Marvell Technology, Inc. (MRVL) Earnings Call Transcript & Summary

Greetings. I'm Vivek Arya, semiconductor research analyst at BofA Securities. Thank you so much for attending our tech talk today on the evolution of cloud datacenter connectivity with Marvell and the 650 Group. As we all know, cloud workloads and Internet traffic are doubling every few years and with AI, ML and the evolution of these new large language models, these growth rates are expected to accelerate. And to meet that demand, cloud operators must scale their data center infrastructure. So what we plan to do on today's tech talk is discuss the new connectivity requirements, how they are evolving, the enabling solutions and the emerging opportunities that they offer. And I'm really delighted to welcome our guests today, including Achyut Shah, Senior Vice President and GM of the physical layer interface business unit at Marvell, who will provide an overview of the market. Really delighted and honored to have Loi Nguyen, Executive Vice President of Optical and Copper Connectivity at Marvell, who will also participate in the Q&A. As many of us know, Loi was also a co-founder of Inphi, a pioneer in optical connectivity that Marvell acquired a few years ago. And then finally, Alan Weckel, Co-Founder of the 650 Group, an industry analyst firm that tracks the data center infrastructure and the networking market. My plan is to start with a few slides from Achyut and then we will get into the Q&A session. So with that, warm welcome to all our panelists and let me turn it over to Achyut to go through some of the slides.

Thanks a lot, Vivek, and good morning, everybody. I'll spend the next few minutes talking about the evolution of the cloud and data center connectivity, the technologies that have enabled the data center of today and what is changing in the data center. And what are the opportunities we are going to have over the next 5 to 10 years in the evolving data center As a quick reminder, I will be making some forward-looking statements today in my presentation. The network infrastructure market continues to grow at a rapid pace. There is a massive amount of data that needs to be processed and transported within the data center and the network customers, the cloud operators continue to invest significantly in growing their network capacity, especially with new applications like AI as the data continues to grow, we expect network bandwidth to grow at a significant pace, about 50% CAGR over the next 5 years. And this growth in bandwidth is accomplished in 2 different ways. You move the network to higher speeds and you also increase the number of ports you have interconnecting the network to get to the scale that you need. And this provides a tailwind to the PAM technologies, which support higher speed ports as there is a transition from the older NRZ networks to the next-generation PAM networks for the future. Now while this provides a significant and tremendous opportunity to our customers, it also leaves them with a significant tremendous challenge to solve. When we ask our customers what keeps you up at night, almost all our cloud customers tell us it's reliability, reliability at this massive scale that they have because for our customers, reliability is revenue. If a network goes down, the downtime caused them in real dollars, not just for their own business, but also for the enterprise business that they host onto their networks. Beyond the revenue, every time some of these large cloud customers has network downtime, it creates headlines, not the flat [ line sort ] and actually has a direct impact on their brand. So reliability for them is revenue and reputation. Now this network that they operate needs to run at 24/7 availability with [ fine line's ] reliability. And this would be reasonably easy to manage if you have smaller networks, a few racks, a couple of buildings but at the scale that they have their network, it's very -- it's a very difficult challenge for them to manage. If you think about the entire data center network is spread globally all over the world, hundreds of locations. You have thousands of data centers already installed today with a few hundred being added every year at a significant growth rate. And this entire network has millions of processing nodes, millions of CPUs and GPUs, multiple tens of millions of network interconnects and supports, billions of users all across the globe. So when these customers have to decide a network topology, had to decide a solution to deploy in this next-generation network, they can not only look at what gives them higher speeds. They have to make sure that the solutions that they pick provide the reliability at the scale that they're working on because the group of engineers at these customers that decides what goes on to the next-generation network is also the same group of engineers that we're going to call it 2 am in the morning when the not -- network goes down, having to fix it to minimize the downtime for their customers. So this decision to upgrade the network, what comes next, how do you upgrade the speeds is a very complex decision. There is multiple inputs that must be taken into account when you make the call. So when making this decision, what is it that our customers want? Of course, you want higher bandwidth, you want higher speeds. You need to -- you had to have the economics of power and cost, and of course, you want time to market, you want to get there quickly. But that alone is not enough. What our customers want is a pluggable multi-vendor ecosystem that gives them the scale and the reliability to run these large global networks the way they need them to run. So this is not a -- this is not a problem that can be locally optimized for a couple of variables. You've got to take a global view of it. And there are multiple features that they look at when making the decision of how to deploy the generation -- the networks of tomorrow. Let me double-click into some of these notes to give you a little more color on that. To get to the scale that they need, you need the ability to upgrade your network asynchronously. And across all of these hundreds of data centers and each of them having your servers, multiple rack, multiple layers of switches, the whole network is not upgraded at the same time. There are different slices of the networks in different data centers upgraded asynchronously. So any solution you put in, you need to make sure that it's backward compatible with the previous generation products, it's forward compatible and sometimes, it's going to be [ app ] equipment that you don't even know what's coming down the line from different vendors over the next 5, 10 years of the data center's life. Sometimes these different layers of the network have different speeds and you need a solution that can negotiate between these different speeds and not just work at a given point of time when you deploy it. It's got to work for the next 10, 15 years over the life of the data center. Another important feature that these customers want is diagnostics and telemetry, which is key for the reliability and the resiliency of the network. They need to be able to know what links in the network are degrading and maintain them and fix them before they go down, so as not to cause a network outage. And if something goes down, they need to be able to find the problem and fix the problem on the fly in the field as soon as possible. The third point I want to talk about is serviceability. When there is an inevitable maintenance required on these networks, you want the flexibility, something that a pluggable solution provides you to be able to change these solutions, swap them out very quickly without affecting the other parts of the network, without bringing the other parts of the network down. So when our customers are looking at the next generation technologies to deploy in the data center, they just don't locally optimize a couple of these points. They need solutions that provide them this entire gamut of features to be able to deploy for the next generation. And this has been proved over the past evolution of the data center over the last couple of decades. So let's take a look at how this has evolved over, say, the last 20 years. Now when I first started in the optics business over 2 decades ago, then gigabits was all the rage, where we could not believe we managed to transmit 10-gigabits over a single fiber in 1 wavelength at the time. And there was a lot of time spent so many something years ago to create a direct drive 10-gig solution. They called it the 10-gig LRM standard, and that optimized the use cases for a CDR-less solution at that speed. Right after that is when you had the advent of the cloud datacenters, they were just starting to deploy, just starting to grow. And the first at scale network that they started to deploy 20, 15 years ago, was at 40-gig, which uses CDR-type solution with an analog NRZ electrooptics. That evolved into a 100-gig solution with CDRs, and that's when the analog hit the wall. Two things changed for them; one, our cloud customers needed a technology beyond NRZ that allowed the continuous upgrade of speed in the data center. But the second problem they faced when they deployed these 100-gig NRZ solutions as their scale began to grow, they began to see problems in network reliability, network diagnostics. And they came and approached multiple semiconductor vendors. I was at Maxim at the time. They talked to Maxim, they talked to a lot of the analog vendors, they talked to companies like Inphi and said they needed a path not just to double the bandwidth from a modulations team perspective, but they needed a solution that gave them the robustness, the resiliency, the insight into their network that would allow them to reliability scale based on the vision that they had for the next decade going forward. That's when the major shift happened from analog to digital, from NRZ to PAM and the DSP entered into the network. The DSP technology is what enabled the networks to go from 100-gigabit speed to 200-gigabit speed with the scale and the reliability that was needed. And over the last couple of generations, the DSP has quickly allowed our customers in the cloud to scale from 200 gig to 400 gig, the advent of 800 gig is already here. Marvell has been shipping 800-gig solutions for about 18 months now. And as you saw at OFC a couple of weeks ago, we've just sampled the first 1.6T generation. And this entire scaling from 200 gig up to 1.6T and beyond has been enabled by this fundamental DSP technology that gives you the speed along with the reliability and the scale that these cloud and data center customers need. And we believe that the DSP technology, looking into the future has enough legs to scale the 3.2T, 6.4T and provide this pluggable multi-vendor ecosystem that our customers need. So the DSP is not an overhead in the system. The DSP is, in fact, the essential block that enabled the network of today for the cloud customers at the scale at which they operate. Now those of you that at OFC caught a lot of noise around, hey, why do you need the DSP, who needs the DSP? But anybody who looked into that a little bit more, dug an inch deep into that layer quickly realized that if you wanted the scale of solution, if you wanted this reliability that the customers want, and you wanted the ability for not just today's solution, but something that scales for 2x, 4x, 8x the bandwidth from today, the DSP is going to be needed. You could always have some niche solutions somewhere on the site. But at a global scale, it's impossible to move away from the DSP. Now as someone very clearly put it, a DSP gives you plug-and-play capability for the future generations. You deploy some systems without the DSP, some kind of linear systems, it's going to be plug-and-play for your global data center network. It's not something these cloud customers want to go towards. So what is it that this DSP provides? Why is it so essential? So if you take a look at the optical module of today, starting on the left, you have the light, starting from the right, you have the light coming into the fiber, you have the electro-optical components, the linear TIAs and drivers, the lasers and the detectors that app, that convert electricity, electronics to optics and amplify the signal. And then you have the DSP that cleans up the signal, provides you the margin that you need and then transmits the digital signal over your switch or your ASIC interconnect. So while this DSP has multiple features, a few that I wanted to share with you. It has digital SerDes that compensates for all these various electro-optical impairments across this large-scale network. You have built a narrow correction at this scale and especially as the speeds get higher, the copper on the boards, the optical components, the lasers, the manufacturing, all of that is going to have less and less margin, which is going to cause errors in the networks over time over the life of the network. The DSP can not only identify but also correct these errors which cannot be done in some kind of linear solutions, which would fairly just amplify the errors, sending them over to different parts of the network, actually causing less quality, less reliability overall for the network there. These DSPs also have multiple telemetry and diagnostic features. They monitor the network. They tell you the quality of the network. And they tell you if something is degrading or going down, so you can proactively fix it before it happens. A significant amount of intelligence is built into these DSPs. And finally, you have some gearbox features that enable you to do rate translation. If some parts of your network are at 50 gigabits or 100 gigabits, other parts are at 100 or 200 gigabits as you asynchronously upgrade these networks, you're going to need something that translates between these rates. That's something a linear solution cannot do. You need a DSP for that. And over the last 5, 7 years, through Inphi and Marvell, there have been tens of millions of DSPs enabled -- that have been deployed in the network today, enabling this doubling of bandwidth over the last 7 years in the network with the reliability and the scale that you need. So while this DSP has enabled the complex, massive networks of today, when we look out to what is happening in tomorrow, when you look at what is changing and what is evolving in the data centers for tomorrow, the need for this DSP is going to continue to grow. So what happens in the future data centers? There are 3 or 4 major trends that are going to happen over the next decade that continues to evolve the network of today. The first one is simple, increasing data rates. As the switch generation moves to 51.2 terabits, the way to get to higher speed but at a much more efficient cost, a much lower power per bit in the system is to use 1.6T optical op solutions. This is something that we demonstrated at OFC at a 200 gig per lambda. And just this optical module compared to its 800 gig previous generation product, provides over 30% savings in cost and power. If you want to deploy a 51.2T switch in the 32 port system, which saves the maximum space, with the lowest cost and the lowest power, you must use 1.6T pluggable optics. That is the easiest path to upgrade your network while keeping the economics in mind. And while the speeds continue to increase, there's another significant transition happening within the data center networks. Today, these DSPs are used in the highest speeds, 200 gig, 400 gig, 800 gig and above. But as the shorter distance, the shorter reach interconnects of the network also good to higher speeds, they face their own challenges. Today, your server to top of rack switch interconnects, the short-reach interconnects between switches, which make up roughly 50% of the interconnects in the data center are all connected with copper. They're connected with what we call direct attach copper cables, DAC cables, passive cables with no electronics in them, simply the copper with connectors on both sides. As these data rates grow to 50 gigabits and then 100 gigabits for these connections, even that copper -- the passive copper cannot drive that distance and you cannot get to from point A to point B. And because the physical distance is remaining constant in the data centers, you now need to add DSPs to these copper cables and make them active electrical cables versus the passive cables of the past. And what we are beginning to see is customers deploying these active electrical cables in small volumes at 20 to 25 gigabits, growing at 50 and then becoming large-scale deployments at 100 gigabits as some of the short-reach interconnects within the data center move to higher and higher speeds. So this provides a greenfield opportunity for DSPs to grow from an application that were never used in to significant scale over the next decade within the data center. Finally, if we zoom out from within the data center network and just look at how our cloud customers are architecting their entire global network. They're moving from a smaller number of large data center buildings to larger, more geographically diverse, smaller data centers to take into account concerns about security, concerns about location of data. And as these last smaller locations continue to grow and run larger and larger workloads, you need to connect all of these locations with long distance, very high-speed links. So the ZR optics that we have deployed today using the Coherent DSPs, which are currently used in smaller volumes are going to continue to grow at a much larger scale over the next decade as a lot of these -- rise of the regional data centers need to be connected with higher and higher speed optics. So as we see, when you scale the speed within the data center, as the shorter reach links need to now start using DSPs, and your fundamental data center architecture becomes more regional and needs to be interconnected with faster links, the need for the DSP is going to continue to explode within the next 5 to 10 years to achieve the reliability and the scale that you need at these higher bandwidths. Now the other application driving this DSP is something we've all heard about, that ChatGPT and AI/ML has been the topic of discussion for the last couple of months. And when you wrap your head around these large language models that -- with ChatGPT 3.3, 4 and then going forward, where they have billions of parameters going to trillions of parameters. The amount of data that needs to be processed and hence, the amount of data that needs to be moved is humungous. And this provides a large volume opportunity, but where the AI/ML network is also more challenging than your traditional data center is it moves at 2x the speed. The traditional networks upgraded 3 to 5 years, while this AI/ML network needs to double the bandwidth every 18 to 24 months. So the technology needs to evolve faster, but it also needs to be reliable technology that is ready when these AI/ML customers are ready to deploy and to enable quick deployment because if you try to look backwards in this and try to optimize previous generations' network, the speed transition has already passed you by. The customers are moving on to the next data node. For this market, it's even more important to look forward and drive to the next feed rather than look backward and optimize minor links in the past. Also, because of this large amount of data being processed, the cost of downtime is exponentially worse if these networks go down. So as AI/ML workloads become a larger part of the data center, you need to migrate to the next speed faster in a much more reliable fashion, and you need to keep moving on to the next speed at a faster pace without having to look backwards, optimizing the past. So let me put all of these in context. I think over the next decade, you're going to have a significantly expanding opportunity for the DSP. You have every generation, as the speeds go higher and we provide a cost power benefit to our customers, they share some of the benefit for -- with us and we get an ASP uplift at the higher speeds. The other thing we see is we are very early in this transition from NRG to PAM within all the global data center networks. There is still a lot more customers who adopt PAM, which is going to provide a tailwind for the DSP adoption over the next decade. And as the data centers of tomorrow evolve with all of these use cases, higher speeds going to 1.6T, a short to reach links like AEC, not needing DSPs, and the rise of the regional cloud needing us to deploy more and more core and DSPs for ZR, the volume of DSPs is going to continue to increase over the next 5 to 10 years. So what I want to leave with before I end is, while data centers need growing bandwidth, scale and reliability are major challenges for them to solve. And a pluggable DSP ecosystem is essential to address their challenges for growth. In addition to the traditional data center network, there is a massive AI/ML application scale that is accelerating the adoption for DSP. And with ASPs increasing every generation and this massive adoption and multiple new applications coming online, there is a great expanding opportunity ahead for the DSPs for the next decade. Thank you.

Great, Achyut. So maybe let's start the Q&A part of the discussion. And to summarize, what we have here is kind of the contrast between 3 approaches, right? One is basically the direct optics approach, right, the co-packaged optics for the CPU approach. Another approach is where we have a linear direct drive, which excludes the DSP. And then the third alternative is the DSP approach, right, that Marvell has pioneered, right, and is promoting very confidently. And we heard about the reliability argument made, right, in contrasting the 3 approaches. I wanted to maybe turn the discussion to you, Loi, and see how do you contrast the 3 approaches from whether it's a cost of power, space or latency, how do they contrast? And then if you could also bring Alan into the conversation after that and see what is your take on these approaches, right, from your meetings at OFC with the customers and other suppliers?

Absolutely. Good morning, everyone. So like Achyut showed you, direct drive is not new. But what direct drive was done was for 10-gig NRZ signal. So it was a limiting case. What is new today is the new proponent of the linear direct drive trying to do it with 100-gig, so 10x smaller bit rate and linear to drive PAM signal. So PAM4 signal is more noisy because the former was signal than an NRZ signal, and the challenge becomes -- because of the linear component is an optics, so linear direct drive cannot compensate for any kind of nonlinearities in the lasers, the fibers, the detectors, the component themselves, over temperatures and voltages. So that -- and also it's a 4-level signal. So that's really the challenge. In the past, there have been 2 attempts to try to do linear for more complex order modulation or compensation and both were not widely adopted because of these challenges. One was the 10-gig LRM. It's a long reach multimode for data center. It was never adopted widely. A few years later, there was a linear drive, direct drive is on analog coherent optics, and I was part of that. We provided the TIAs and driver for that interface. The ACO was very difficult. It was almost difficult to interact, fine-tuning every channel, fine-tuning every line card. And the ACO was replaced by the DCO, DCO means digital coherent optics as soon as DSP became available and offered enough -- low enough power to be used inside of local optics. So those are the kind of challenges. So even on the panel, right, the [ Inphi back force ] team, the speaker for the span at OFC, he said that linear drives require very careful design of every single channel, require fine-tuning of every channel, require testing every line card, every module, every system. So to me, that was very hard to scale. And like I show you, right, customers want time to market, reliable resiliencies and plug-and-play and all of those. So that's the challenge for the linear direct drive. The benefit, of course, is that it's lower power. So I think that there may be some limited use case for the limiting linear drive, just like 10-gig, some limited yields. But I don't see widespread adoption of that technology to be able to replace the DSP for all applications. That's just not going to be possible. The second one that you asked about is co-packaged optics. Co-packaged optics is a totally different story. Let's -- please do not confuse linear drive and co-packaged optics. The co-packaged optics -- the premise is that you reduce power by putting the optic very close to the switch. And from the signal integrity standpoint, that's a good thing. But co-packaged optics faces a much higher barrier from the adoption standpoint, even higher than linear direct drive actually because it is not pluggable. When we spoke to our customers, whether the cloud providers or the AI/ML OEM, the one thing that they say consistent time and time again is I want pluggability. I want pluggability. I want multi-vendor ecosystem. Low latency is good, low power is good. But I need to get my system up and running, 24/7 availability. And I don't know how to deal with, I am a -- if one of these co-packaged optics tower to go down. So those are the different kind of take. So I think co-packaged optics is a longer-term horizon. Technology will need progress, co-packaged optics needs to prove itself out over the next, say, 5 years, that this is the real technology, not a laboratory demonstration, to be able to deploy it in any kind of manner. So those are the kind of my take on linear drive and co-packaged optics. I'll let Alan to finish it.

Yes. No, thanks so much, everybody, for having me today. So I think when we talk about CPO, I'll just kind of work the question backwards, that's a future technology out there, and it's multiple generations in the future. Right now, we're moving so fast on pluggables and so fast and kind of the speeds and feeds that CPO just isn't going to keep pace there with the reliability and the proven technology out there, especially when you talk about the fact that most of these data centers have some sort of brownfield in them, whether it's the cabling plant or the lower level or higher level of transport out there. And then when we look at a linear drive out there, there will be a niche for it, all technologies get deployed in some way, shape or form out there. But if we look at the pricing issues for the hyperscalers today, it's really about the move to 800-gig and then immediately after that, 1.6 terabit. And so you need that kind of ecosystem in play, which will be kind of traditional modules out there as we move for speeds and feeds. So there's places for all technology, but I think we kind of know where the road map will be over the next, call it, 2 to 4 years or next 2 generations of product.

All right. The one question right -- is that we are at a place where the adoption of AI/ML and obviously generative AI large language models is taking off. And typically, these kind of technology inflections, can they have an impact on the way people think about connectivity? So if you look at your really large switch competitor, right, they lead the cloud switching market. They've obviously tried to make their case for linear direct drive and co-packaged optics. So one question, a follow-up to you, Loi is, do you think they can integrate the DSP into their ASIC, like is that kind of integration possible? Where, yes, you still have a DSP, but it is integrated into the switch ASIC so that can become potentially a competitive issue. So what would be the pros and cons of that kind of approach, right? And then kind of part B of the question is, that your other -- I shouldn't say competitor, more kind of a collaborator on the AI and AI and ML side, right, who is leading the accelerator market. They have spoken very confidently about the need for lowering latency. And every time you introduce another component like a DSP, you increase the latency. So if you could address it both from the switching side, can they improve in an ASIC? And then from the latency side, but are there disadvantages, right? If I'm using accelerators from the same company, do I even need interoperability, right at all? So if you could address it from their perspective because that will have a major bearing on the direction of this market.

Okay. So that's -- those are good questions. So on the first question about the SerDes to be integrated by the switch, that is happening now. Basically, on switches have SerDes and the SerDes getting better and better with each generation. And then SerDes can drive the long distances provided it has the DSP on the other side. With SerDes trying to go drive long distances without a DSP on the other side, and you face all kind of the issues that I brought up earlier with the linear direct drive. You need to drive through 10 inches of copper traces the 20 DV loss, signal, the eye become close. But electrical design, any kind of noise on the electrical channel will be amplified by the TIAs and the driver, the linear amplifier into the optical domain. And once those noise gets amplified in the optical domain, it is very, very hard for -- to go back and then have a switch ASICs sitting at 10 inches away to be [ away ] and compensate for that. That was why the DSP was invented in the first place. So whether the DSP integrate into the switch ASIC or not, it's the same problem. With respect to CPO, yes, the DSP, obviously, need to be incorporated in the switch ASIC because -- but in this case, you don't have the signal integrity problem because the optic is driving next to the switch ASICs. CPO as I mentioned earlier, is facing different kind of adoption hurdle. And in terms of latency, latency is certainly is something that, that customers want to reduce. But think about a 30-meter fiber, the latency is 150-nanosecond for light to travel through that fiber. And then you add to the latency that go through the switch itself. So the latency in the DSP itself tend to be in the low single-digit, it's not 20%. It's not 30%. It's not 50%, that's for sure. And so, yes, latency is important. But yesterday, there was a panel and NVIDIA speaker said that I want pluggability. Yes, I would like to reduce latency, but I need my system to work. So DSP is still the [ appeal ]. Yes.

Maybe Alan would appreciate your view as well on this, because you have the market, right, really being driven by people on the computing side, which is NVIDIA. So can their desire, right, and their architecture which prioritizes lower latency, can that create a disruption in the market that challenges the role of the DSP? And then sort of a similar question from a Broadcom perspective, since they lead the market on the switching side. Can their preferred architecture, right, challenge the role of the DSP as well?

Yes. That's a good question out there. So I think we should kind of ground in the data of AI, right? AI bandwidth is growing north of 100% per year. So something even higher than exponential. And that's what the real challenge is in these kind of scale-out architectures, and it creates both an internal and external network. So 2 challenges show up, higher bandwidth per year and then these internal and external networks. And so when we kind of optimize for that we kind of end up back with pluggable modules, right? In the internal network, you're going to have a purpose-built set of switches and connections there that run at tremendous speed and latency is important. But scale and reliability are more important. And then when you get to the external network, it really is all about speeds and feeds. And that's where you're pushing from, call it, hundreds of gigs to terabits per server out there. And that incremental speed is really about pluggables and really about moving the speed forward, not kind of optimizing for a little bit of latency or a little bit more performance on some other metrics.

And just one last follow-up on linear direct drive. So Loi, do you see any niche at all where linear direct drive makes sense? And do you see what might be a niche today, getting bigger over time. And if it does, how is Marvell positioned then?

Good question. So for the -- I think the market is expanding rapidly. So there's opportunity for basically every vendor, every technology to look for niche markets to insert their technology in. So linear drive could work in application where the switch ASICs and pluggable are not so far apart, like in a switch that's 10 inches. But maybe something like a NIC card where the distance maybe 2 inches, there could be a niche application like that. Also realizing that the whole SerDes and the line speed moving at a different speed, like Achyut said, right? So often time -- well, let me draw back. The ASIC SerDes go up in -- to the next speed about 4 or 5 -- 2x every 4 or 5 years, whereas the lines speed, the optical line side speed doubles speed every 2 years. So you only need a gearbox to translate between these 2 speeds. And that's the Marvell DSP. Only when the host ASIC SerDes and the optical line are the same that you can even think about using linear direct drive. So there is that niche where the line side and the whole side is the same where the loss between the switch ASIC or the NIC and the optical is low, there could be applications for that kind of -- for that type of application. But customer, again, they want pluggability, multi-vendor, [ inter arm ] standard. Right now we need [ a director ], there no standard, they are no ecosystem. I mean they're going to try to get to that, but it's going to take light a while. And meanwhile, the AI/ML moving up to 200-gig and 200-gig can be very, very difficult, what we need to try to do. So some niche application [ title ].

Great. Let me bring Achyut into the conversation. So OFC created a lot of excitement, due and undue. I was hoping if you could give us your perspective on what you heard in. And where I would like you to kind of focus is does the introduction of generative AI change the landscape in any way that challenges Marvell's DSP dominance? Do you see a path of growth for DSP? And I think one thing you guys introduced at the conference was the 1.6T solution. So if you could also talk about what Marvell's roadmap looks like in this business.

Absolutely, Vivek. And actually, both the questions you asked about how regenerative AI changes this going forward landscape for the DSP and the 1.6T are actually interconnected questions. So what AI/ML does is it does change the landscape, but it changes it in a positive fashion for the DSP, which is actually going to drive more and faster adoption to DSP because the fundamental challenge for the AI is growing the network bandwidth at a much, much faster pace than everybody has been used to in the last decade. So they want to get 2x speed every 18 months. So we're going to have to come out with a new DSP at a much faster pace and it's going to add not only significantly to the volume, but what the DSP does well is getting first to the next speed. Then when you try to sort of an optimized solution, remove the DSP, characterize the links. That's more of a backward-looking activity. And in AI, you don't really have the time to look backwards because in that 18 months, you've already gone on to the next generation, right? So if you just take 800 gig, for example, large customers today in AI/ML. Marvell has been shipping our 800-gig DSP-based modules in AI/ML to multiple customers for 18 months already. So when something like linear comes and says, next year, we might have a standard, you've lost 2 years already. This is already backward looking, right? We introduced 1.6T at OFC, there was a lot of excitement for the customers around it. The first customers to deploy 1.6T are going to be in the AI application. So we're sampling our customers right now, they're going to go to qualifications later, a year. And by next year, they're already going to be deploying 1.6T in the AI/ML networks. So yes, AI/ML changes what the DSP needs. It makes it more complex and it makes it much faster to adoption. So it actually is a tailwind and a significant growth path for Marvell's DSP for the speed and the reliability that these AI/ML networks need. Now the first application for 1.6T around which there was a lot of excitement at OFC is in AI/ML. We were the first ones in the industry to show 200 gig per lambda solutions in the market. And so I say next year, this AI/ML moving to 1.6T. And then right after that, as the 51.2T routers -- switches are operated in the system, you'll see 1.6T moving on to that also. So I actually think AI/ML is fantastic news for the Marvell DSP.

Got it. And just a follow-up to that. So let's assume that DSP is going to be the preferred solution, right, for all the reasons that Loi and Alan and you referred to. But what about the competitive landscape within the DSP market? Because one of your very large competitors also announced an 800G solution, right, with an integrated TIA and driver. So how does Marvell fare from a competitive perspective, Achyut from your perspective? And then, Alan, I would like your kind of independent perspective on how you look at the landscape right now.

Certainly. I mean, frankly, any market worth getting into is going to have competition, right? It just sort of validates that it's a real market to be in. It's a good place to be. And so there is no market without competitors. They keep driving us. They make us better. At Inphi and at Marvell, our goal has always -- be the first to market. Again, yes, they've come out with 800-gig solution. Like I said, we've been shipping for 18 months. And by the way, our solutions have had an integrated driver from day one. So that's not -- we've had integrated drivers from 400 gig, 800 gig, even at 1.6T solution we demoed at OFC has an integrated driver. So that's sort of been a core competence at Marvell. The TIA is an interesting piece because there is something called over integration, right? And if you look at all the different links in the network, you have shorter reach, 30-meter links, you have 2-kilometer links, you have FR4, DR4 standards. And integrated TIA inherently has less skill -- it has less flexibility and less performance than a discrete TIA. So our customers -- we talk very closely with our customers, there's 5, 6 big customers in the market. We don't put out a merchant solution. When we talk to them, they wanted discrete TIA, and that's what they've been using for the last 18 months, like I said, because they want that 1 solution with the DSP and integrated driver in the discrete AI to work across all the distances, all the links in a scalable, reliable fashion. Now as you go forward, would they want to optimize for some specific links where an integration might make sense? Yes, but then that's not the only thing you do. There are other features they want for that specific link, to optimize that solution. So I think we work very closely with our customers. We are hearing what they're saying. Our rule is always going to be first to market. And from an 800-gig perspective, yes, there is competition. There's always going to be competition, but we are in a very good position. With all the ones that have shipped, hundreds of thousands of these at scale, running with the reliability proven. And the question of integrated [ AI ] makes a good press release, but you got to be careful. It actually ends up hampering you from the scalability and the reliability you might need for all the data center applications. So you've got to actually optimize these working very closely with your customers for specific links. So you don't always equate more integration to better performance. So it's a delicate balance you have to hit there.

Competition is always good for innovation. And when we talk about going into these AI networks, we need a lot of innovation out there. So that's a good thing for the industry out there. It drives everyone faster to the higher speeds. When we look at kind of the higher speed markets, I think we kind of touched on this a little bit earlier out there. Next year, the largest hyperscalers will be moving to 1.6T, both on the switch and the optics for AI and other parts of their network, while some other smaller hyperscalers will be moving to 800 and then we'll have the rest of the broader ecosystem, say, moving to 400. So this is the first time really at an industry level where the market has gone even beyond bifurcation, and we're going to see multiple adoption curves out there. So we should expect to see this at a market level, right? On your bleeding edge, then you have following right behind that and then kind of the standard adoption curve. And they're going to be at 3 different speeds and 3 different technologies just because of how fast the hyperscalers are moving or how large the AI opportunity is. That's new for the market. It hasn't happened in Ethernet in it's 20-plus years of being a data center technology.

All right. And maybe related to that, Alan, staying that you -- in the last year or so, we have also heard about the active electrical cable market, the AEC market, right? There was a small credo who kind of got that up and running early. Where are we in the adoption phase of AEC? How large of an opportunity it can be? And I should maybe -- then I would like your perspective on how Marvell is positioned in AEC?

Yes. So AEC is kind of the natural progression there of server access. We're moving from passive cables to active cables as we increase the SerDes speed and port speed and that's being driven by 2 things. We've mentioned the first one already, kind of the AI phenomenon. But also remember, we've rapidly moved from PCIe Gen3 to 4 to 5 on the server side. And so that increases your bandwidth kind of coming off these servers in a way that the industry wasn't used to in the past out there. So the speed is what's driving that. In terms of market size, it's going to be a multibillion dollar market when you talk about kind of the cable solution out there. When we talk about it from just kind of the DSP or chip level, it gets north of $1 billion. So this is kind of a market that comes out of the DAC market where no one cared. And by adding in technology, we're adding value in the network, and it will kind of again, create a multibillion-dollar ecosystem there for server connectivity.

Achyut?

I agree with everything Alan said in terms of the scale of the market. It drives the size of the optical business we have today in terms of our availability of the market. And at AEC, it's a very interesting opportunity for the DSP. These customers -- the physical distance that are remaining constant in the data center as the speeds keep going up. But even beyond scaling the speeds, what you realize is at the density of the network they want, with all of these copper cables plugged in, they need to actually start using thinner copper cables. And thinner tables actually have more loss. So that need to go to sort of higher-density racks with pillar cables to optimize the TCO of the data center actually is driving the need to add this DSP into this passive cable and make an active electrical cable, like Alan mentioned. Our competitors have done a good job actually starting and getting ahead in the market. There are some few niche applications that have ramped already in the last year and this year. I think the big ramp, the growth is going to come over the next 2 to 5 years as these technologies gets into multiple customers. The ecosystem, I think the same fundamental mindset that we have on optical for pluggable multivendor ecosystem is what the customers are going to need for these cables. At the end of the day, it's just another interconnect technology for them, regardless of whether it's fiber or copper. So our -- Marvell's view on this is just like on the optical side, we want to be a merchant DSP vendor. We are working with 5, 6, 7 module vendors who provide the copper cables and the connectors, integrate our DSPs into that and provide a multivendor ecosystem for our end customers. So we have -- at OFC, we've done, what, 3 or 4 demos. We have multiple other customers we're working with. Our DSP with multiple cable vendors is already in the hands of multiple large cloud customers that they are testing them today, qualifying them today, and we expect the ramp to start sometime next year. But I think at the very, very beginning of this market, it's sort of this optical DSP was maybe even 10, 12 years ago. And in scale, over the next decade, it's going to rival the size of the optical market, so it's a great new growing opportunity for us. And coincidentally, the market moved from NRZ to PAM4 and analog to digital in the optical domain of 50 gigabits, and it's sort of similar to what we're seeing right now on the electrical side. There are some niche applications for AEC at 25 gigabits NRZ. But as you get to 50 and then 100, you really see this growth of the active electrical cable that uses our DSPs. It's all about a multi-vendor pluggable ecosystem and scale and reliability just like it is on the optical side, Vivak.

And then maybe the final piece of the puzzle when it comes to data center connectivity is on the DCI pluggables market opportunity. Maybe, Alan, if you could educate us on where we are in terms of the cloud datacenter adoption. And then Loi, I remember the first time we met, tried that at Inphi. Empire was one of the pioneers in coming out with the COLORZ module which was deployed very extensively at Microsoft at that time. And the industry is now getting towards this 400-gig ZR and ZR+ standard. So maybe if Alan, you could educate us on where we are in terms of the adoption curve from the data center cloud operators. And Loi, if you could give us kind of the technology aspect of where we are from a technology adoption perspective.

Yes, absolutely. So we're moving kind of from the early adoption into mainstream adoption there on ZR and DCI. And this is primarily driven by the hyperscalers. So this is all greenfield connectivity between facilities that could be within a metro or within kind of a short-reach metro out there. And we're also starting to converge other layer 1 technologies onto Ethernet. So a similar theme that we mentioned on AEC and AI, this is kind of a new greenfield opportunity for Ethernet. It increases the TAM at the switch and router level and then the optics level there. So you're going to get DCI to become a multibillion-dollar opportunity driven by cloud, and I'd say this is kind of irrespective of AI, right? You're building these DCI networks in parallel or adjacent to what you're doing on the AI side out there. So that provides the incremental story and TAM expansion for Ethernet in a fairly significant way. The market size there will be bigger at some point, this will become a $5 billion to $10 billion market switch plus optics, but we're kind of in the -- getting past the $1 billion phase right now.

Yes. So I'm really super excited whenever anyone talk about DCI and ZR optics. We pioneered that at Inphi. So a little -- let me step back a little bit. In the own days, cloud provider had only global data centers and these are the data centers that consumed 150 megawatts of power, big, big campuses. But as cloud computing became more ubiquitous and now ChatGPT and so on, and also privacy laws, the personal data center needs to stay within borders. And resiliencies of data center, 24/7 like Achyut talked about. Cloud providers are thinking differently how they build new data centers. Instead of building these 150 megawatts data center, which would be very difficult to platform now in, say, in the middle of San Francisco or the Virginia corridors or L.A. So -- so instead, what they do is they'll build dozens of smaller ones. But when I say smaller one, I mean like 35 megawatts class. We're not talking about small, 35-megawatt class of data center. But a dozen of them, half a dozen of them in a region, separate by, say, 80-kilometer or 100-kilometer or maybe 40-kilometer apart. And these data centers, they can be connected by high-speed links and not the ZR optics that Alan talked about. So I would say I totally agree with Alan, we are in the early stages of adoption. 400 gigs is the first high-performance coherent optics that can be put in a small form factor pluggable. And that's going to replace all kind of like car boxes, chassis because why? Because they are much lower cost, much lower power, takes up less space, higher face based density, interops standard, all of those things. The thing that surprised me and which I think I should think about it, suddenly is that ChatGPT actually bring more customers to the regional clouds. And so we actually begin to see an uptick in demand for our [ pluggables ] as we speak today because of all the traffic that's coming into these regional cloud datacenters.

Noted, as a lot more inference activity is done in those clouds.

Yes.

Absolutely. So I think we have run through the gamut of connectivity inside the data center and between the data centers. I have a number of other questions that I've gotten from the audience. But I think we -- for the purpose of this call, we have captured some of the key points, and we can close the call here. But I really want to thank Loi and Achyut from Marvell and Alan from the 650 Group, right? And to everyone in the audience for participating in this call. And I think the slides will be posted on Marvell's IR website following the call. And as always, if you have any follow-ups, please feel free to contact me, and I'll be getting back to people who asked the other questions during the call. So with that, Robbie, let me turn it back to you to close the call. And thank you all, and have a good day.