AI’s Next Scaling Test Is Connectivity, From Rack to Region

AI scaling has run into a series of rate limiters: first compute, then memory, and now connectivity. As inference and agentic workloads move more data across greater reaches of the network, the physical limits of copper are moving closer to the silicon. At the same time, data centers are rebuilding links for AI-to-AI traffic.

At Marvell Investor Day 2026 in New York, Daniel Newman, CEO and Chief Analyst at Futurum, sits down with three leaders from Marvell: Sandeep Bharathi, President, Data Center Group; Xi Wang, Senior Vice President and General Manager, Connectivity Group; and Russ Esmacher, Senior Vice President and General Manager, Data Center Interconnect.

Together, they explore why Marvell is positioning itself as a connectivity-first company and how copper and optical technologies are evolving together to address the demands of AI infrastructure at scale.

The panel’s common thread is that connectivity decisions are now made years in advance. Customers deploying 800G and 1.6T today are already co-designing 3.2T with Marvell. Meanwhile, scale-across links between AI clusters are ramping toward volumes the high-speed optics industry has never shipped.

Key Takeaways:

🔷 The copper wall is moving closer to the silicon. Bharathi says AI workloads are pressing against both the memory wall and the copper wall, as the space to move data in and out of chips and trays gets tighter. Optics is moving into scale-up and even inside the rack, while copper remains a reliable medium wherever distances allow.

🔷 Scale-out planning now runs two generations ahead. Hyperscalers are deploying 800G in massive volume and shipping 1.6T, based on decisions Wang says they made with Marvell years ago. Today the teams are defining 3.2T, with 400G-per-lane PAM4 and coherent-lite as the two paths under evaluation.

🔷 Scale-across is redefining data center interconnect. Esmacher describes purpose-built links between AI clusters carrying latency-sensitive, largely inference traffic. He puts the bandwidth requirement at roughly 15x that of traditional data center interconnect. Because these links start as pluggables from day one, they are pulling 1.6T and 3.2T forward.

🔷 High-volume manufacturing is now a design requirement. Esmacher expects scale-across to push a market of nearly one million units this year to well over three million within three years, a volume he calls unprecedented for high-speed optics. Marvell designs for packaging, signal integrity, and yield from the start, a discipline he traces back a decade to Inphi.

🔷 Deployment velocity is what hyperscalers are buying. Wang points to application-specific requirements, generation cycles compressed from three to five years down to 12 to 18 months, and volumes 10x to 100x higher than before. Marvell invests in operations and supply chain and designs out dependencies across foundry processes and laser types. As a result, some parts ship in volume just three months after sampling.

🔷 A complete portfolio lets Marvell support every architecture. Bharathi describes Marvell as “the Switzerland” for these technologies, working across copper and optics, NPO and CPO, and custom and semi-custom silicon. Newman notes that this connectivity footprint extends Marvell’s opportunity well beyond the XPU socket itself.

Infrastructure leaders who choose multi-generational connectivity partners now will be able to scale AI clusters at each new bandwidth generation without redesigning from scratch.

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

Learn more about Marvell at marvell.com.

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.

Sandeep Bharathi:
We have multiple charts on goal. The portfolio gives us the freedom and the capabilities to execute on all this.

Russ Esmacher: 

We've spent the last decade connecting data centers but miniaturizing the technology.

Xi Wang: 

Many years of multi-generational co-design model that we developed with hyperscalers to really understand their need.

Daniel Newman: 

Hello and welcome to The Six Five. We are on the road here at Marvell Investor Day and it's been a heck of a day. Great conversations we heard from all of the leaders of Marvell and it's good to be here as The Six Five having a conversation with a number of leaders, a few. that have been here with us before. Sandeep, Ji, welcome back. Thank you. And a new leader that just joined pretty recently, I think, you know, going on seven months, I believe. Ten months. Ten months. All the way there. I just made that up. Seven months, ten months. It's pretty close. Russ, thank you guys all so much for being here.

Sandeep Bharathi: 

Thank you for having us.

Daniel Newman:

 It was a day about connectivity, you have to say that. Matt didn't bury the lead at all, came straight out, said, hey, this market is being shaped. And in our work, in our market sizing, intelligence, and evaluations, we're seeing the same thing. Rate limiters, we've had them throughout this whole build out. Started off with compute, and then it went to memory. And we're definitely seeing network and connectivity as the next big wave. And Marvell, you guys seem really well positioned. So before I jump into all the questions, you guys feeling good?

Sandeep Bharathi: 

Absolutely.

Daniel Newman: 

So I, as I sat there and listened to Matt, I said, you know, there's a lot of really good numbers in there, but that 80 billion midpoint number heading into, you know, 2031 FY, because you guys run on an interesting fiscal, just a huge number, just a huge number. And having followed you guys now for close to a decade, You're not a company that comes out and says these things if you don't absolutely believe it's possible. So congratulations on the progress. I know these are big days. I'm guessing you guys have spent the last 48 hours in a rehearsal. So I appreciate you making a few minutes here for me. Sandeep, I'll start out with you. I kind of started the question there, what I was just saying. AI, and especially inference and agentic, are creating so much need to move data faster. So connectivity has become a central constraint, but it's also a huge opportunity. Talk a little bit about how you're seeing that.

Sandeep Bharathi: 

Yeah, excellent question. Thank you for having us again. It's an excellent day so far. Yeah, Matt said it outright out of the gate, right? Marvell is a connectivity first company. And again, I talked about how the workloads are changing more towards the two fundamental walls, which is the memory wall and the copper wall. And scale-up optics is becoming more and more important into the rack. So it's all distance based. Now, what is happening is the connectivity challenge is coming all the way to the tray as you connect more XPUs, accelerator, accelerator attached, because the real estate that you want to get data in and out of any chip or even a tray is becoming constrained by the physical limitations of what copper can do. And as you look at that distance and the huge amount of bandwidth that needs to go either between Accelerators to accelerator attach or to memory, you need to get the bandwidth constraint in a given location and optics seems to be the definite natural choice. So as a result, connectivity and scale-up optics is not only within the rack, but inside the rack. And what makes us very well positioned to address this is our excellent IP portfolio that you have, our silicon photonics portfolio, all our modulation schemes that we talked about at Investor Day. So naturally, the copper wall has shifted and moved close to the silicon. Of course, copper will always exist because it's a reliable medium over many years, but optics is the name of the game and connectivity is focal and central to it.

Daniel Newman: 

Yeah, one of the things in my read-through is you were very focused today on the word and. In a lot of cases, when new technology comes, it starts to feel like you're asking, choose one or the other. And you saw the and copper and optics. And even when you talked about NPO and CPO, it's like, it depends on what you're trying to do. It depends on, are you focused on highest performance? Are you focused on having highest reliability? And kind of throughout, you were very, just as a company, you were all very focused on, we want the world to understand that the portfolio is very complete. And not everything, you know, it's like they talk about frontier models, right? Sometimes using the most advanced model, it's like driving an F1 car to go get coffee at Starbucks. You just don't need to do that. And I think you guys know that depending on all your clients, you're seeing that. In your portfolio, you know, especially scale out networks. I mean, this is an area that is moving incredibly fast. Bandwidth requirements continue to grow. You know, kind of curious, because your customers are in deployment mode of current generation. But in this world, I mean, they have to be thinking a lot about next generation. How do you help your customers sort of get what they need to get done now, but knowing that they're almost immediately going to have to be moving generation to generation?

Xi Wang: 

Yes, scale-out is definitely a very big opportunity that has drove massive growth for Marvell. Scale-out is the defining moment for all these hyperscaler data centers to train their model, to inference, and also to provide the networking to their customers. And we are in this unique position to support all these hyperscalers with their scale-out need. So you are absolutely correct. They are deploying 800GB at massive volume, but also shipping 1.6T solutions. But those decisions actually are made years ago. And we made those decisions together with them. How to define this system? How to define a scale out? What is the best way to deploy that? So those are decisions made years ago. Now we are working on the next generation and then the generation after that. So today we are working with hyperscalers on defining what is needed for 3.2T generation, as I mentioned earlier today. is we are looking at 400 gig PAM4 or the Cochrane Lite as two approach to 3.2T and those are the day-to-day conversation we have with them and this is based on many years of multi-generational co-design model that we developed with hyperscalers to really understand their need and also like them to understand our technology and then the product breadth and the depths, and then we can build that solution for them, not just for the current generation, but two generations ahead.

Daniel Newman: 

The speed, though, from generation to generation, right, we continue to see that timeline somewhat collapse. But of course, you guys are sort of setting the pace. You know, we talked, the market wants to talk about setting a model pace. But in some ways, what you build is sort of setting the pace for how fast we can go. You know, of course, Many different approaches of how we can train the largest models and brute force moving data from point A to point B, but we all know that efficiency is increasingly in vogue right now. And these trillions of dollars that we throw around, it's amazing. We've gone from an era where we talk about a trillion or more dollars, and we talk about it so, like, passe, so casual. It's just a trillion. We're just 10% of GDP now for the entire U.S. just spending on AI CapEx. But it's going to revolutionize the world. And I think Dave said from the stage, these workloads are profitable. And so I think that's a really important caveat for those that are not seeing it that way. You know, the industry has watched large transport shrink to pluggable. You were holding those up on the stage. There's a lot of show and tell today, by the way. Very proud. It's the geeky semiconductor industry. It's that moment, you stand on stage and you hold that. You have to have proof points so people believe you. Yes, and now you can 3D print them. I'm joking. It was only that easy. We know they're all real. But you know, the demand in the data, it's creating just, it's unfettered demand. I'm just kind of curious on scale across, like how is scale across different than sort of interconnected people are historically, you know, paying attention to?

Russ Esmacher: 

Thank you for having me, first time with you guys. And that's a pretty pervasive question I get all the time, is differentiate between classical DCI, which is so old, it's almost 10 years old, and what's happening in the new application, which is scale across, which is completely redefining what it means to connect data centers, right? So as you alluded to, we've spent the last decade connecting data centers, but miniaturizing the technology, focused on power efficiency, It's just really focused on cost per bit and watt per bit and really done a great job there. But now data centers are not just being connected, they're being purposefully connected AI cluster to AI cluster. And only workloads going across them are AI based. It's none of the compute, it's none of the store, it's not even some of the training, but really a lot of the inference based extremely synchronized specific latency sensitive is driving the scale across. The bandwidth, as I mentioned today, Some range at 10x growth, some are as high as 20x growth. I think it's 15 times the bandwidth of DCI. This is custom-built, one, for pluggables to start with, so we're not going to start with boxes and then go to pluggables. We'll start with pluggables immediately, and the speed of the pluggables bitrate pushes physics to the limit. Coherent DCI has always pushed physics to the maximum. The highest bandwidth, the highest ball rate, highest performance. Now we have to do it in a small package and it's pulling 1.6 T forward. As you saw today, 3.2 T forward. And so our customers need this capacity between the AI clusters. They desperately need the bandwidth. They're putting the money in the fibers. You see the mass investments going in, millions of kilometers of fibers going in. I view it's the responsibility of Marvell to make sure that capacity is delivered the lowest possible watt per bit and cost per bit and the maximum capacity. So very exciting time. scale across, I believe, I'll have a title change here soon. I'm going to talk to Sandeep about this. Instead of DCI leader, maybe scale across leader. So we'll see if that goes through.

Daniel Newman: 

Also a professional career development show. I like that. But this one actually for both of you, Sandeep and Russ. So we just talked a little bit about kind of the picking the right technology for the need, right?

Sandeep Bharathi: 

Yeah.

Daniel Newman: 

And so one of those things really is where copper and optical inflect, where they sort of, you know, how that decision, I guess, is being made, you know, I'm just kind of, you know, want your read is how do the two work together, meaning electrical, optical, right? Because again, it's probably not going to be all one or the other for some time. How do they work in together? And how are you guys deciding how these clusters sort of interact across different connectivity?

Sandeep Bharathi: 

I'll take the first part and Russ can address as well. We have a front row seat to all the hyperscalers, how they're designing these ecosystems, right? It doesn't matter what the distance is, whether it's scale across, scale out, scale in, scale up, right? And where the technology is, you know, Russ's technology team has capabilities to understand the silicon photonics piece and how it interacts with the rest of the ecosystem. And if you want to do a stack up, They need to really say, okay, how does silicon photonics and CMOS integrate together, and what kind of copper hybrid bonding needs to go on, right? That's at the basic level. It's the physical level. Now, how do we integrate lasers inside for the optics piece? Now, the distances are shrinking on copper, and then it becomes not only an architecture play, whether we need a DSP or we just do a linear engine. All those with the customers, we come up with an architecture, and we look at it three to five years out. It's not decided on the fly because we understand where the use cases are and For high volume production, we need to really understand the technology so it can be deployed in the millions, billions of field hours. That's the way we make these decisions. So technology, markets, and customers are all interplaying together.

Daniel Newman: 

And it was interesting because you've talked about this throughout, but you really are servicing now all four. Yes, absolutely. And again, there was an obsession that we noted a lot in our research that people were very focused on Marvell as the XPU provider. And there was such a miss for such a long time that, you know, TAM and opportunity beyond the XPU itself, right? And I think the Google deal did a really nice job of sort of, well, it confused some people initially, like of actually sort of defining how big the opportunity is. Because you can be coexisting with a different XPU that's not necessarily one that you make, and it's still beginning a many, many of billions opportunity for the company over a very long period of time.

Sandeep Bharathi: 

Absolutely.

Daniel Newman: 

I know, Rush, you may want to add something there.

Russ Esmacher: 

Great question. So one piece of that is, volume in manufacturing. So I mentioned today the unit volume driven by the scale across really accelerates what would have been a nice DCI growth to almost a million units this year for a total market. And it really puts it into hyperdrive and takes it well north of three million units by the next three years. While she's business, those aren't really big numbers. Those are quaint numbers for she's business. For the high-speed optics, it's unprecedented. And so in the past, when we were doing coherent optics as an industry, it was, can I get this much of an extra dB out? And that's a really important thing. And the question I believe that Marvell asked properly a decade ago, and even within PHI, is how do I make these at high volume? What's the packaging? What's the signal integrity? How do I think about making these not one at a time, but making these multiple per minute? Thinking about that as your design parameter versus what's the absolute maximum dB I can get out of it is a conscious decision you make up front. And now with the densification of data centers, the distances by definition are shrinking between data centers. And so it's not about the tenth of a DB that I'm going to get out of it, it's can I make $10,000 a week? Can I make $20,000 a week? I need to service this demand and design from the beginning, the sheet of design when I sit down with the team. They already know this. How do I make this a mass production, high yield, high quality, great supply chain of suppliers and do it very reliably every time? And this is not new muscle that Marvell is just now developing. It's been there for a decade and I feel very fortunate to be part of this team who gets this as the market ramps very rapidly.

Daniel Newman: 

Yeah, a massive portfolio for innovation. I mean, that is for certain. Just talking a little bit about Sandy kind of started alluding to this. The hyperscalers and the companies you work with are the biggest in the world and they're making the biggest bets. Another trillions of commitment that are being made. And they're working very closely with you. They're working very closely with you to make decisions across the portfolio to handle you know, their entire offering. It's going to become codependent on you and this ecosystem. Kind of what are those decisions right now for these deployment cycles that you're super focused on that you think, or at least that you're in the, don't give anybody any private secrets, but you know, like that you're basically reiterating to these folks, like these are the keys you have to get right now. If you want to be sure you're in the best position from the next generation to the next generation.

Xi Wang: 

There's a lot. 

Daniel Newman: 

Give me like top 10. Make it a list.

Xi Wang: 

I will make a short list.

Xi Wang: 

I think from the very beginning is right now the hyperscalers has a very diverse requirement. It's not a one-chip and face-off approach anymore. So you need to design a specific product for their specific applications. So that is a diverse requirement. The second of all is, it used to be the cycle is three to five years. Now it compressed all the way to 12 to 18 months. So every generation coming after each generation 12 to 18 months. So that massively reduced the cycle. The third thing is the volume is much higher, sometimes 10 times or 100 times higher than the previous generation. So with all these requirements, and then they fundamentally go through one simple thing is how you can enable them to deploy with velocity. So that is why Marvell is a very different company. What we do here, as Russ just mentioned, is from the design phase, we built in those parameters. We thought about how we can have high volume manufacturing. We invest in operations. We invest in our supply chain to make sure that velocity is in the right mindset. Give you some example. Some of our parts, after three months of sample, we are shipping in minutes. That's just how crazy the requirement is. But on top of that, we are also doing something fundamentally different. When you look at this as a supply chain ecosystem, there's a lot of dependencies. How, from our design, we can remove those dependencies and making sure that everything that we design is 10 times potential, that we can increase the volume, including the different foundry processes, and a different type of lasers that we can support and a different design approach, right? All in all, together, that is how we work with the Hyperscale customer to understand their need and how we can provide the product that they can deploy with Velocity.

Daniel Newman: 

And the environments have become quite heterogeneous, haven't they? Yeah. But in the end, that really bodes well for custom. And as Matt talked about, semi-custom, which probably worth just a quick shout to, because that obviously gives you guys another unique edge is the semi-custom portfolio that you can put alongside the fully custom cloud-optimized silicon you're developing.

Sandeep Bharathi: 

Yeah, absolutely. I think Matt has used this term before, like we are the Switzerland for all these technologies, which means we can work with multiple customers on multiple technologies, and we are not saying pick one. We have multiple shots on goal, and the portfolio gives us the freedom and the capabilities to execute on all this.

Daniel Newman: 

Absolutely. When those sockets aren't close to the XPU, you also build the connectivity layer that makes those sockets useful. So there's a lot of good technology packed into that. Probably a joke in there somewhat. But we're going to end this on a serious note. Russ, G, Sandeep, I just want to thank you all so much for being part of this, taking a little time out. Hopefully everyone out there can appreciate digging deeper out there. And of course, if you didn't check out Marvell Investor Day, all the information is now out there. And of course, This conversation summed it up nicely. Appreciate you being part of our Six Five community and joining us for this on the road episode here in New York at Marbell Investor Day. But for now, we got to go. We'll see you all later. Bye bye.

‍