdom on all these larger digital silicon chips that sort of get a lot of excitement and news.
I think at the end of the day, it always comes down to the I/O, and that's what we're really good at.
Great. No, I got to give you guys credit on the XPU-attach term you guys came up with first and definitely delivering there. We'll stay tuned for the investor day, but for now, you guys have said that you're on track to get to that $10 billion-$11 billion custom ASIC fiscal 2029 number, and it's-
Yep.
Custom sales are doubling next year.
It's on a good track. We've been very consistent and programmatic and prescriptive on this one because I think in the past, some of the custom expectations would run a little wild. Sometimes you get third-party data. Somebody posts on X. There's some tracker and the correlations, and I totally get it. There's a need to try to contextualize the market, and I think our approach, we flipped at the end of last year, was we're just trying to tell you what we're going to go do, and then we're just going to go do that and give you a number. I think for the Investor Day, we probably need to give a range for the future just because I think there is a range of outcomes.
Even the base case would be very compelling and certainly compared to a year ago at our custom AI day or even earlier, two years ago at our broader AI event in New York, I think custom sort of only floated up, but the rest of the business has ripped also beyond, I think what we had even thought was possible.
Great. Dan, as semiconductor investors, we get fixated on gross margins, and the nature of the XPU custom business is because you guys share the IP with your customers is intrinsically lower gross margins. My question is really around navigating, is there a structural floor to the gross margins for custom business where you walk away from certain sorts of sockets, and how you guys are optimizing the portfolio around the gross margins?
Yeah. So there's multiple parts to the business. As Matt keeps saying, it's a multi-pronged strategy not geared on one particular socket. We've got a great merchant business with a clear leadership position, with an enviable market position, and a margin structure that reflects that. As we think about custom, not all custom opportunities are built the same. In an XPU-attach, think of it as semi-custom. Think of it a lot of it is Marvell IP that's just customized in several ways for the environment that that chip goes into that has a very different margin profile than maybe, say, an XPU opportunity. So you've got different parts of the business. As custom grows, as the mix of custom grows, you will see incremental step down from a gross margin standpoint. But the operating margin opportunity will grow, and you see that in the current environment.
You see a slight tick down as we're ramping custom into the back half of the year, slight tick down from a gross margin standpoint. But in the most recent quarter, you see 180 basis point improvement in operating margin. You're going to see sequential increase in operating margin in Q3 and again in Q4, getting us to the low end of our long-term target range of 38%-40%. At a similar range of gross margin as we window into the next year, calendar 2027, fiscal 2028. You're going to see that operating margin walk from touching the low end of the 38% - 40% range to touching the high end of the 38% - 40% range.
We'll have more to say about this as we get into the Analyst Day and give the ranges of outcomes as it relates to merchant, as it relates to custom, the interplay, and what it ultimately means at a structurally larger scale for how the margin profiles, particularly from an operating margin standpoint over that window, to drive EPS growth in a very attractive way. We'll have more to say about that, but clearly the mix will play out, and we've got multiple aspects of the business, both very attractive merchant business as well as an attractive custom business that is going to produce a very nice EPS trajectory over time.
And maybe just to add, I think Dan said almost all of it, but just to add for some investors that are new to the story, the custom business, and we articulate this from day one when we acquired Avera, is while the custom business carries a lower gross margin, it has the ability to deliver commensurate operating margins with the overall company. Then you might say, "Well, how is that possible given the differential?" Again, this is for people new to the story, the way it works and how we account for it is a custom program, you are getting customer funding to do it because it is for them, so they are going to pay you a bunch of money to basically go in and make this custom chip for them. In some cases, the NRE is very meaningful because the development costs are high.
When you do that, the way we treat it from an accounting standpoint is an offset or contra R&D. So in the end, even if the product is carrying a lower gross margin, you get that offset, so the OM and the flow-through is actually very competitive. So, that is just part of the dynamic. That is why, as Dan said, as we are ramping with a little bit of margin compression in the second half this year, and it is not a lot, it is not a crazy amount. It is coming down mildly, but we are delivering very strong operating margin leverage through there. I would just say for next year, usually you get beat up on gross margin questions when it goes down.
We were having a meeting today where people were saying, "Geez, next year you guys said it would hang in there at the same level as the exit rate, but your custom business is doubling next year, but you are not doubling as a company, so how is that possible?" I think that is a positive story. We are managing it, is what I would tell you. We have margin-rich opportunities within the rest of the business, that is kicking in.
Remember also, before we were saying maybe you can probably assume margins come down a little bit, gross margins next year, a quarter or two ago, that is when Marvell revenue was like $13.5 billion next year, with custom about the same as it is today. As we have raised from $13.5 billion - $18 billion, most of that raise has come from interconnect and switching and storage, and some from custom. Some.
But custom has been pretty consistent in terms of the double, so that is helping us. I think the key point I want to make sure everyone knows here is, we are managing the business. We are not just going to print the results that just show up. We are going to also manage the business, and that means cost reduction, mix, allocation, price, the whole nine yards, to make sure that we are managing a competitive gross margin for the company despite mix under the hood. Now, if custom just goes and becomes half our business someday, of course, the company gross margin is going to come down. But the flow-through on OM and EPS is going to be so massive, it will be worth that for every investor, okay?
But I am saying even in the short term, next year, given the mix we are looking at today, and that could always change if things bias up or down, we are actually able to manage with a lot of growth in custom, an operating gross margin flat-ish with OM expansion. Not a bad story, at least heading into next year, and we will see how the year plays out.
Dan, on the capital allocation front, Marvell historically has been acquisitive. Cavium, Inphi, Celestial AI, more of a tuck-in type acquisition recently. So when you look at your priorities of dividend versus share purchase, debt serving, and also M&A, and you guys have a tremendous IP portfolio. How do you rank order where we go from here?
Yeah, as I think about capital allocation, I typically think of it as three priorities. First, we are going to orient towards growth. We are going to invest in innovation. We talk about a healthy pipeline from an R&D standpoint, organic growth. We will complement it from time to time, as you point out, from an inorganic growth standpoint, but pillar one of capital allocation is grow the company. Second, strong and healthy balance sheet. Flexibility in an environment like this is important. Third, return excess cash to shareholders, and you see that balance between both dividends and share repurchase.
In an environment like this, where you are seeing a major inflection shaping the end markets, and the end markets are going structurally larger, as part of growing the company, we want to make sure we have and preserve the flexibility to put order in the supply chain and retire those key pinch points of constraints before they materialize and be a rate limiter of growth. So there is another element in terms of growing the company in an environment like this. We work deeply with our supply chain. We have got a great ops team, supply chain team, and a good rhythm around this, where we are looking out on the horizon and making sure capital is deployed in a way so that it facilitates the incredible opportunities we see on the horizon. Grow the company, strong, flexible balance sheet, be disciplined capital allocators to investors with excess cash.
Awesome. One last question for Matt. Matt, as you look beyond fiscal 2029, once the current wave of XPU and XPU-attach programs matures, what's the next multi-billion dollar product category Marvell is building towards or we're under-appreciating?
Yeah, I think the biggest one is going to be, and I'd say this in totality, is the scale-up opportunity. I think if you look at the silicon switching TAM, then you look at the optics attach, at some point, you're looking at optics attach around the switch that's commensurate, I think, to the ASP of the switch. Then if you can bookend the solution, then you're putting that same amount of optics content on the XPU, GPU side. So it's quite meaningful depending on where it biases upwards to. But that whole area, which is in scale-up and then also higher layer networking and scale-out, again, that's a business we don't talk about as much. We're growing into that.
But both of those in total, I think switching plus optics is going to be very big, and I think when you look at that whole franchise. Then, of course, you've got on the scale-up, both UAL and ESUN, so we'll see how that plays out in the mix. But I think that one is a very large future TAM, where we're investing very heavily, and I think when we look back, that'll end up being one of the big additional growth engines for Marvell. But even the existing businesses we're in. Like I told you, DCI is going to scale across. That's got a dynamic that's going to crank the TAM up. On the traditional even interconnect business, as we go from on scale-out, which is our biggest business, and you go from 800 gig to 1.6T to 3.2T, you're getting more ASP per generation.
And the content, the attach rate of optics per GPU and XPU is going up significantly from generation to generation. And the whole industry's been shooting behind the eight ball on transceiver shipments. You can see this. Transceiver shipments keep floating up dramatically because I think there was not a good understanding of what those attach rates would look like. I think also CapEx has floated up. There's a bunch of reasons. But if I look out over the next few years, this interconnect story for us on traditional optical DSPs, TIAs, and drivers into the module ecosystem, that's going to have a very strong year in calendar 2028 and 2029. And as you have product transitions, like I said, into 3.2T, and the attach rate keeps going up, that should continue for some time.
Don't count out the existing businesses growing at 50% a year plus off a pretty big base. You're right on layering in new things. I think the switching plus optics combination on a three, four, five year view out and where a big meaningful chunk of revenue could come from that we don't have today is I think that's an area that's very exciting for us and right in our wheelhouse.
Awesome. We'll wrap it up. Matt and Dan, thank you for coming to the Citi conference.
Yeah, thank you.
Thank you.