Reinventing Silicon for the AI Era: Marvell’s Vision for Compute & Connectivity

How are advances in silicon and connectivity fundamentally reshaping the future of data infrastructure as AI accelerates?

‍

From Marvell Industry Analyst Day 2025, hosts Patrick Moorhead and Daniel Newman are joined by Marvell’s Chris Koopmans and Sandeep Bharathi for a conversation on how Marvell is reinventing data infrastructure for the AI era. They explore the macro trends shaping the industry, Marvell’s strategic approach as connectivity becomes as vital as compute, key partnerships, like their latest with Celestial AI, and the roadmap for Co-Packaged Optics adoption—all underscored by insights on staying competitive in a dynamic AI and infrastructure market.

‍

Key Takeaways Include:

🔹Macro Trends in AI and Infrastructure: Marvell’s perspective on data infrastructure shifts and how the company is responding over the next 3–5 years.

‍

🔹Connectivity and Compute Strategy: Insight into Marvell’s core strategic pillars, especially as connectivity becomes equal in importance to compute.

‍

🔹Co-Packaged Optics (CPO) & Copper: How CPO will impact traditional copper and Marvell’s product roadmap, plus implications for the broader market.

‍

🔹Marvell & Celestial AI: The vision behind Marvell’s collaboration with Celestial AI and how it aligns with Marvell’s long-term strategy.

‍

🔹Navigating Industry Skepticism: Approaches Marvell is adopting to stay ahead and address uncertainties in a crowded AI landscape.

‍

Learn more at Marvell.

‍

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

‍

‍

Or listen to the audio here:

Disclaimer: Six Five On The Road is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded, and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

‍

Patrick Moorhead:

The Six Five is On The Road here at Marvell headquarters. It is Analyst Day 2025. Daniel, it's all about AI, and it's amazing that infrastructure is leading the charge. Forever, it was always, you know, infrastructure, it's a commodity. Software is ruling the roost and driving everything. AI hits and here we are.

‍

Daniel Newman:

It reminds me to my 2019 op-ed that I wrote the last week of 2019 with my trends for 2020 when I said semiconductors will eat the world.

‍

Patrick Moorhead:

There we go.

‍

Daniel Newman:

I got one right. I was maybe a little early but then you know we Over the next five years, Pat, I mean, gosh, have we spent more time in our lives between media, television, you know, talking to the industry, talking to investors. Everybody wants to know about what is going in what socket. And that's going to be a big part of the conversation today.

‍

Patrick Moorhead:

I've always loved chips. having worked for a few companies there. But listen, one of the debates that always happens during these inflections is what part of the stack is most important, right? Is it compute? Is it storage? Is it memory? Is it networking? And the reality is long-term, they all matter, and they all have to be strong. But it's very clear where we are right now that, of course, compute matters, but connecting all of that compute matters as much, or I think some would argue, even more. And that's really the topic of today. We have two amazing leaders. They've been on here, on the show, on The Six Five before. Chris and Sandeep, welcome back on the show. Thank you. Great to be here.

‍

Chris Koopmans:

Yeah, thank you. It's exciting. Yeah, it's great to be here, Pat.

‍

Daniel Newman: 

You set that up really nicely. walked away feeling like there's so much more to the story. You know, Chris, you and I have been sat down numerous times on the Six Five here. I feel like we kind of end up having the same conversation a lot. But I feel like that's because it's so important to remind the market that all of these things really do have to come together. But connectivity is a big bottleneck right now. I know people like to talk about compute, and for three years we've obsessed about it. And then, you know, we talk a lot about energy, but another one is we have to move all this data around. So Chris, maybe start out just talking about what are those big macro trends, the big three to five macro trends that you and Marvell are paying a lot of attention to as it pertains to data infrastructure?

‍

Chris Koopmans:

Yeah, I mean, if you think about data center scale computing, which is what we're talking about here, accelerated computing is really data center scale computing. One accelerator or one XPU really can't solve the problem. It can't do all the calculations necessary for any of the jobs in machine learning today. So you need tens, hundreds, thousands, millions of them working together. And that becomes a connectivity challenge. So data center scale computing is fundamentally a connectivity challenge because the computation, the bottleneck, is often not the compute element. It's often how fast can the communication between processors and between processors and memory actually happen. And that's what we try to solve here every day at Marvell. And it's not just a one solution for everything. There's every distance, every speed, every technology, copper, optical. You know, we're talking about inventing new technologies to move data in bits over a millimeter is a very different technology from something that needs to move over a thousand kilometers. And ultimately, what we're seeing at Marvell is just an explosion of innovation that is driving connectivity forward. And those solutions find themselves in every AI system in the world. I mean, you could say today that there is no model that was ever trained. There's no data center in the world today that's not running Marvell Silicon in connectivity.

‍

Patrick Moorhead:

Yeah, the amount of technology that you laid out today was awesome. I had to catch my breath a few times because it was so broad. But Sandeep, one thing I'd love for you to do is maybe simplify it strategically. What are some of your core strategic pillars as connectivity becomes as important, I would say more important than compute, but as important as compute?

‍

Sandeep Bharathi:

Yeah, excellent question. So from a core pillars, I mean, connectivity that just like was described is even more important that its data always needs to get moved. It cannot be static. The innovation that we need to bring to our connectivity portfolio, whether it is on the scale up switches or scale out switches, and more importantly, all the connectivity products that we do have on PAM-4 DSP or optical DSP products, the TIAs and drivers that drive these signals, the error correction capabilities. Then we talked about data center interconnects, which is across 20 kilometers and beyond. And then there is, those are all coherent modulation techniques. So, at a very high level, there is a method for the movement for shorter distances, like Chris was saying, one meter, it's very physical-based, Serdes-based. Then you go into algorithms for error correction for larger distances. Then you look at optics, you know, copper and optics need to coexist, so we need to invest in both copper as well as optics, and you heard some of that in our industry analyst day. The pillars being, if you take a look at it and bifurcate the segments, it's really by distance, and it is whether it is copper medium or optical medium, and what investments we need to do for all of these aspects. Those drive the innovation that informs our product roadmaps.

‍

Patrick Moorhead:

I always pause over the past year or so when I see I think of scale up as an example as a few feet or something, or it's in the rack, 11 feet or something. I know I'm just making up numbers, but now it's 50 meters, which is absolutely crazy because essentially you can have a giant model that's the one memory planar surface that extends over that. breaking new ground. And when you showed the Surtees roadmap and how it used to be every four years, right, like clockwork, and now it's just cha-chink. Everything is up and innovating every part of that cycle. It's truly impressive.

‍

Daniel Newman:

It's really interesting too, as you were catching your breath overall with what was presented, just the breadth of everything that's being offered. I think you hit it, Pat, on scale up, scale out, scale across. This is such a massive opportunity. We've talked a lot of times about this TAM growth. You showed a great slide early on, Chris. You had your TAM in 23, and then you're like, oh, this is a big number. How are we ever going to get there? And then you had 24, and it was like, I think you said it was about $400 billion in CapEx growth over just two years. And now the forecast, you hear numbers being thrown around of $2 and even $3 trillion by the end of the decade. I mean, you guys have to be licking your chops, preferably speaking, about that opportunity, because you really do service you know, not just that XPU socket, but all these connectivity sockets. Was it two-to-one, three-to-one opportunities? And speaking of, you know, just the overall growth, you guys have been growing really well organically, but you've also been really acquisitive, you know, over the years, whether it's Cavium, whether it's InFi, and many other acquisitions have come along the way, and now you've announced another one in the optical space, Celestial. Talk a little bit about Celestial AI, why you made that deal, why that's the right partnership. I think, was it 5.5 billion was the number on that deal? All in with the earn out, yeah. Well, I mean, they're going to hit the numbers, right? Absolutely, absolutely.

‍

Chris Koopmans:

Beat them. So, yeah, so it's a good, maybe I'll put it into the context of other acquisitions that we've done. So, in 2021, we acquired Infi and Inovium, and they built the Scale-Out Optics and Scale-Out Switching. Scale-Out is the part of the network that is connecting across the data center. It's sort of rack-to-rack communication. And it's connecting, you know, hundreds of thousands of XPUs together. And that business, when we bought it, you know, was in the, it was probably, put the two together, was probably doing something like $600 million. We've grown that like multiples over that period, more than 5X over that period of time. Now it's doing, you know, more than $3 billion this year, as it's grown, you know, 50% plus CAGR for the entire period of time since we built it. That's the scale out portion of the network. Excellent business, excellent team, excellent roadmap. You heard from them today. It's going to keep going. Now, the scale up today is sort of more like connecting a fabric of XBUs. You know, like you mentioned, it's getting bigger, but you know, within a tray, within a rack, or maybe a couple of racks in any order of magnitude higher bandwidth is the way to think about it to connect those together. And today, that's almost all passive. There's no switch, there's no optics. Marvell is organically investing in the switching for UA-Link and vSUN for that network. And Celestial AI really brings the optical interconnect fabric for that scale-up network. Because at some point, copper must always give way to optical, just in terms of how much bandwidth you need to push over a certain distance. And Celestii has invented technology that is purpose-built for that mission. It was purpose-built and designed from the beginning for that scale-up network to be the best overall bandwidth, lowest latency, lowest power, to actually enable that to move to optical. And when you do that, you break new grounds. Distance, as you mentioned. Now you can start going to the next rack over and making an even bigger cluster. So that's really what that acquisition is about. We're as excited of that acquisition as we have been of these other acquisitions that I mentioned that have been so successful. Actually, the opportunity for this is probably even bigger than that. So super excited about that. We expect that to close in the first quarter of next year and really start participating together right away.

‍

Patrick Moorhead:

Yeah, you really are, and I choose my words carefully, an acquisition machine in the ability to buy something and then scale it and grow it. And I know this wasn't just slideware, but I liked how Celestial AI magically fit directly into your roadmap. in what you had showed today. And I know it's a lot of hard work to get it over the line, but with all the LOIs and things that they have already, it looks really good. So Sandeep, I want to talk about CPO. It was funny. The discussion about CPO has been around forever, and there was a large GPU conference. that literally it got everybody talking about it. Can you give us the lay of the land about it? Like, when is it really going to be here? And what does it mean to copper, right? Copper has been dying for 10 years, every year. But it just seems like technologies just keep improving, whether it's powered or DSPs, just to keep that going.

‍

Daniel Newman:

Are you calling copper the mainframe?

‍

Patrick Moorhead:

Well, maybe.

‍

Sandeep Bharathi:

I don't know. No, it's an excellent point that you make. CPO's moment has always been, yeah, it's going to come, it's going to come. But honestly, what has happened is, look, copper will be there for speeds that are currently at 100 or 200 gig. It has the right distance and what we call in SerDes parallel and serializer, deserializer for copper, it's really three metrics. Power, how far it can go, that's called the reach. And then it's measured in bit error rate as well, how reliably bits can be transmitted for long periods of time. But as the speeds have gone from 200 to 200 to 400, the distances that it can go are radically less. And at this point, it is an inflection point where some other technology needs to come and take over. And copper had always existed with what we call the beachfront, which is if you have a single die, there's only so many escapes that copper can make within that die size. But what you saw with Celestial AI or even our own efforts on the light engine, it has come to a point where you need to move about 14 terabits from an XPU or a switch. And that cannot be done with copper. So with the 3D packaging advances in that and even innovations in how coupling can happen with the optical fibers, that's very close to the silicon. And that has taken time for making sure it can be reliably and volume manufactured, as well as all the stability that it should come up with, whether it's power stable, thermal stable. And CPU's moment is here because of the build out of the rack part, where the capability of compute has been far expanding, and then connectivity needs to catch up. And optics is right here and right now. And if I just take a look at it, there will be different kinds of modulation schemes and everything will coexist. You've heard many terms in the industry about modulation and whether it is near-packaged optics or onboard optics. All of these will exist because of the reliability that is necessary for the data center operators, as well as cost. So I think it will definitely coexist, but a lot more will switch over to CPO over time.

‍

Patrick Moorhead:

Yeah, one of the key objections that you helped remove, along with the industry, was the reliability question, right? If one of the modules goes down, it takes down my entire switch. Like, that's unacceptable. And I think you had a very good solution. Cost, I think, was the other. I think it just shows kind of different strokes for different folks. People are looking for different types of implementations to take in, if nothing else, differentiate what they're offering. I think that's super important.

‍

Daniel Newman:

It's interesting, he got me thinking of Matt Garman last week at reInvent when we were talking about quantum, and he says, five years away from being five years away. I think Copac's optics is coming. It's coming a little bit quicker than that, and I think it's a really sizable TAM. And that's one of the things I really like about the Marvell business that I sometimes think is misunderstood is the opportunity in every area that you're in, I think the TAMs are somewhat understated because I think as these kind of compute numbers keep getting bigger, people don't as quickly adjust the things that are all around it and then you start to go, oh my gosh, how much more networking, how much more storage, and you look at how much, of a challenge we have right now with memory. We were talking about spinning disk drives and how everything right now that can be built. And so you're in so many of these great areas. But let's talk about the doubters. Let's talk about, Chris, a little bit about the skeptics. Marvell faces it sometimes directly. Sometimes it's an industry-wide thing. I think I did 50 television interviews in the last 45 days where I talked about, is AI a bubble? I'm not kidding. It was like the same question every day. I started to get frustrated. It's like we're still having this debate, but at the same time, as fast as we're building out infrastructure, there are still some questions when it comes to what workload, what are people going to pay for it? But there doesn't seem to be any hesitancy to keep building. How does Marvell sort of think about this, balance this, and make sure you stay ahead of any risk associated with a slowdown?

‍

Chris Koopmans:

Sure. I mean, look, I think you really have to take a long view and ask yourself the question. I mean, the bubble that's most commonly referenced is the dot-com bubble. Well, does anybody question today whether the internet and all the telecommunications infrastructure that connected the world was valuable and useful? No, it's very valuable and useful and it created like trillions of dollars of economic value over the years. Similar here, I think we're building the infrastructure for the next wave of innovation. And that's what this capital investment cycle is like. Now, most capital investment cycles do go through some ups and downs. It's not a straight line up and to the right. There's a And these are not small companies that are spending this and building this out. They're very thoughtful about what they're doing. You're not building AI data centers just like overnight. You're contracting power, land, space, water, energy, like long term to be able to sort of meet the needs of your customers. And you're measuring real time. Is the demand still there? Are customers finding real use cases? Is my revenue growing? Is it profitable? Are they finding use cases to use this and to turn that into revenue for them? So, I think two things are true. Number one, this investment is going to continue to grow over time, and it's going to prove to be very, very valuable over time. Number two, it may not be a straight line. It's more likely to be that, in fact, as I showed earlier, every time we look at it, it's bigger. There'll be a time where when we look at it, it's a little smaller. And our forecasts will modulate, if you will, but it's not going to start shrinking anytime soon. I feel like that's a few years away before it gets smaller, if ever. I think that most, you can pick any investment cycle over time and look at it. They tend to be 10 to 15 year cycles and they aren't all straight and up and to the right. But Our view is that right now, the way you know when you're in an upswing and that it's not coming to an end is when you feel like you're always behind no matter what you're doing. Sure. I mean, I think Sandeep and I come to work every day and feel like we're behind. We've got more bookings. Oh my gosh, we're already behind.

‍

Patrick Moorhead:

I thought you were going to say you were going to look at the 101 traffic.

‍

Chris Koopmans:

Well, you could look at the traffic in Silicon Valley, but ultimately, when you have this immense pull from your customers to do more, deliver more, get it faster, need higher speed, if we could build it, we could sell it, if we could make it go faster. And that's really, especially all of these connections and all of these different links of all the I.O. inside the data center, if we can make these breakthroughs, it'll unlock new value inside the data center.

‍

Daniel Newman:

Well, I think Mark Zuckerberg said it pretty well when he said in this particular supercycle, I would rather be over-invested than under-invested. And I think that is going to be the quote that most of these large hyperscalers, Neo clouds, enterprises, I have not met a CEO or a board yet that hasn't put AI as its top priority to gain productivity and efficiency. So it's a great place. to be as providing the compute, the connectivity, and so many other parts of that stack. Chris, Sandeep, I want to thank you both so much for being part of The Six Five here. It's always great to have you both on the show.

‍

Chris Koopmans:

Thank you very much. Thanks for having us. Great to see you, as always. And thanks for attending our Industry Analyst Day. Definitely. Thanks for having us.

‍

Daniel Newman:

Great day. And thank you, everybody, for being part of This Six Five. We are on the road here at Marvell Industry Analyst Day 2025, December. Pat, this is my last trip for work. Congratulations. Yeah. Isn't that nice?

‍

Patrick Moorhead:

It is.

‍

Daniel Newman:

It's nice for you, too. But there's going to be a whole lot more. So subscribe and be part of our community. We love all of you checking out the content here on The Six Five. But for this show, this episode, it's time to say goodbye. See you all later.

‍