Great. Thank you. Happy New Year, everyone, and thank you for attending JPMorgan's Virtual Fireside Chat here at CES this week. My name is Harlan Sur. I'm the semiconductor and semiconductor capital equipment analyst for the firm. Very pleased to have Matt Murphy, Chief Executive Officer and Chairman of Marvell, and Ashish Saran, Senior Vice President of Investor Relations, here with us this morning as well. Marvell stock was one of our top picks in the semiconductor sector in 2025. It's one of our top picks again here stepping into 2026. Leadership in cloud data center with their networking, compute, storage, custom ASIC solutions, currently seeing a strong tailwind from AI and accelerated compute, strong recovery, and new product ramps in the core service provider and enterprise networking franchise.
So, Happy New Year, gentlemen. Thank you for joining us this morning.
Thanks, Harlan.
Happy New Year.
Yeah, happy New Year.
Nice.
Let's start off at a high level. You know, the market is concerned around the sustainability of strong data center CapEx spending beyond this year, right? We all know the CSPs and the hyperscalers are going to grow their CapEx 50%, 60% this year, right? The acceleration of all of this started in the second half of calendar 2023, expected to continue strong into this year with the top four hyperscalers and CSPs growing 50%, 55% this year in calendar 2026 versus 65% growth last year. You've articulated, Matt, a strong data center segment growth outlook over the next several years for Marvell, and we'll dive into the details of this in a few minutes.
But given the lead times of AI infrastructure buildouts by customers, their mission-critical AI, XPU, and networking initiatives, you actually have unique visibility over a multi-year period of time with your order book also now potentially extending into calendar 2027. I think your customer silicon commitments are actually a good proxy for their data center spending trends and intentions. So do you see continued growth in data center infrastructure spending into and through calendar 2027? And what would you tell investors that are concerned around an AI data center spending bubble? What are they missing?
Yeah, no, great, Harlan, and thanks for the lead-off. Yeah, first of all, on the concern side, I mean, what I would just say we're seeing at the moment would be what I would call the exact opposite of a slowdown. We talked about back at the earnings call about a month ago, very strong bookings for the company, and then actually, you know, about a week later, I also went out and talked about the fact that, well, we had had record corporate bookings in Q2, record corporate bookings in Q3. Our Q4 quarter-to-date, like a month ago, was actually running ahead of where we were the prior quarter, and it's remained very strong even through the break.
What that means is really we've been laying in the backlog, the firm orders now to really support the growth that we talked about for next year in our fiscal 2027. So that's very encouraging to see. And what I would add to that and the discussions we were having at the end of the year and even a couple of meetings I've already been in, there's very detailed planning happening now for calendar 2026, fiscal 2028, calendar 2027 relative to the capacity that's going to be required, the program ramps. I mean, it's very, very detailed at the moment. So everyone's very much shifted their view based on this sort of very, very detailed infrastructure build happening to a multi-year outlook. And then there's also discussion of what does 2028 look like as well.
So it's short-term bookings are on fire. Backlog is laying in. We feel very confident and even more confident about our prospects for this coming year, but certainly for the year after. And we can talk about that. We have a whole bunch of great shots on goal that I think will actually accelerate growth in our fiscal 2028 versus fiscal 2027.
Yeah, you know, and as AI and accelerated compute again started driving this adoption curve in calendar 2023, it's also driven an incredible transformation of Marvell's business over the past 36 months as you're benefiting from the strong demand pull from accelerated compute and AI, given your leadership in networking, compute, storage, and ASICs, right? Your overall data center business is up 3x since calendar 2023, right? And from our perspective, our performance for the team is technology and product leadership. It's breadth of the product portfolio. It's execution. It's better demand and deployment by your customers, right? So help us understand the factors that drove the strong data center results last year, outlook for this year, and over the next few years.
In particular, can you just further elaborate on the assumptions behind, like you laid out to us at earnings, but your assumptions behind customer revenue forecast to grow +20% this year, but more importantly, to double in calendar 2027 over calendar 2026?
Yeah, great. No, a lot of good stuff in there. So let me lead it off. So first of all, you know, we've been. This was not an overnight sort of success that came out of nowhere. I mean, we were planning and building the business. If you even go back to 2019, Harlan, when we did the acquisition of Aquantia, we sold our Wi-Fi business. We bought Aquantia. The next year, we followed it up with Inphi. So a bunch of, you know, Innovium after that, and then the two recent ones we've done. So a lot of inorganic activity to get there. Also huge organic spend, right?
If you just look at Marvell's outsized R&D budget, I mean, we've invested through all the up and down cycles we've had over the past few years, and now we've got a $2 billion-plus R&D machine that's absolutely cranking. And so the 3x growth we've seen over the last few years in data center, that's from investments prior to that. And just to set the stage, yeah, this past year, the company, based on our Q4 forecast, will grow in excess of 40% year over year. And that's, by the way, I would note that's with the divestiture of our automotive business contribution from the new acquisition. So actually, growth rate's probably more like 44% or something if you just pro forma it. So really, really strong growth.
For next year, and the CapEx numbers keep moving around and floating up. We used about a month ago a 30% base case, right, for the cloud CapEx next year. We said that our optics business would continue to grow above CapEx. How much we'll see, but certainly that's been part of the strong booking story for us, but that'll be definitely above CapEx. Our custom business, which I'll get to in a second for the year after, we said we'd grow 20%, which is off of a really strong growth last year, which is over a double plus, you know, the prior year was also a double, so having a transition year, but still 20%, I think is very reasonable, and then the rest of the business, which is our switching, storage, kind of other management products, that's also going to grow like 15% this year, so that's a nice setup for data center for the coming year.
And I would just note even with the divestiture, the comms and other bucket is going to grow about 10%. So that's the outlook for this year. And I would say, again, just based on the order backlog, I feel a lot of increased confidence in achieving that.
Okay, so now you go to the next year. And again, we've got, again, who knows, but we have sort of just put in a plug of 20% CapEx growth in fiscal 2028, calendar 2027. So again, optics should absolutely keep outperforming, especially with some of the program transitions that we're seeing there. And then the question you asked in your upfront was on the custom business, right? Well, how is that going to double? It's only growing 20% this year, although like I showed you the trajectory, it's just been absolutely up and to the right.
Here's a way to think about the custom business doubling in fiscal 2028. And it's not as daunting as I think maybe some people's initial reaction. Okay, the first is we have a very strong second half setup in custom for this year, fiscal 2027 that we're in. E e said the exit rate would be about $2 billion on a run rate basis, okay, of custom. So just kind of take that number. And just that's a reference point is we've given is roughly $1.8 billion for the full year, but then for the year we're in, the exit rate being at $2 billion .
Let's just assume that grows at CapEx. Just 20% on that, right? The kind of the base. It could be better, who knows, but just use that. So that's $2.4 billion . So if you want to double from $1.8 billion this year to next year, that gets you to $3.6 billion. So you got to add another $1.2 billion. But remember, we also have very strong ramps happening in fiscal 2028 on our XPU attached business. Okay, we have 15 different sockets that are going to be going into production starting this year and next year and the year after. That's going to contribute meaningful revenue because even a subportion of those we talked about, which is the NIC and the CXL categories, those are going to be like a billion plus each the year after, right? So that's going to flow in.
Again, how big are they going to be? When are they going to be timing? We're still figuring all that out, but those are coming. Okay, so that's going to be a big chunk of the growth if you look at the $1.2 billion, and then really what's left over is another big XPU program, which again, those are larger and episodic and we'll see, but that program is also tracking and we're planning for production. The magnitude we don't know, but if you actually back it all out, I think it's a reasonable set of assumptions we have on growth. We're not putting in, so I think the mindset that like, hey, this whole double is predicated on like one socket and if it doesn't happen, there's a lot of moving pieces because again, we've built this custom business up from a couple of sockets, a handful to like +20% , right, that are working their way through the system, so I think that's still a very, very good base case.
And then you take the switching, storage, other. You put another 10% on that guy. And then we just had common other, like just a GDP, just to make it simple. So if you do all that, you basically get data center growing at like 40%, which is 2x CapEx, which we've done before in ramp years. And then the rest of the business, I think, is very conservatively forecasted. I think this whole base case of like 25% next year and 40% the year after on data center. Obviously, take the other business more like 30%. I think it's a great setup we have. We're very comfortable with the setup, Harlan. Our visibility, the programs, the quality of the business that we've taken that we've won. And we'll see. We'll see how it develops.
No, that's a great update and good to see you reiterating all of the prior targets that you put out and even better to see that I think seems to be coming together, orders, visibility, backlog, and so on.
Yeah, everything's loaded up in the last 30 days, Harlan, just in terms of just what the outlook is and what people are thinking, so.
That's perfect. Let's switch it up a bit and focus on your networking leadership, right? Both traditional scale-out networking and new opportunities within scale-up networking, right? While Marvell's scale-up business is well understood with your market-leading optical business, rapidly growing Ethernet switching business, scale-up looks to be a big focus, right, given your recent acquisitions of Celestial, today's announcement of your acquisition of XConn Technologies, as well as your organic investments in UALink switches, right? So can you just elaborate on what Marvell is doing to enable rack scale solutions for your customers?
Yeah, I think that's the right way to say it. It's a scale-up opportunity, but really the way that customers are thinking about this, and I would say even since we've announced Celestial, I think that sort of framework has really kicked in even with our customer base in terms of how we're thinking about approaching it. So yeah, the first thing is on the scale-up and the rack scale solution. I mean, we've been investing here. I think a lot of the progress we made got a little lost last year in the discussion maybe around custom silicons since that was such a big part of our revenue growth last year.
But the reality is, look, on the copper side, there was a big transition that's happened in the last couple of years where the initial deployments of, and products that were released to market on AECs and PCIe retimers were all based on NRZ modulation. And so our view a few years ago was we should time our strategy around the transition to PAM, which we did. And so we had a lot of success on our first-generation products, both our 100 gig per lane AECs as well as on our 1.0, sorry, on our PCIe Gen 6 retimers. And we did a lot of discussion about a month ago on level of customer engagement, product release, roadmap. And that business is still on track to like double plus, right, in the coming year. And that's all copper-based today, but that's still part of the solution and will be there.
Of course, beyond that, that's going to continue to grow. We added Celestial, right, which adds a photonic fabric now capability to us, which is when scale-up moves out of copper interconnect and moves to optical to have like a bookended solution where you can have a photonic chiplet integration on the ASIC side or the XPU side as well as on the scale-up switching side is going to be critical. We'll get to that later if you want to talk about it, but that's very different than the narrative and the approach that's been going on scale-out for years, which was already served by optical, right? That's already been served by optical on pluggables. This is a different transition because it's not served in the optical domain today. It's served first passive copper, then active copper. Now it's going to move to optics.
We're well positioned there. And then to make it all happen, you need the fabric, you need the network. And so the scale-up switching really becomes mission critical. And there's two approaches we have there. One is based on an industry standard called UALink. We're investing there today organically and have a plan in terms of tape out and sampling that we've articulated for next year, but we'll be very competitive on that product. We'll get to XConn in a moment, but that is going to help us a lot. And that was actually a key reason to join forces with them is that this is a veteran team, Harlan, of experts in fabric switching. They come from a very strong background in PCIe. And so the benefit of this acquisition actually is we get a PCIe business on switching we did not have before. It's something we've wanted to do, but these are complex technology. It's hard to do it yourself.
We also get CXL switching as well, which was another piece of the puzzle on CXL. But probably most importantly is this can really help accelerate our efforts on driving a very compelling roadmap on UAL. And then actually our team also is comprised of experts, which came from Innovium that have Ethernet background. And so there's another standard called ESUN, which is Ethernet scale-up networking, which we're also working on. So it actually just turbocharges our team and our ability to go execute a very full roadmap across multiple switching technologies with all the best in class interconnect in it. And so as you sort of roll this through, the TAM opportunity when you start adding all of these pieces is just deca billions of dollars, right?
So it ends up being kind of the single biggest new incremental TAM for Marvell. So you could now see why we've been very active here on the M&A side. And by the way, I think we've been able to do that and still preserve a lot of firepower financially, right, in terms of buybacks and capital allocation strategy. And that's very consistent with our overall corporate strategy. And it kind of flowed, by the way, from the automotive divestiture we did last year, right, which was sharpening our focus and then doing that combination of return of capital to shareholders and then some smart M&A. W e're super pumped, right, about bringing together all these inorganic teams into the current Marvell fold.
You've seen what we've been able to do in terms of the acquisitions, how we're able to bring the teams in, get them and really expand their market opportunity significantly beyond what they were able to achieve on their own. Let's be clear with XConn. We're getting an existing set of design wins, book of business revenue from that acquisition, which is great. I think that the team and the talent is really going to help us not just this year and next year, but really as we extend the roadmap over the next five years. That's how we're thinking about it. Very, very compelling.
And Harlan, maybe just to add, I think, and hopefully this is coming clear, we're in a very unique position if you think about scale-up. We're basically one of the few companies which is going to offer the entire solution, right? The interconnect, both copper today, right, and the switching copper today moving to optical in the future. In fact, if you look at the design win, which comes along with Celestial AI, I think we think we'll probably be the industry's first company enabling basically optical in scale-up, right, in high volume with the tier one customers. That's a very, very unique position. And then we're in switching, I think the interesting thing in scale-up is it's a brand new market, right? No one has any really significant shares.
So in scale-out when we entered through Innovium, there was already a large embedded competitor and we've done well and we've really massively grown the business, $150 million, doubling to $300 million, now $500 million. So it's done great, but it was fighting uphill against somebody who's already in the market, right? The really nice part about being this fast to enter scale-up switching is no one has any, that's a new market. And that's a massive opportunity for Marvell. So it's a very, very unique position we built. W e're really all in on scale-up is the way to think about it.
The nice thing about the XConn acquisition is that Marvell has been focused on the retimer part of the PCIe market, right, and various constructs within the CXL market, right? And XConn brings in more of the core switching capabilities.
So your portfolio is with XConn, your portfolio is pretty complete, both from a PCIe perspective as well as from a CXL perspective, right? And it sounds like based on the press release, XConn is going to be ramping some of their PCIe and CXL products second half of this year. They've got like +20 customers, right? So very synergistic from customers, from your portfolio versus their portfolio. Is that kind of the right way to think about it?
Yeah, absolutely. Yeah. It absolutely fills out the portfolio. I mean, think about CXL. We have a very strong position in expanders and accelerators, but this adds the switching and pooling, which is all connected to each other, right? If you just look at the block diagram coming out of CPUs or XPUs, and if you want to gang up more memory, you're going to have to have a switch in between at some point. And so it literally fills out almost a chipset full solution. W e can really do a lot with it. But I think the fundamental strong engineering background of the team and that's coming over is going to serve us very well. They're just very experienced here.
And it's like anything in semis, Harlan, when you find the right group of people that have been dedicated to doing something very specifically, you can look at it on paper. It's like how many people tried to do what Inphi did, right, as an example. I mean, even when I was at Marvell before we acquired Inphi, my team was saying, "Oh, we can go do that. We can go hire some guys and build some TIAs and drivers and build a PAM DSP." And it's like, "No." I mean, you really have to know what you're doing here. And Jerry Fan, who's coming over, he's the CEO of XConn. He's going to have a very strong role here. I mean, he actually is a Marvell veteran. He was the PCIe back in the day, he was the PCIe IP guy in central engineering.
I mean, the level of detailed knowledge the team has from core IP to the switching fabric to the standards and how it all works together. And really their concept was designing these products for the AI era, which is much higher radix and feature sets to support, in theory, to support scale-up networking in the very early stages. So they've had a very strong view of the market. And I think it's going to be a great addition to Marvell. But yeah, it does fill out the portfolio very nicely across the board.
One of the areas on scale-up where the team has been sort of really leveraging your DSP expertise, your networking connectivity expertise, right, putting those two together and attacking the copper market with an AEC solution, right? You've talked about having wins with two major tier one hyperscalers. Is that 800 gig or 1.6T? And you've talked about your AEC and retimer products growing 2x this year. Are we talking a few hundred million dollars in sort of revenue scale? Is that kind of the way to think about it?
Yeah. O n the first one, so the 800 gig products are what's ramping, and those are based on 100 gig per lane technology. That's where the ramp is happening, but we're going to be migrating those over time. And in terms of sizing, I mean, I think that's right. I mean, I don't remember exactly what we said to Ashish, but in terms of the it's multi-hundred, right? If you add that plus the retimers.
That's right.
Yeah.
That's right.
Okay.
Yeah, I would say Harlan, in the near term, which is just this coming fiscal year, which is calendar 2026, fiscal 2027, our retimer plus our AEC business is going to be basically, call it in the multi-hundred-million-dollar range. That's just a very near-term number, and if you go out a few years, this is going to be several hundred million between the two product lines, right? That's a very big, very large market, and we're basically hitting it at the right time where this transition to PAM is actually happening in a very big way.
Yeah.
I mean, that's part of the double year over year. I think there feels like there's some upward bias to that, so.
What is the--
That's right.
What is, -- I mean, obviously we all know that you have leadership in optical DSP, and it's that DSP architecture, which is a big differentiator even when it comes to things like copper and AEC, right? But can you guys just give us one or two metrics where Marvell kind of has been differentiating themselves on the AEC side? Just curious.
I think the way you want to think about, Harlan, is it's the same reason why we were able to be as strong as we are in the optical space, which is we essentially build embedded DSP, which can basically be used by any optical module manufacturer, right? And build a module because we have enough basically signal to noise available where it can essentially overcome any of the typical inefficiencies as you go through the entire signal path, right? We're doing the exact same playbook, essentially, in the copper market, which is what the customers want. What the customers want really is multiple high-volume suppliers, right? So think about the module equivalent being a cable OEM in the world of AECs. In fact, we had a press release in early December where we launched what we call our Golden Cable Initiative.
The idea is very similar to what we run the optical module market very successfully, which is, hey, it's a standard product. It has enough capability that multiple cable OEMs can manufacture using our reference design and all essentially offer a very competitive product to the end customer, right? It all comes down to you basically have to have a DSP which has enough high signal-to-noise available where you don't have to tune the product, right, for every single cable. That's really the big differentiator.
Let's talk about scale-out Ethernet, right? I mean, you're on track to drive $300 million of switching silicon revenues this year with the Teralynx 12.8T switching platform. You're now ramping your 51.2T Teralynx 10 switching platform, which is expected, which is comparable to Broadcom's Tomahawk 5 chip. And you guys are expecting your switching revenue to grow to, I don't know, $500 million plus this year. Could you guys just provide us an update on customer design wins beyond your primary customer and share some insights on the team's product cadence and roadmap over the next several years here?
Yeah. No, thanks. I mean, we started this business with Innovium. Again, this was a case where we looked and internal team said, "We can go build something." And we could. It would just have taken a long time. And you had a dedicated group of folks that had put in a huge amount of effort with private venture-backed funding to go get there. And they were the best asset. And they've done great with us. W e started at 12.8T with a lead customer. We've ramped that successfully. It took a lot of the Marvell know-how to make that happen and drive it into high-volume manufacturing, which we did. We followed that up with our 51.2T product, which is Teralynx 10. That product is ramping very aggressively into production this year. And this will be its first real major year.
We released it probably a year and a half, two years ago in terms of sampling, but it just takes time to work its way through the culture, but it's a very high-performing product. It's all the Marvell IP, Marvell Certies, etc., and our 100T part is coming out soon too, and so we have a committed roadmap there, and yeah, the revenue growth has happened, and it's at kind of $500 million plus this year. Probably just keeps going, and we've also been expanding the customer base. We'll see that later, but yeah, given the product traction and the success and people can see that it's in the market and it works, so we're aggressively investing in this part of the business, for sure, but it's great to see the revenue actually coming in. We have both kind of where we thought it would be at this time.
Yeah. With the revenue growing, right, from $100 million to $300 million to $500 million, Harlan, that's certainly come along with an expansion in the customer base. And just kind of going back to the XConn acquisition, kind of the other reason we did it is, I mean, the overall level of activity for our switching team is just increasing very significantly. And so the beauty of what that lets us do really is we've really basically got multiple parallel teams now. In effect, in our switching business, you've got the traditional scale-out switching business, which is what we're talking about today. So that team just gets laser focused on just that particular market. Obviously, some of those learnings fold over into our ESUN initiative, which is also Ethernet-based.
And then you've got this very unique, very focused effort, which is purely on UALink, where you can basically leverage a different switch architecture for this very memory-intensive set of operations, which is really what you need in the scale-up market, right? So that's kind of the other beauty is that it really lets us double down even more on the scale-out market, right, which is the Innovium business. And we can just keep increasing the cadence. So 51.2T today and 100T product is not that far behind, right? So I think this is where we can really go push the accelerator.
I think the other thing I would say that we see is that I think we have a unique combo here now because on UAL, the die size is much larger than PCIe. It's a whole different beast. I mean, if you look at a fully equipped scale-up switch we would be building for UAL, say like at 115T, the die size gets comparable to like a TL10, like an actual 50 or 100T Ethernet switch. So the reticle buster almost type of size, which is a whole different skill set. I mean, quite frankly, in Marvell, we've got kind of three groups that have been doing this, right? We have our ASIC group, advanced node, massive die, our switching group, and then also some of the processors that we've done. But really on the switching and the ASIC, it's been the two big ones.
The difference is then you say you take large die, that's a unique skill set of the industry. That's where a lot of the other switch vendors, quite frankly, all fell apart, right? There was a bunch of people trying to make Ethernet switches at 3, 2, 6, 4, 12, 8, and all of them failed and were unable to be successful. One, it's a whole different skill set in terms of designing these products. We can do that. The second is that the I/O intensity in the I/O die area is massive on scale-up. And so again, you're talking about real critical SerDes integration. Using third party is going to be very difficult there. I mean, we had that issue with Innovium. They were using third party SerDes on their 51T part. It didn't work. It did not work, and we knew this in due diligence.
We had to rip it all out and put our own stuff in because it literally was not feasible, I think you get a couple of things, so I think getting our UAL effort boosted with XConn is going to help, but then we can help them because of our expertise in these other areas and pulling, and then finally, just the ability to actually load the software on it. I mean, again, we're used to doing things like SONiC and SAI and supporting networking standards, so we have a whole software team, so this isn't just building like a widget, this UAL product or ESUN. This isn't just like a one-off chip. It's a whole embedded solution, and then you start talking about layering photonics into it.
How are you going to do that integration? And then how are you going to be able to be the guy on the other side on the ASIC side to make it all talk? I hope you're getting a sense of what this is really going to take. And I think to Ashish's point, yeah, there's going to be a lot of churn. It's a new market. How's the standards going to evolve? We're just scaling up our team to invest across the board so we can be there and guide our customers to the best solution to get to market. But to do that, if you come in as just a one-trick pony, I think it's going to be very difficult in the end to actually get the product to be manufacturable and ship in high volume. And we did see this with Innovium. The product we acquired at 12.8T was not manufacturable.
We get it manufacturable. Marvell did it, and we're able to do that with our own expertise, our own experience, and so again, that combination, what we've seen with these M&As, has been just very positive for the outcome, I think all these hang together in a very clean way.
Let's maybe shift the focus to your custom ASIC business, right? Marvell, second largest custom ASIC supplier in the world. We often have to remind investors this, but Marvell has been doing ASICs for over 25 years. Strong position in storage, enterprise, service provider, ASIC solutions, now building share in ASICs for AI-based compute workloads. Even before AI, you guys were building a ton of ASICs for data center applications, right? You've got now multiple custom ASIC XPU programs, either ramping or in your pipeline. The market has always thought about ASICs as fixed function, customer-specific, right? Yet we're seeing with AI that several of the AI frontier model developers are actually using or potentially migrating their frameworks from merchant GPU to ASIC XPU architectures, right? Some examples, right? Anthropic, frontier model, training on Amazon's Trainium XPU platform, right? An ASIC XPU platform where obviously you've co-designed with Amazon, right?
We've seen rumblings about OpenAI potentially migrating its frontier model to Trainium as well. And then, of course, you've got Anthropic migrating some of its workloads to TPU and so on and so forth, right? So obviously, if your customer's XPUs were truly fixed function, these types of AI model porting would not be possible. So help us understand how flexible, programmable are these XPUs? And do you see more porting of AI model training inferencing to ASIC XPU architectures?
Yeah, I agree with how you explained it when you gave the setup 100%. I actually think I'm surprised this narrative is still out there. This was something that was kind of on people's minds like two, three years ago, as well, could custom ever take off because it's a fixed function? And then not just us, I think the industry was sort of like, no, I mean, look, this isn't an HDD controller. I mean, it's not some so by design, they've designed them this way. And I think just the way to look at it, I would say, is there's just no way you would have seen this level of mass deployment. I was talking about industry-wide. Forget about just Marvell. Just mass deployment industry-wide of custom XPUs. It's absolutely happening on a massive scale.
You would never see that happen, right, if it was a one-trick pony type of solution that was tailored for just one model, just one type of application. So no, I mean, I agree with what you said, that there's a tremendous amount of programmability that has been built into these and that I think that's become effectively a non-issue. There's other reasons, right, why merchant XPUs are the right solution. And I totally agree with that. And that's going to continue to be a massive, the biggest part of the market. But the notion that, oh, these are fixed function, they're never going to go anywhere, it's already taken off to be a huge portion of the accelerator market. In 2026, the numbers are going to be really big, right, for everybody.
Yeah. I mean, Harlan, if you look at it, right, look at the ramp these products have done over the last four to five years. And it's not like there's a new one coming out every six months versus we know that the underlying models have been changing very quickly. So despite all the models changing, custom products have continued to grow. I mean, it's almost like that's your ultimate proof point, right, that, hey, they're still growing despite the fact that the model is changing. So if they were truly fixed function, that would have not happened. So yeah, I think that question just doesn't make any sense anymore. I think I used to get this question quite a bit a couple of years back, but I think the evidence pretty much tells you that no, these are not.
These are fully programmable devices as flexible as a merchant solution.
As the focus on AI compute workloads accelerated, started in 2023, that enabled you guys to rapidly engage in custom ASIC XPU programs with three out of the top four global hyperscalers back in 2024. Fast forward to today. And you now have secured over 20 XPU and XPU attached custom chip programs in your pipeline across many customers, right? Your first lead XPU customer ramp started second half of 2024, drove strong growth for the team. Last year, the follow-on program to that current program is going to set the stage to help you guys drive +20% growth in your custom business this year. You talked about having purchase orders for all of 2026 for this follow-on program.
Given what looks to be strong demand for this follow-on program post your customer's launch of this program, has your order book and backlog for 2026 continued to improve? Just curious.
Yeah. Well, first of all, yeah, I think you summarized it well. I tried to summarize it back a month ago, and there was still some level of noise afterwards for that reason. But basically, yeah, I actually have a month ago, even before that, we already had purchase orders and backlog support growth that we're talking about. And that's still consistent. W e feel very good about the growth there. We feel very good about the product transition that we're managing. And like you said, we built this business up to +20% very complex, large competitive chips that we're building for the world's most important companies for this AI infrastructure buildout. And all those are going to roll through the system.
I think, like I say, consistent with my other comments about backlog and order coverage and business momentum, that also applies as well to the custom part of our business. And I'm still sticking to what I said a month ago, but certainly I would say incrementally across the board on all of our product categories, including custom, incrementally more positive and bullish and confident about the growth that we're driving for this year. And I already gave the setup for the year after, which I think is actually quite reasonable as well. I think we set a good baseline for everybody to go off and execute to.
We're very excited to be very, very excited entering 2026. I mean, the team came off a great year. Like I said, +40% revenue growth, seeing the hard work and labor of the team actually delivering, and then being able to do all kinds of things on the aggressive side, focusing the business, doing acquisitions again, customer engagements. I mean, I was having calls over the break about how we can get together in the new year. And some of those were to do with the acquisitions. Some of those were to do with capacity planning. Some of those were to do with how do we pull in the roadmap. So yeah, super strong level of activity. And the Marvell team is firing on all cylinders right now.
There's been a lot of focus recently on SRAM-based XPUs, right? Much more efficient, cost-effective for low user count, lower parameter, large language models, primarily suited for enterprise applications, right? People don't realize this, but the Marvell team actually helped to architect and design one of the first SRAM-based XPU chip designs, right? And you've continued to innovate with your embedded SRAM architecture with an AI-optimized 2-nanometer SRAM memory platform. What's your view on SRAM-based XPUs use cases? And does Marvell have any XPU designs within its pipeline utilizing its 2-nanometer SRAM architecture?
Yeah. No, I mean, it's a key IP, Harlan, that nobody talks about or seems to care about. But I remember in 2019, in the Aquantia due diligence, okay, that this was highlighted as this key strength of the team. And they had a world of experts from IBM lineage that had developed the most compact, power-efficient SRAM. And the view back then, which was very prescient by the team, was that, by the way, not just for XPUs, which was really not thought of in the same way, but certainly for complex processors, but also networking, right? Same thing. If you look at our Switch products. And so it's a key piece of IP. And actually, we spent some time on it at the AI custom event we did last year.
Mark Kuemerle, who came through us through Aquantia, actually walked through a whole presentation on it. But basically, we had like 50% area improvement, 66% lower standby power, so kind of giving us a massive advantage in terms of our design. Yeah, that's a key IP also for XPUs. It's something our customers know about and they lean on us for. It's just part of the toolkit. But it's definitely a differentiator. It's one where people just think SRAM, oh, that's like a 1980 technology. It's not a. Embedded SRAM is a key part. Yeah, you can see that on some of the companies out there that have developed competitive XPUs. They've leveraged that big time. That's a capability that we bring to the table.
Given the strong growth profile that the business is looking for, is the team concerned about supply constraints looking into this year, into calendar 2027? If so, is it more around 3-nanometer wafer supply, 2-nanometer wafer supply, CoWoS memory? Any concerns around potential bottlenecks, potentially slowing your potential rapid growth profile over the next couple of years?
Yeah. I mean, I think the way to think about it is we've been in a world, Harlan, for probably the last five or six years that's been fundamentally constrained in different ways. The first wave of it was during the pandemic. And that happened for all the reasons we knew about, which caused manic order patterning and order patterns. But we made a shift back then at Marvell, which was really set up as a tactical supply chain management approach, which was sort of get a bunch of suppliers, bid the lowest price, bluff. And sort of my recognition was that that's got to change. I mean, our business was changing at that time. We're going for the cloud market. So we retooled the whole thing. And that's when Chris Koopmans took the lead running operations.
The reason I bring all that up is we fundamentally retooled and built back up our supply chain framework with long-term strategic partnerships. We actually burned a lot of CapEx. I don't know if you remember this, but our free cash flow took a little bit of a hit in 2021, 2022. We put in big investments in substrate supply, backend, CoWoS, real strategic partnerships. So all that is now in place, right? And we got some of that stuff set up through like 2030 and beyond. That's really helped us. But also just our proactive approach, how we manage our DOI. So we went from pandemic, and we had a little breather with the interest rate little scare, right, which caused demand to drop. And then we immediately was ChatGPT, right?
All of a sudden now, like you said, we've tripled our business, right, from that time in data center. We tripled our business. We actually got that much extra supply. This past year, if you actually look year over year, we had one of the strongest industry growth, right, of any company. We were able to go get that supply. Yes, it's going to continue to have constraints and bottlenecks. I think the way we set it up, I mean, by the way, I also put in place a five-year supply forecast for all of my strategic partners four years ago. We actually update it annually for them. We meet with them. I've been giving my suppliers this view.
And if I look at 2026, the numbers that I gave them like two, three years ago were about what we need. Mix is a little different, but total revenues, Harlan, we've been driving a $10 billion revenue company here. I know there were times when it looked like it wasn't going to happen and who knows and what's going on with Marvell, but we've been planning for success this whole time. I've been telling our supply base, this is where we're taking the company. And I've also given them, now you can imagine, they've got five years from now, right? And so the one reason we've been able to get what we need is they believe in us. We've delivered to what we told them we would do. And they view us as one of the fastest growing opportunities in semis of companies to bet on.
We really have a different approach. W e're going to manage all that. But yeah, you name your thing, and there'll be some supply constraint here or there. We're going to work through it. But all the kind of outlook I've given you is predicated in the belief as we believe we'll have line of sight to go get what we need.
We're just about out of time, Matt, but I wanted to leave the last couple of minutes for any parting remarks that you have to your investors or potential investors.
Well, yeah. No, first of all, thanks for joining. And happy New Year to everybody. Again, the company had, I think, very strong results last year. I think we have a very compelling outlook for the next couple of years. I think what you'll see from us, and you started to see it the back half of the year as things started to settle down a little bit on the custom side, is we've really built this just incredible networking interconnect franchise, right, that is really in the very early stages of growth. And I think that if I had a message for investors, I think that got lost in the, can you ramp the one socket, which really started back in like 2023, if you remember, right? And I had this call with you two years ago.
And in the December 2023 earnings call, the stock was down, and people said, Matt's not confident in the ramp of the XPU. And then we had the call in January 2024, and I'm like, and we had the chip, we had just gotten it back. And we brought it up over the break, and it turned out it worked first-pass silicon, and we got our forecast. So things have improved. W e really kind of got in that story, which I get because it was such a big growth for us, right? It took the company from $5 billion -$6 billion to like $8 billion. That's great. That's in motion now, right? All that is in motion. We had our AI event, our investor event in June. We laid it all out for you guys, very consistently executing against that.
I think what you'll see and what you should get excited about as we unpack it, and the acquisitions are just kind of part of this, is really where we're taking our networking story. Scale up, scale out, AECs, retimers, switching, interconnect, DSPs, LPOs. Name the alphabet soup, we will be there with the leading product. Pick a new acronym. We'll be there. So you should expect a very strong year from the company and a very aggressive year in terms of product development, customer engagement. And we're going to go drive it really hard in 2026. So it's great to see everybody. Thank you for the time.
Great. Thank you, Matt. Thank you, Ashish. Looking forward to another strong growth year ahead for the team.
Yeah. Thank you.
Thanks, Harlan. Thank you.
Okay.