Marvell on What Comes Next for AI Infrastructure

Ten years ago, data center business made up less than 10% of Marvell's revenue. Today it accounts for roughly 75%, and Matt Murphy calls that shift 'betting the farm.'

At Six Five Summit: AI Unleashed 2026, Matt Murphy, Chairman and CEO of Marvell, spoke with Patrick Moorhead and Daniel Newman about what’s next for AI infrastructure now that compute is no longer the only bottleneck.

Murphy's central argument is that connectivity, not compute, is the new constraint the industry is only beginning to confront. Marvell's 2021 acquisition of Inphi positioned the company early for the shift from electrical to optical connectivity, and its 2026 acquisition of Celestial AI, led by CEO David Lazovsky, extends that bet into co-packaged and near-package optics ahead of what Murphy expects to be a wave of multi-rack, then multi-data-center, scale-up.

The memory market tells a similar story. Marvell's CXL-based memory expansion technology predates the current memory shortage but is now seeing faster adoption as customers look to stretch existing memory further. On custom silicon, Murphy says Marvell's "XPU Attach" components, memory expansion, NICs, and security silicon sitting around the accelerator itself have quietly become a bigger opportunity than the accelerator socket everyone talks about, spanning 15 to 18 different hyperscaler designs.

Murphy's read on where the AI buildout stands: early innings. He points to NVIDIA's $2 billion investment in Marvell this March, tied to NVLink Fusion, as a signal that the ecosystem is still expanding rather than consolidating around a single winner.

Key Insights:

🔹 Connectivity is the next bottleneck after compute and memory, according to Murphy, as AI systems scale from single racks to multi-rack and eventually multi-data-center clusters.

🔹 Marvell's 2021 Inphi acquisition anticipated the shift from electrical to optical connectivity; its 2026 Celestial AI acquisition extends that bet into co-packaged and near-package optics.

🔹 Marvell built memory expansion technology years before the current shortage based on the CXL standard. They're now seeing accelerated adoption as customers try to stretch existing memory supply further.

🔹 "XPU Attach" custom silicon spans 15 to 18 designs across four major hyperscalers, worth up to $2,000 in content per accelerator

🔹 Murphy calls the AI buildout "early innings," pointing to NVIDIA's $2 billion investment in Marvell this March as a signal the ecosystem is still expanding.

Murphy's bet is the tech that’s still maturing–connectivity, memory disaggregation, and custom silicon, are what determine how far the current AI buildout can actually scale.

Watch the full video at sixfivemedia.com, and be sure to subscribe to our YouTube channel so you never miss an episode.

Catch the rest of Six Five Summit: AI Unleashed 2026 Day 2 coverage at sixfivemedia.com/summit.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology audiences.

‍

Matt Murphy:
As far back as 2020, we were saying this: basically every cloud's different, every cloud's going to be its own market, its own architecture, and everyone, ultimately, all the pieces of it, will require some level of customization. We said that five years ago, and now it's happening.

Patrick Moorhead: 

Welcome to Six Five Summit 2026. It is AI Unleashed. It is day two. I'm 

Patrick Moorhead, joined by my bestie, Daniel Newman. Today, we're kicking off things by talking about X infrastructure behind the next phase of AI and discussing what will it take to keep pace with the scale of the build-out. We're joined by Matt Murphy, CEO of Marvell. Matt, welcome to Six Five. You've been on the show before.

Matt Murphy: 

Hey, thanks, Pat. Great to be here. Hi, Dan.

Daniel Newman: 

Hey, Matt. It's been a minute, but you're a wily veteran of the show. So it's great to have you back here. You kind of heard Pat in the setup. I mean, look, it's all the things. And if you look at Marvell's, you know, your MNA and your build over the last few years, it seems like you were kind of playing all the right cards. You know, as we keep talking about, where's the constraint? And you guys are playing in all of them. Look, let's start at a high level. I mean, we are in the middle of the largest infrastructure build-out in history, probably the largest technological revolution that any three of us young men will experience in our lifetime. You know, from where you sit in the ecosystem, right in the middle of the action, curious kind of what is your overall observations of this transformation that's going on? And do you think people actually underestimating this even still as big as this is getting?

Matt Murphy: 

Yeah, well, again, great to be here. And I think a couple of things to think about. I think the first is we've been on basically a 10-year journey here, right, at Marvell. We made a pivot to what we called the data infrastructure market, which we kind of named 10 years ago. That really didn't exist as a sort of a semiconductor end market. But the belief we had basically was that all these millions and millions and billions of units of devices that had shipped and had created data, we're going to create a whole bunch of data that was going to need to get sorted through and monetized and ultimately transmitted, moved around and stored. And at that time, the advent of cloud computing and data center technology was really taking off. So that's where we pivoted the company. It was less than 10% of our revenue back then. It'll be 80 plus percent of our revenue, not in the too near future. The company has grown over 5X over that period. It's not a new thing for us, but I think to your point, the AI application became the killer app of data infrastructure. we're at the top here or like, you know, can it keep going? I mean, we've, I think we've all felt that way since chat GPT dropped back in the end of 2022. So from our standpoint and what I see in the market, you know, being in this business day in and day out, we are still at the very, very early stages of the deployments. And, and even more importantly, I think the very early stages of really having, as a broad ecosystem, the technology required to truly scale AI to the levels it needs to. And we can talk about that, but Pat referred to it at the beginning. You know, you had the compute, and that got all the attention, and it was sort of like, who can make the best GPU and XPU and custom ASIC, and we could talk about all that, and processor. And then the memory and the storage has really been sort of a pronounced super cycle, if you will, that has been sort of unprecedented in the last year or so. But what's coming next and what we're in the middle of is now, to the point, the connectivity that needs to get unleashed and the interconnect that really enables the memory, the compute, and all of the data processing and memory processing that's happening to actually move between the chips, within the racks, across the pods, up at the upper layer of the network, scaling across data centers, and ultimately back to where the consumer is benefiting from this. So we're at the very early stages of that, and I'm happy to talk more about it.

Patrick Moorhead: 

Daniel referred to kind of the early bets that you made and photonics. I mean, you absolutely crushed it, right? You acquired Inphi five years ago and you're looking very smart for doing it and being able to build that out. Let me ask you this, and a lot of it's around the debate of the copper wall and things like that, but what did you see then? And how do you see the transition from electrical to optical connectivity playing out for here? There's a lot of talk about CPO is the ultimate destination, the three versions of that. But how do you see this? How do you see this playing out and when?

Matt Murphy: 

Yeah, you're right. We closed Inphi in April of 2021. We announced it in the fall of 2020. And that had been a company actually since December 2016, which was about five months after I became CEO, that we were interested in. I had known the company for some time. And the stars aligned in 2020. But that was at a moment of inflection where inside the data center, there was a massive transition happening. on the optical side to PAM-based DSPs, which basically was the modulation technology and the architecture that was required to really move to the next generation of high-frequency communications over optics. So we got a great DSP business. But by the way, it's not just that. Within those modules, you also have to have high-performance analog. which is typically silicon germanium based TIAs and drivers. And we're going to come to that next because that technology, which by the way, I was very familiar with from my Maxim days. I mean, we were one of the pioneering companies actually at Maxim to develop silicon germanium technology in our own fabs. And I managed those product lines. I had these kinds of components 25 years ago. Now, a much, much more crude state of silicon germanium that exists today, But these broadband analog components, actually, we'll get into this later, have become a key component, not only of the optical-based DSPs, but now when you go to NPO, which is near package optics, and then CPO, for linear drive, you're going to have to have the silicon germanium technology. So we got that from EnFi. And then on top of it, we got silicon photonics technology, which was used, designed by, brought to production by Marvell and EnFi together. And that was used in long distance communications called DCI, which is for between data center, you know, long reach applications. But we've shipped millions and millions of units and have 15 billion hours of reliability data over the last decade in silicon photonics. So now we're sitting here at this advent where all of a sudden, silicon photonics, broadband analog components, DSPs, all of that is the fundamental technology you need to not only build the scale out network. But as you go to scale up and then even scale in, those are the key building blocks. And I think people are now realizing Wow, first it was moving to DSP-based optics, now it's actually moving directly to CPO and NPO. We've been doing this for 10 years. And so Inphi was a part of it. And to accelerate that, we did another acquisition at the end of last year of Celestial AI, which had a very, very competitive purpose-built CPO and photonic fabric solution. that really we then combine those two teams together. So we've got kind of the best of both worlds. We've got 10 years of development on our side, on DCI and then NPO solutions, Celestial coming in with CPO. And basically, we have, in Marvell now, the most broad, diverse, and competitive silicon photonics and optics team out there. which is by the way not a standalone product because you actually want to connect those optics to your XPU. if you're gonna go do custom silicon on one side, or we can work with third-party or merchant GPU companies to integrate the technology. And then to move the data around, you need to send all the data through a switch, which again, we can communicate, we can attach our optics to ethernet-based switches, UAL-based switches, or even NVLink-based switches. So all having all these pieces under one roof is going to prove to be very, a very compelling thing for our customers because everyone's trying to figure out how to take advantage of all these diverse technologies that are required to really drive, you know, thousands of GPUs and ultimately, you know, hundreds of thousands of GPUs and millions of GPUs to communicate with each other.

Daniel Newman: 

And the clusters are just going to keep getting bigger, aren't they? Dave, by the way, the Celestial CEO joins us for a session here at Six Five Summit, everybody. So make sure that you tune in for that.

Matt Murphy: 

He'll have a great perspective on that. And he's leading that entire combined effort for us now. So Dave doesn't just run the celestial business, he runs the Marvell silicon photonics, he runs the DCI module business for us, and he's responsible for our entire switching platform. So we have one executive that's kind of got the end-to-end ownership of this. So yeah, he'll be very exciting to listen to. He's right in the middle of this entire technology revolution.

Daniel Newman: 

By the way, there's another… bottleneck that I think Marvell is addressing or attempting to address, and that's the memory wall, right? I mean, every, you know, connectivity is a big challenge. Memory is also a big challenge, especially with the scale of inference. Just in your viewpoint, why is this such a hard problem? And, you know, is, you know, you kind of hear about everyone's working around the memory wall or, you know, architecture, like, what do you see there? Is that, is that happening?

Matt Murphy: 

Yeah, a huge amount of activity there. And it predates the memory shortage, by the way. So the memory expansion technology we fundamentally possess now was all organically developed at Marvell. So this was something that we decided to do on our own. And the first effort we made there was with kind of an industry standard technology that emerged about five years ago called CXL. and basically CXL at the time, this is pre-AI, guys. It was envisioned for industry standard servers. And Pat, you know this business very, very well. And you remember traditional CPUs, x86 and ARM, all have a fixed number of memory controller ports. And so what was happening even in standard servers is when people wanted to add more DRAM and more memory, you'd have to buy more CPUs. which didn't make a lot of sense. And so effectively people wanted to get put, put any created standard. And the idea was you could put a CXL memory expander. or even later a pooling device, but basically gang up larger amounts of memory, not have to scale your CPUs with memory. And you could do those things in a disaggregated way, which is where disaggregated memory came from. So that's been happening. Now, AI actually kind of accelerated the use cases for this type of thing, because one, in inference, again, you're gonna wanna have disproportionate amounts of memory attached to your XPU for KV caching, that's one. And then on top of that, so that whole trend is happening. And then on top of that, with the memory shortages that people are seeing, everyone's getting creative. So we're actually seeing a faster adoption now of customers that were already designing us in on some of our solutions, trying to go faster, because basically it obviates the need to buy as much memory as they thought before, if they can put some level of memory expansion capability in between. So that's become very strategic to us and we've got multiple customers on either custom-based memory expansion or we have a whole standard product line of CXL expanders, CXL switches, and CXL retimers. So it's an end-to-end kind of offering we have. And I think people were wondering, is this really going to take off after the x86 kind of application slowed down when AI hit, but it's actually, it's on turbo charge. And we've called this out as like a, you know, like a billion, multi-billion dollar kind of business for us in the future. So it's become a real thing. And we have absolute product and market leadership here.

Patrick Moorhead: 

Yeah, it's interesting. I think we met 10 years ago, right after you started. And back then, I think Marvell, 10% of its revenue was data center. And here we are today with memory solutions, connectivity solutions, and a Computex Jensen calls you out as the next trillion dollar company. I guess more editorial, congratulations. You keep making the right moves. I want to talk about custom silicon. We've chatted about this a lot. You do a lot of it. You have a lot of IP in there. For those who don't live and breathe it, like us on here, what does it mean and why do hyperscalers, why do they continue to invest in it?

Matt Murphy: 

Yeah, it's interesting how that's evolved. We got into this business through an acquisition we did in 2019 of a company called Avera Semiconductor, which was a spin out of Global Foundries, which had all of its roots. And it was IBM's original custom silicon design team, which was a very successful team. They really needed to to be able to operate at the leading process nodes. And so when GlobalFoundries decided to focus on mature and specialty technologies, they spun the group out. We put them right on TSMC and on our technology platform, and we pointed at this data center market. And we ended up winning a number of custom silicon sockets pretty quickly in the data center. And Pat, to your point, there was a debate just a few years ago whether whether these would ever go to production. There was a debate whether custom XPUs could actually penetrate a reasonable part of the market. If you fast-forward to today, companies like ourselves and a few other large peers have taken into production very complex custom XPUs that are being deployed and now running training workloads on them or inference workloads on them, and customers are using them. We had sized this two years ago that maybe you'd have like 25% penetration of custom silicon versus merchant. I think the prevailing view is that number is probably going to be higher in terms of units. And the reason that happens is customers have found reasons why they believe, for their own workloads, which they know better than anyone else, that for a portion of their fleet, they see a lot of economic advantages and technical advantages and architectural advantages to doing some of that themselves. Now, the notion that custom is going to take over all of the market, I've been not of that opinion consistently for a long time. It will coexist. It'll be a part of the market. That's where we come in. By the way, you mentioned Jensen and Nvidia. They did do a $2 billion investment into us earlier this year, and part of that agreement was actually us being able to use a lot of their very rich IP in our custom products, so we can actually interoperate with their merchant products. So, you know, they're not fighting it either, and they see that this is the way the market's evolving, and they're just trying to make sure that the ecosystem ultimately supports the best possible technology that gives the best performance and returns for our customers. We're very active in this area. What gets talked about a lot is the accelerator itself. There's a lot of excitement about that. A lot of people want to talk about that. A lot of articles; pretty much every day you'll see something. And we're in that business, and customers rely on us for that. But there's another category that we basically called out and defined ourselves, which we called XPU Attach, which is all of the key custom silicon components around the XPU, which some of those, Daniel, are memory expansion. But we also see the NICs, or the network interface products, also being customized, security products. I can go on and on. And so that whole category, which was looked at as, oh, maybe that's just too small, it's too nascent. Our design wind momentum here is significant. And if you think about these different ecosystems that have now developed, the TPU ecosystem, the Tranium ecosystem, the MTIA ecosystem, there's several of these now. And by the way, we also do XPU attach, which can work with somebody else's custom XPU, or a merchant GPU, by the way. Some of these get deployed on servers from both. And so that business is doing extremely well for us. And customers see real value. Because ultimately, and as far back as 2020, we were saying this: basically, every cloud's different. Every cloud's going to be its own market, its own architecture. And everyone, ultimately, all the pieces of it will require some level of customization. We said that five years ago, and now it's happening. And I think we said a year ago at our AI day, we had like 15 or 18 different designs across all four big hyperscalers, plus others of these XPU-attached products, which can get up into the $500,000, $2,000 content level per accelerator. And this is all Marvell IP, by the way. These are chips that we design, build to spec, typically. It's not a lot of RTL or design from our customers. Sometimes it is. And so we can add a lot of value here, especially in conjunction with the XPUs we have, but also just the broader the broader business we have with these hyperscalers.

Daniel Newman: 

Yeah, we've entered the era of abundance, I call it. There's this kind of perpetual narrative in the market that someone has to lose for someone to win. And I think that's been wrong the whole way up. We continue to revise our forecasts up, Matt, but we have close to 700 billion of cumulative just XPU between now and 2030. And I think that number continues to rise with every quarter when we revise it; the number keeps getting bigger. You know, I think we've- I mean, we did a recent data center CapEx forecast. I think it went from like 10.7 trillion cumulative between now and 2030, Matt, to over 12 trillion in just three months, just as we keep watching this grow. So, and I think your attach story got missed for a long time, but I do think the market's beginning to appreciate it, which is funny because you've been doing it and saying it all along.

Matt Murphy: Right. Well, I think it's funny. Some of it actually, because of your, point you made, there was a point in time, I think, that's gone away, but where it was viewed like there's three sockets, it's a one or a zero. If you have one, you're great. If you don't have one, you've lost. And then everything else is an excuse. So I actually think in some ways when we articulated our XPU attach strategy, people thought it was like, hey, just go look over here. And we're like, that's fine, but we're going to kind of do it like we always do it, guys, very consistent, talk about our business, frame the opportunity, go execute against it. And if I look over the last decade, we've been very, very consistent and very accurate in how we sort of frame these things. So there was no other conspiracy theory on this. We basically said, look, There's XPUs, and we're doing well there, and here's how big this is. There's all these other sockets we've won, and it's not one or two; it's like, you know, a dozen-plus, couple dozen. And they will generate, you know, meaningful revenue for us. And they're also very strategic, right? Because they ultimately are like a very bespoke part of our customer's architecture that gives them advantage in what they're trying to do. Then when you combine that with our strength and connectivity and switching, then you're talking about a very nice architectural end-to-end approach we can take and share a lot of that IP as well between all these different solutions. Customers see that, especially in an era where you can't miss; you've got to execute on time, and you've got to have large, reliable suppliers to count on. It's absolutely not a one-guy-wins, and one guy loses. This thing is absolutely at this point; this market is not a zero-sum game. There's enough market growth that the key participants will all, I think, do really well.

Daniel Newman: You can all win, right? It's the air of abundance. You guys are in the tray, in the rack, in the data center, and across the data centers, and there's opportunities in all of those. So, gone pretty deep, and appreciate that, Matt. As we sort of wrap this up and get into our day here at Six Five Summit, I want to ask you a bigger question, just a broad viewpoint. There's debate constantly, you know; Pat and I go on CNBC or different places, you know; you do the same. People will say, what inning are we in? Or how early are we? Or how far into this AI revolution are we? You know, I've proclaimed that we're still in the pregame tailgating. I've heard people go on and say we're in the third inning. Just kind of curious, like, where do you think we are? How early or late is this? And how would you define, how would you answer that question?

Matt Murphy: 

I'd refer to it as early innings. You know, I don't know if I can get that precise, but clearly there's momentum, right? There's, you know, things are happening. And, you know, we, I mean, we, when ChatGPT dropped, you know, and it was sort of like early 2023, we were trying to figure out what's our content, you know, how much, how are we attached to this? I mean, I knew, I knew we had design wins. I knew our content because I knew even back when we acquired Inphi, I remember doing diligence on these guys in 2020, and they showed me the whole team, all of their GPU clusters, they had won. I mean, we saw them all. But how do you quantify? That was definitely a pregame early inning, right? No question. And then, by the way, it was crazy at that time. We said in May 2023, we're going to do $200 million this year in AI and $400 million next year. And it was like our stock went up like 40% in one day on that. I mean, we're doing, you know, our data center business is, you know, like north of $2 billion a quarter right now, just to give you a sense. So we're definitely progressed. But when I look going forward, the TAM opportunity is massive. And the technological advancement right now is still early. I mean, the real big one is, this is where the connectivity comes in, guys, just to kind of wrap this. You still haven't seen mass deployments of GPUs and AI accelerators scaled up. You haven't seen it. This is all in front of us. I mean, think about the compute and memory power that's going to get unlocked when you can gang up in daisy chain now, multiples of DPUs inside a rack, multiple racks and pods together. Scale across is one that we didn't really talk about, but that DCI application I talked about, which was just sending data traffic between data centers. You're going to be able to fairly soon coherently connect up clusters in different data centers and have them operate as one. So this era of connectivity, I'm telling you, is going to unleash a whole new wave of innovation. It's going to enable new use cases. It's going to enable costs to come down, performance to go up. None of this has happened yet. So when you hear about, oh, the scale-up market, it's going to be big. Yeah, because it's the next way you can actually drive the scale of compute that's required by the AI market. So that's why I think it's still very early. And we're still looking at right now, like, for example, CPO. Give you one last one. Yeah. That's coming. And we have certain customers that are going to adopt it. But what's hit us in the last six months is that NPO, or Near Package Optics, probably will hit first in a bigger way, and then CPOs coming. So it's all coming, but it's just not going to come at once. So I think there's many, many years in front of us here. And I mean, I'm just getting off of our annual strategic review. We do it once a year, this time of the year, every year since 2016. I was CEO for five weeks, put it together, and reviewed the whole portfolio. Deep dive. I'm doing it right now. I'm telling you, I've never seen anything like this in terms of the TAM in front of us, all the solutions we can go after. From our standpoint, we're very early innings in what we can go do as a company, but also where the AI market is in terms of its technological advancement relative to the silicon that's required.

Daniel Newman: 

It's a great answer. So I'll submit it for our audience that if the game is really long, if you're willing to acknowledge that this is like a cricket game that can go like two days, it's the early innings. If it's a shorter game, maybe I was right, and we're in the pregame. I'm not putting words in your mouth. But the fact is, the utility, like where actually industries and stuff are using it, is really just getting started. The build-out is probably, like you said, a little bit further along. Matt Murphy, chairman and CEO, thank you, Matt, so much for joining us.

Matt Murphy: 

Yeah, great to see you guys. Thanks for having me on.

Daniel Newman: And everybody, stay tuned. Day two, it's on. Stick with us.