Lumentum Holdings Inc. (LITE) Earnings Call Transcript & Summary

Good morning, everyone. I think we're going to go ahead and get started. So I'm Kathy Ta, I'm the Head of Investor Relations at Lumentum. It's great to see all of you here, standing room only in our event this morning. We are working on getting more chairs for the folks that are standing in the back. So I'm very happy to have our executive team here to talk with you today about how Lumentum is enabling the AI revolution, and this is our Second Annual Lumentum -- 2024 Lumentum Investor Technology events. So I am told that more chairs are on the way. It's great to see the turnout for our event. So first, we have a few safe harbor comments. We are going to be making some forward-looking statements. Those statements are subject to some risks and uncertainties. You can read all about it on our website. We are going to post these slides at 8:30, so you can read the forward-looking statements then. And then I'll go to -- okay, it doesn't go ahead. Today's speakers. So we have 4 of our execs going to be speaking with you today. Alan Lowe, who is our CEO; as well as Wupen Yuen, who is our President of our Cloud and Networking Business Unit. We have Dennis Tong, formerly known as the Cloud Light CEO. Now he is the Group VP and General Manager of our Cloud Networking platform. And then finally, we'll have Wajid Ali, talking about some of the financial particulars of our business. And later, at 8:30, we'll have a Q&A session, and Chris will join us on stage to help address the questions you may have. When it's time for Q&A, please step up to those standing mike that's in the middle and the isle, and state your name and your affiliation and we'll take your questions. And now I'd like to turn the stage over to Alan.

Thank you, Kathy. Well, this is -- reminds me of when I had an Investor Day in 1999. This is different. Really is different. I was on a panel actually yesterday, and the question was, is this similar to 1999? Or is it real and sustainable? And it was 4x4 real and sustainable. So I'm super excited to be standing up here in front of each and every one of you to talk about what's going on at Lumentum and the impact of AI specifically to what we're doing and how we're addressing it with our customers. Next page, please. We're going to go the old-fashioned way. We're going to talk a lot about AI, but it's not just about AI. It's about the sustainable growth of AI. And in fact, a little bit about how Lumentum is using AI and really addressing the speed and scale of developing new products, manufacturing our products and how do we go to market more rapidly to meet the -- address the rapidly growing needs of the hyperscale data centers and the infrastructure providers. In addition, we're not going to talk too much about this, but Industry 5.0 is real. It's really taking advantage of what photonics can do for how things are manufactured. We're making lasers today for EV batteries and for solar cell manufacturing. But more importantly, we're addressing with our imaging and sensing business, really machine vision. And all of that data that comes from our LiDAR applications and machine vision are being uploaded into the cloud to then be able to make things better for those industries themselves. So it's super exciting, but we're going to really focus most of the time on AI and how we're addressing that rapidly growing market. As I said it's a rapidly growing market, one of the reasons we did the acquisition of Cloud Light, and we wish we had done it earlier, but Dennis was a tough negotiator, and it took us a little bit longer than we needed or wanted, but it's a big market today. If you look at the various parts of what we call the $4.5 billion market, transceivers and Data Center Interconnect, which is going to become more and more important as the needs for power are making data centers have to be built further apart. And so Wupen is going to talk a little bit more about that. But there's also the transport network both between data centers as well as what we are able to do, and Wupen again will talk about this, too, is the advanced switching inside the data centers to really drive a different level of power consumption and latency, which really is needed by the hyperscalers and the infrastructure providers. So when you look at the combination of Lumentum and Cloud Light, it's really a win for our customers, and I'm super excited to have been out talking to our customers for the last 5 months about what we can do for them, not just on next generation of transceivers, but really what is the fundamental architecture needed for next generation and next, next generation. And again Wupen will talk more about that. But I see this really as an acquisition of 1+1=6. And that's really because of the fundamental core technology that Lumentum has, as well as the -- what Cloud Light brings with respect to time to market, speed to ramp and manufacturing infrastructure and capabilities that really provide the best-in-class capability that our customers are super excited to see. The other thing that we're worried about is the geopolitical environment of the world. The acquisition provides us with a major manufacturer, that's headquartered in the U.S. with capability outside of China. And when you talk to the hyperscalers, they've tolerated a lot of the Chinese build manufacturing, and they're going to continue to do that, but they really want to make sure they have alternatives that are outside of China and Lumentum provides that capability for them. Next please. As I said, the now new Lumentum and Cloud Light together is really first-to-market. And as shown by the chart on the left, you can see the 800 gig is 30% of the market last year. The transceivers we shipped in the December quarter, 75% of those were 800 gig. The balance of them were 400 gig. And so really the capability that Cloud Light brings with respect to development of new products at leading-edge will gives us confidence, gives our customers confidence that we'll be first to market with 1.6 terabit. Taiwan manufacturing, Dennis is going to talk a little bit more about that, but we have a large campus in Thailand that has an extremely capable team and proven manufacturing infrastructure to be able to ramp quickly. This is what the [indiscernible] says that in 2028, will be the mix of products. I really believe that 1.6 terabit will be a significantly bigger part of the market than as the hyperscalers will shift to 1.6 terabit even faster than what this chart says. And I think, that's really because the economics make sense. But the new GPUs need more and more data faster and faster in connecting those GPUs and clusters together really will drive the need for higher speed than even the 800 gig. So we're excited about where we are and about how fast that market is growing. I talked about this a little bit, but maybe taking it one step further and talk about. When I go out and I talk to the hyperscalers, what they talk about to me and how we're addressing that. Really time to market is critically important to them. Having that capability as soon as they can and having that ramp very quickly with high quality and low cost and low power consumption. Power is such an important part of the design of new transceivers and the capabilities that I think between Lumentum and Cloud Light bring, we really do focus on how do we drive the power down, the cost down, quality up and the ramp as fast as we possibly can, so it's an exciting time, and customers resonate to that. I already talked about geopolitical concerns. We have wafer fabs in the U.S., in Japan, in the U.K., providing technology for the hyperscalers and the infrastructure providers with the manufacturing footprint. It really enables us to give them confidence that we will invest for them and be there for them, when they need to ramp very quickly in Thailand outside of China. They also want to talk to us about what comes next. And I'll let Wupen talk more about this, but we've had in-depth conversations with customers about is CPO the right thing? Are there external light sources? Or what are those next-generation things that really provide the data to the GPUs, when they need it at the speed they need it, which is really quite amazing. So together, I do believe that we provide a solution to these customers that is really unparalleled. I suggest that you, after this session, go across or down the street to see our booth. We've got a lot of interesting things going on there. We believe we have one of the first to market 800 gig ZR, not just on normal ZR, but also the separate modes that trust the needs of data centers moving further apart. So extended reach, there is an interconnect, metro 800 gig. So for the short metro hauls, very capable 800-gig as well as lower speed but longer reach. So long haul 400- and 600-gig. So all these modes that we demonstrated at our booth, which really is getting the traction from our customers. Already talked about Co-Packaged optics and solutions beyond, 1.6 terabit and 3.2 terabit, we have demonstration in our booth talking about that, which really, again, as the hyperscaler is excited about. The capability not just to making transceivers but having technology -- fundamental core technology at the chip level to be able to do the different kinds of designs that our customers need. As I said earlier, power consumption is critically important for these hyperscaler customers. And so we're going to be just demonstrating a linear receive optics, which really cuts down the power consumption of transceivers at 200-gig per lane. So these are the capabilities that enable 4x200 for 800-gig, but also 8x200 for 1.6 terabit. And so we're demonstrating those in our booth as well. And then as more and more data needs to get to the edge of the network, and we've been deploying tens of thousands, if not hundreds of thousands of 10-gig tunable SFP plus that have been in the market for probably 15 years, we're now introducing an extended reach 25-gig tunable product that enables the access, as well as the wireless providers to upgrade the speed at the edge of the network because of the needs for getting that data to the edge as rapidly as possible, which will then again drive demand for metro and long-haul networks demand. So I suggest you go to our booth in the convergence there. Next. So I do truly believe that we are positioned to win. I think we're making the right investments for R&D and manufacturing infrastructure. We have a broad portfolio of transceivers today, as well as samples to customers and development agreements with leading-edge customers to provide them with what's next. In-house capabilities. When we talk to our customers, that one of things they are concerned about is, I don't want to rely on a single source of EML supply from Lumentum because Lumentum has a very large share of EMLs to all of the transceiver customers. So we have EMLs, we have VCSELs, we have Silicon Photonics in the combination of what Cloud Light brought to the acquisitions we've done in the past, the IPG technology, IPG Photonics acquisition as well as the NeoPhotonics where we brought teams of Silicon Photonics together, enables us to rapidly turn Silicon Photonics design. So we can do whatever our customers want given their use case and given their desire to have diverse technology as well as diverse supply chain. And so we're uniquely positioned with that. And then as I said before, wafer fabs throughout the world, Japan, U.S. and the U.K. with the infrastructure for manufacturing capability in Thailand, which really has a lot of traction, and I'm very busy with a lot of customer meetings in the rest of this week. So I think we're getting the traction we need and want, and I do believe we will grow faster than that 30% CAGR, certainly on the data center and hyperscaler side. So with that, I will turn it over to Wupen to talk a little bit more in detail. Wupen?

Thank you, Alan. So I'd like to share kind of my excitement of the AI machine learning world. I want to give you a few fundamental concept, right? There's a lot of jargon, a lot of technology being talked about. I want to give you some fundamental reasons why that's the case and fundamentally what's changing, and then we can talk about some of the ramification technologies that can fit into this picture, right? So #1, the biggest change is really that the workload of AI machine learning now has really growing so much faster than what the [indiscernible] the scaling can actually happen, right? We all know that Moore's Law is reaching its limit, right? Used to be, you can just scale up, go to higher density nodes and you can get -- all the compute together on single chip not anymore. And fundamentally that barrier is causing a major change in the optimal industry, number one. Number two, now you have to imagine then, given that's the case, now we're really shrinking, imagine we have a national network today, AT&T override have tens of thousands of nodes, right? The connections all coming together. Now all that idea is strong into data center, right? Because now you have tens of thousands of nodes on GPUs. Everyone's running at a really high speed. How we're going to actually connect them together at the very lowest cost per bit, lowest [indiscernible] per bit possible, right? That whole idea, if you think about the overall picture, we talk about optical switching, you talk about these different ideas of NVLink versus Optical and all these different places optics is playing a role, that's all basically thinking about how we're going to transmit for the bit to -- from location A to location B at a lowest nonblocking, lower latency way. So the whole ideas of portfolio technologies that Lumentum has, is going to play a role now, because everything from long haul, optical switching, WSS, reference routing to even pumping and high speed transmissions, and high data rates, stuff like that, all going to be converging into data centers. I want you to remember that because you're going to see more and more since actually happening, right? The third one actually is even more important, is now optics is part of compute. It's not just connecting things, it's actually part of the compute. And then we call it why called cloud speed because now because of power compute, now the data rate evolution is now tied to the -- how we can handle the workload growth, right? Number one. Number two, then the cycle time is now shrinking. The life cycle used to be like 4, 5 years in the cloud space, now it's 2 years. It's now tied to every release of the GPUs, right? This is a really, really big deal, right? It thoroughly challenges the scale, the technology, the go-to-market, manufacturabilities of the technology and the scale of the company, right? And therefore, thinking overall picture here, I think Lumentum is well positioned because of technology base, all the way of optical switching down to the small chips, with indium phosphide silicon photonics, lasers, technologies to optical switching, amplification and so on and so forth, right? So these are the kind of the big ideas I want to give to you before we jump into the details of why 1.6T, one why CPO? Why this, why that? But these are the 3 big things I'd like to share with you. So next slide, please. So we talked about this already, right? So fundamentally, the fundamental driver of this is that large language models require so much computation, right? And the Moore's Law of scaling is really stopping, from 16 to 7 to 6 to 5 to 3, the scaling has become much more challenged now, right? And therefore to do the other compute, you are forced to break it up into multiple chips, right? It's not just the computation itself, but also all memory access, right? So that fundamentally basically meaning you can use co-ops to co-package, functions for GPUs and memory. And like the -- Nvidia just announced, now we're going to put a whole NVLink linked up, 72 GPUs altogether, is trying to create a large computation unit. That's how it can be very effectively doing a lot of amount of competition. But despite that you still have the need to go beyond a single rack, go beyond a single GPU. So all these connections has to be made somehow. The optics really is the only scalable approach to accommodating this ever-increasing -- rapidly increasing AI demand, right? And this is fundamental reason why you're going to see a lot more optics coming through. If you asked me in 1999, I got the feeling of like, let's talk about the Internet coming in, changing the optical industry for the last 20 years. And now the AI is going to be supercharging that for this fundamental reason. So next slide, please. So now traditionally, the left side is the kind of traditional cloud networking, right? It's kind of switch based, cloud server based, this is all well known. It has a telecom business, carried different business for the last 10 years, everything is going pretty well. It was a large volume and so on and so forth, right? The AI showing up to the world now has really a big effect because we never had this much competition before, right? So this new competition node has to be carried somehow, right? So that in itself is really driving up the use of optics, right? And fundamentally, also though, there is a very big difference, is that the AI clusters are built on non-blocking, right? So non-blocking basically means that there's harder connectivities. You cannot have -- because GPU is really expensive. [indiscernible] you cost a lot of money, right? So everybody wants to make sure that GPUs are fully utilized and all the switchings are actually nonblocking. Non-blocking basically means a lot of links, therefore, a lot of use of the optics, right? So basically, that itself is causing a jump, right, in the usage of optics. So that's kind of all the excitement we've seen so far, right, since last year, OFC to now. But there's a hidden one, which is not even talked about here yet, which is not yet in the numbers, is -- when you start to open up the GPU to GPU and GPU to the memory interconnect, that's going to be another factor 10 of bandwidth open up. Today, Nvidia, for example, using is NVLink to connect them. So it is still all copper-based, right? But as they scale GPUs, there's a scale to reach the limitation of the silicon or copper-based interconnects between GPUs, now will have to come optics as well. When that happens, that's going to be another fact of 5 to 10, right? So there are 2 fundamental big things. Today, it's just scale and volume and non-blocking switching and so on and so forth, large compute node. Tomorrow is going to be opening up the NVLink to the optics, right? That's going to be a really big deal yet to come, right? That has not -- it's not in anybody's forecast yet. And that's where CPOs, all these low-cost, low-energy technologies will come into play. So to carry all these fast-growing traffic, right, this is tried-and-true technology. It's basically let's just go to the next highest speed possible. Why? Because typically, when you scale to a next higher data rate, you'll use the same kind of optics like a single lane used to be 100-gig, now its 200-gig. So immediately your cost per bit dropped by nearly half, not exactly half, but nearly half, right? Secondarily, actually used to be the case that when you go to a higher data rate nodes, your power doesn't go -- consumption doesn't go up by 2x, right? And therefore, you actually also get a power consumption advantage as well. But that is also slowing down, right? Because Moore's Law is kind of being reached, the generation to generation power reduction is getting less and less, right? But overall though, still to carry the traffic, when the GPU is growing so much faster, the only way to do is actually use the highest speed optics possible. And therefore, you see Nvidia is being the leading edge 800G and then whoever is doing -- using our AI training workload, they're going to use the highest optical as the only way to be carrying doing this competition in a cost-effective way, right? So we see the 100G deployment skyrocket because of AI machine learning in about 1.5 years, we're going to see 1.6T skyrocket also because of AI machine learning, right? So that speed is really going up. Now Cloud Speed has different meaning, it's also now the cycle time is shrinking, right? It used to be for connection, now it's for compute. That's the big idea here, right? Optics now is part of the compute. And now it's not every 2 years you're going to see a generation. If I would not be surprised at all because the competition, the computation pressure between these GPU companies, that cycle may even get strong, even faster, right? So everybody is now competing to bring the highest GPU possible. And at the same time, we, as optics vendors will also be pressured to bring out the latest optics possible to pair up with the compute, right? So that's what we call a cloud speed. Next slide, please. So, and then, of course, the AI machine will be the first way to think about this is going to be like waves and waves of technology leverage, right? So first phase of the high-speed data rate 1.6T as an example, it's going to be happening late this year, 2025, in the last 2 or 3 years, it's going to go away. They go to 3 to 3.2T. But there's going to be waves and waves of, let's say, smaller users, later technology adopters that will be using 1.6T technologies, right? So you're going to see waves and waves of this kind of adoption of higher data rate. And so in aggregate, you have a very fast ramp, right, along a few years of lifetime and coming down. But we think each of this is kind of a several waves kind of combined together into a big wave, right? So that's going to -- what's going to happen. And this will fundamentally then grow our TAM, right? So you now have an AI machine-based fast ramp attempt followed by smaller, but I guess, similar amount of the TAM in combination, right, to -- for every generation of technology. So next slide, please. So this is an interesting one. We recently, ever since the GTC event last week, question talk about, hey, now Nvidia is linking all these 2 together, now 72 GPUs in a rack, rather than just like one server box, what's happening here? What's going to happen to the use of optics, right? So -- and somebody told us that it is up, right? It's actually very interesting. So basically now what we see here in the world is this. Single more optics, right, because the bandwidth is growing very fast, the only scalable technology going forward is going to be single more optics, right? So they will basically be more and more used single optics. Meanwhile, the copper technologies, you can see Nvidia's announcement of the [indiscernible], it's still the lowest cost, lowest power per bit technology. So now they're trying to increase the use of copper, but that doesn't change the use of optics because the interconnection, the distance that are being used are still favoring optics. So fiber optics doesn't really reduce, it actually goes up, right? And the only way to kind of temper down the volume increases by using higher data rate, 1.6T, for example, what's actually squeezed in the middle is actually the multi-mode. Multi-mode used to be connecting the servers to the middle of the road, switch and stuff like that. That's actually gradually being replaced a way by shorter distance, by copper, and longer distance by single mode, right? So what's happening here is single, multi which Lumentum is really, really good at is actually going to see a larger percentage of market going forward in this transceiver form and eventually it comes more like co-packaging, optics kind of form. So that's what we are really, really focused at. Now the multiple technology will have a different use cases. We will talk a little bit later. It doesn't mean you'll go away, but it's going to be a different use case later. So think about it basically higher data rate, larger GPU clusters, you're going to see more using more optics at the networking level, connecting level in the short distance, probably a little more copper, right? And multi-mode would be squeezed a little bit, but the technology of multi-mode will be used to some other activities.

When you say that [indiscernible] in the framework at [indiscernible] 72, that copper back. You say that it moves to optical because the need to get the bandwidth up or the need to get to reach beyond 72 [indiscernible]. Is it the need to reach up? Or is it just getting the data rates on the copper to get too hard?

Yes. So probably the latter.

The reaches. Basic connecting node to node, [indiscernible].

Yes, will come later on questioning. Yes, yes. So next slide, please. Next slide, please. So this is just a physical summary, right? So basically, what happens here, you can see single mode can cover everything, all use cases, everything, right? So it's -- and also single mode because you build data center, right? The cabling is actually a very big investment. You don't want to just change the cabling all the time. And therefore, single mode, if you want to choose to build a new data center for AI or machine learning, you will choose the structural cabling, you're going to use single mode, right? You're not going to see multi-mode for cabling. And therefore, I would imagine, most if not all new data center will be built on single-mode fiber, right? Because the copper has the advantage, but multi-modes being squeezed in the middle, as I was just talking about, right? We imagine that the use of single mode will increase. And again, that's what Lumentum is really, really good at with all the scale technologies and road map to support it. So next slide, please. So then we'll talk about this again, right? So with bandwidth growing so fast, really the only way to grow further, to grow to scale that is to use the highest data rate optics, right? And therefore, we're now talking about the 200G per lane optics that will be giving you the 1.6T, right? And meanwhile, actually, in the electrical domain, it's also now moving from 100G per lane to 200G per lane, right? So they're kind of pairing up and that's actually the lowest cost, lowest power consumption way to carry the traffic through. So that turning point is fairly challenging technology as well already, right? Because doubling the bandwidth in a short amount of time is actually not easy, right? But we do see Lumentum technology to actually carry out to either 400G and 800G per wavelength technology, right? At some point, you can argue you might have to use the coherent technologies, in order to carry further on the bandwidth per wavelength or per lane, right? So now you can see -- you can give you another picture here now. The evolution of technologies is not only confined data centers. You're going to start to grab the technologies for transmission, local transmission actually moving into a data center as well, right? It's something that could happen in the next 5 to 7 years, right? Another point here really is this, going forward, the intellectual connectivities at 20G per lane may be more limited now because, again, the Moore's Law is reaching its limit. 200G per lane of studies beyond that is going to be very challenging. The question becomes, what do you do after that, right? And this is where optical connectivity is. This is where the opportunities of opening up the NVLink or any kind of inter-chip connectivity links using optics coming to play, right? That's the big idea we talk about 5 to 10x of the volume at that time, right? So that's going to be a mixed huge opportunity, huge challenge for optics on cost scale power perspective. So next slide, please. So this kind of gives you the flavor, right? Today, 1.6T, we have the EMLs, the silicon photonics, the PDs and CW lasers, all pairing up. Today, all these are getting designed into either modules, our own modules, our customers' modules, end customers modules and so on and so forth, right? So this is ongoing now, right? And this will see volume in 2025. Still [indiscernible] based and so on, right? Meanwhile, a lot of people are seeing the same problem we just talk about here. How are we going to scale further, right? You compare the power consumption of 800G to 1.6T, the power scaling now become minimal, right? There's a huge manpower consumption associated with digital signal processing. You hear that people talk about LPOs, LROs, all these different things and CPOs, it's all trying to solve the problem of power. We have power consumption, right? And so that's one area where people say, "Hey, can we integrate the whole thing together, remove all these DSP in the process and lower power consumption, good idea, right? We're supporting it. The external LifeSource technologies and Silicon Photonics now will be supporting that. It will take some time to prove it out, but that's element be a very important trend going forward. And that will then lead into the chip-to-chip connectivities based on optics, right, which is, again, to overcome both the power consumption limitation also to overcome the scaling of electronic link speed, right? Go beyond that, we're going to do, right? So that is the dual reason why people are talking about optical connectivity. Again, that will be a huge opportunity for us to drive up the volume even more, right? And of course, in that domain, there's another opportunity actually to connect them using VCSELs, right? You don't have using single mode optics in this case, because distance is really, really short, right? You could connect all those chips, high-speed chip with VCSELs. And that's another opportunity now to say, hey, maybe that's a good way to construct a low-cost, low-power consumption, optical connectivity in a short distance at chip level, and Lumentum is actually engaging all these discussions, all those activities. So we're not just looking at 1.6T. We have 1.6T, 3.2T, 6.4T and optical connectivity at chip level, how that's going to change our business composition and the volume going forward. So next slide, please. So actually, Alan talked on this already, right, as that -- if you look at each GPU-rated cost like a couple of thousand watts or something, increase amount of power consumption, right? So today, the limitation now of networking, the power is putting a big -- imagine how much compute we can put into the data center. So now you hear the word about the data center, the traffic used to be confined within the data centers now is leaking out of data centers, right? They need to be connected with dispersed the data center to connect them together, not at an inter GPU level of connectivity, but still very big, a lot bigger than what it used to be. So DCI now carry a different meaning that it could become a connectivity also due to AI machine learning clusters growth, right? So we see a lot of [ users ] there to start to drive the DCI market higher as well, right? We're at the early phase of that. We haven't seen huge amount of upside, but we can see that coming. So next slide, please. So in that case, we're getting really excited about the ZR. ZR used to be a cost reduction, efficiency improvement, management simplification play for the hyperscalers. Now it could be looked at, hey, I need to scale my AI overall bandwidth to connect different data centers. How we're going to do it? Lower cost and without having to putting a huge amount of DW infrastructure, right? So now we're seeing more and more of those kind of opportunities. Therefore, we're really excited about the 800G ZR products. Now we're actually sampling to the market, right? You're going to see some demonstrations. And now we call it work cloudification. Basically, people to build more and more point-to-point networks just like cloud to building the network. And then we're going to see more use of the pluggable transceivers in the transmission domain, right? That's a very exciting opportunity for us. We think the AI provision will start to drive the higher usage of that technology as well. So next slide, please. So finally, I come back to my first big idea is that you're thinking about you have 10,000 of more GPUs. Now you have to connect them with the lowest cost in lowest-latency way possible, right? Now the optical switching is coming into play, right? So how are we going to actually have this nonblocking, scalable -- scalable not just in terms of ratings or connectivity, but also for data rate, right? When your data is changing so fast every 2 years, if you have to change the cabling all the time, it's going to be really expensive. And power consumption is also a big challenge. And therefore, how about using optics to connect them, right? The optical connectivities or optical circa switch, right? So this is a technology now is being developed and we do see that will become a pretty important technology in the AMS learning clusters going forward. Going forward, you can even imagine we have the GPU level optical connectivity at a full data rate speed. You might use DWM in conjunction with optical switching, which basically means waferless selective switch kind of technology plus the free space, full fiber switching capabilities. That's also actually in the [ carts ], right? So imagine the data center at a time will really become like a telecom network, DWM wafers level of routing and so on and so forth, right? You can start to put the picture together. And again, this was getting me very excited because all this is used in Lumentum technology. Really, the wholesale from big networking to denser networking and everything that we have in our portfolio will be applicable to the future of AI machine learning evolution. So next slide, please. So to summarize, right, the story is in the next 3 to 5 years, you're going to see more and more of the speed increasing from 800T to 1.6T to 3.2T to even 6.4T. And the technology is all in the optical chips, how we're going to package low cost, how we're going to scale up really quickly at every 2 years of product life cycle. Meanwhile, the AI machine learning traffic pattern is encouraging the use of optical circa switch, and that will become a reality in the next couple of years, right? So that's kind of really in the next 3 to 5 years, right? And meanwhile, the leakage of AI traffic out of data centers is actually causing the growth of the DCI technologies, which you have amplification, you have ZR kind of technologies that's being used. We're going to see that going up as well. And finally, when the GPU traffic can really open up and carry by optics because of the limitation of Moore's Law interface speed, that's where you're going to see optics not going to next level of compute. Now we're going to see exponential growth of optics, the use of optics, 5, 7 to 10 years from now, right? Again, Lumentum has all the technology base, the scale, the road map to support this evolution. And that's what we're here for, that's what we're so excited about our future going forward. So I'll pass this out now to Dennis to bring us back into today and tell us how we're going to actually implement the near-term road map with transceivers.

Thank you, Wupen. Good morning, everyone. My name is Dennis Tong. And I joined Lumentum through the recent acquisition of Cloud Light. We started Cloud Light back in 2018 with a belief and a goal to set up a manufacturing platform to help -- perhaps we should go to the next slides -- with a goal to set up a manufacturing platform to help scale Silicon Photonics technology quickly. And based on that platform, which I will show in the next slide, we tend to differentiate ourselves by focusing on the latest generations of transceiver product. Can you -- can you go back one page, please? Yes. So what does that mean, right? That means if you look at this chart here, in the year before September 2023, 800G product represent approximately 37% of our revenue. One quarter after that, 800G product represents 75% of our total revenue. That indicates our ability to quickly migrate from one generation to another generation based on the platform that we have set up at Cloud Light. Next page. Before Cloud Light, if you look at the Silicon Photonics ecosystem, it is pretty fragmented. Starting from the left, you have a wafer foundry to make your design. You then pass your wafer to a typical [ ICOSAT ] which will do wafer level testing, dicing. And after that, you ship the dice individual device to a contract manufacturer to make your product, be it CPO based applicable transceiver. In Cloud Light, we aim to put all this process under one roof. And we believe that by taking ownership of all this process, we can enable data transparency throughout all the process -- across all the process. And by doing that, we can achieve better yield management, better efficiency and more cost-effective manufacturing. Another beauty of this platform is that once you have it set it up, because of the nature of silicon photonics technology, you don't really have to do that much change once you migrate from one generation to another generation. And that allow us to achieve fast -- faster product qualification and rapid time to market. Next page, please. On our platform, we are able to develop some very interesting in-house automated equipment. These are all designed by our Cloud Light team. And each of these equipment are customized for proprietary in-house precision micro-optics assembly process. I've shown a few examples here, including laser diode to lens alignment, silicon optical bench to the PIC alignment and also fiber attachment to the silicon photonics, all very critical in productizing silicon photonics. And together with our wafer-level testing, also the bare-die DSP co-packaging capability, we believe that we are offering a fully integrated product development capability here. Next page. Yes. So we talk about 400G and 800G. These are our road map. We launched in volume production 800G product last year -- earlier last year. And then the plan is that we will launch 1.6T product second half of this year and 3.2T 2025 and beyond. What I also want to mention is that we talk about silicon photonics product. At Cloud Light, we also offer a full range of multi-mode product, dating back all the way to 400G. And if you look at our track record of consistent deliver generation after generation of different products, we believe we are on -- we are very confident to continue to do that to support customer needs. Yes. Next page. Yes. So we talk about capability. We talk about product road map. I want to shift gear and talk about scalability. Shown here is Lumentum's manufacturing site in Thailand. Each of this building here, the footprint of it is about 7,000 square meter, okay? Phase 1 is already in full operation, and Phase 2 offer a lot of flexibility for further expansion and growth here. Yes. Next page. And then our customer go visit the site. They love it not just because of the scale of it but because of the way we manage the facility, including process control, Kaizen process and the culture of 0 defect. Next page, please. And the more exciting thing is that we are executing our plan to expand this facility per our previous announcement. Yes. Next page. Okay. So to sum up my part, we believe, Lumentum, we are in a unique position to capture the AI growth. We offer a broad range of product like what we have just described. Wupen has covered a broad range of component, and our Thailand facility certainly offer flexibility for us to grow and capture the massive AI opportunity. Thank you.

I think I've been introduced to everybody. Wajid Ali, CFO of Lumentum. Thanks very much for being here this morning, and I'm really sorry that the chairs we were promised at the back haven't come through. But we'll have to talk to somebody about that. So we'll present our financial model kind of post acquisition of Cloud Light in a couple of slides. And I think what you'll see is that we're taking a very balanced approach to the customer activity that Alan spoke about. I think all of you probably caught on to the fact that he said he's pretty much busy this week with customers, working through opportunities that we're seeing given the acquisition that we've just done. Those weeks have been very similar for the last 5 or 6 weeks as I've seen it. And I think because of that and the road map work that Wupen and Dennis' teams have been working on, we've taken a look at what we think our business is really going to look like from a mix standpoint moving forward. I'm sure this is not a surprise to any of you, but just to kind of fine point it and put some numbers around it. About 3/4 of our business was cloud and networking exiting fiscal year '23, and the other 25% of it was industrial products between our consumer business and our lasers business. And moving forward, our expectation is that our cloud and networking platform between all the opportunities that we're seeing, and quite frankly, what you'll see on the next slide, where we're investing our dollars and deploying our capital, our expectation is that will be greater than 85%. Not that we don't expect our lasers business to recover and for the new ultrafast products that we're investing in to support some growth in lasers as well, but just given the -- I would almost call it the step function opportunities and revenues that we're seeing from customers and multiple of them just between the hyperscalers and AI infrastructure providers, those step function opportunities are really going to shift the mix of our revenue base as we look forward. We'll start to see some of it in our fiscal '25, and it will be more pronounced in fiscal '26 based on the customer activity that we're having right now. Okay. So we're investing for growth, and we're investing for growth in the cloud and networking platform. And you hear a lot of talk from us on our earnings calls around lowering our fixed cost base. And really, the reason for lowering our fixed cost base is not just to improve our operating performance as a company, but it's also to make room for the R&D investments that we are supporting for different flavors and variants of 800G and for the DCI opportunities that Wupen talked about as well as 1.6T and 3.2 generation of products. And so we're not backing off from those critical R&D investments, but we are lowering our overall fixed cost base so that we can all do it within an envelope of operating expenses that's reasonable for the business. You saw the Phase 1 and Phase 2 pictures that Dennis presented. We've already started to take action on investments that we need to make from a capital standpoint, whether that's buildings or taking a look at what new tooling we need and what new equipment that we need in order to support the growth. And so we will see some elevated CapEx over the next 12 to 18 months versus our historical CapEx investments. But those CapEx investments will be very much focused on our Thailand facility and the equipment and infrastructure that's needed to support the transceiver growth that we're seeing. As all of you remember, we've already made a lot of investments in our Sagamihara facility. And so although we'll probably need a little bit more incremental investments for the 200G EML growth we're expecting to see, the majority of the CapEx will go to support our facilities in Thailand. And through those capital deployments, our expectation is that we're going to grow top line revenue. And you'll see that much of the operating profit improvements that we're expecting are going to come through that growth of top line revenue not just because of the incremental margin dollars that flow through from those sales but also the improvements in manufacturing capacity utilization. On the last conference call, we also stepped up our commitment to accelerate NeoPhotonics synergies. And so if all of you remember, we had come in with that acquisition committing to $60 million of synergies. 6 to 9 months later, we re-upped that to $80 million. And on the last conference call, we were able to increase that to $100 million of NeoPhotonics synergies. Again, opening up room for us so that we can reinvest in all the opportunities we see with the cloud and networking platform and also to keep us within an envelope that allows us to maintain a reasonable and balanced business model. So what you can see on this chart is the consensus estimates for fiscal year '24. And what we've done is you see 2 blue bar charts on the right. You don't have to take pictures. It will be out on the web. And many a times, as I've been invited to speak at fireside chats, I've constantly said that our goal #1 from a financial standpoint is to get back to double-digit operating margins. And for us to be able to get back to double-digit operating margins, we really need to be north of $400 million a quarter of revenue just given the shift in mix that we've had that all of you are very familiar with. And that double-digit operating profit really will come from the growth in margin dollars that come from incremental revenue but also through some of the COGS efficiencies that our teams are working through as well as the fixed cost reductions that we're also seeing. And the way to think about that north of $400 million of revenue is really just kind of recovery in our telecom end market, some recovery, some natural recovery in lasers and some normalization in the product transition that we're going through with the former Cloud Light business. So kind of once that normalizes, we'd be able to comfortably be north of $400 million a quarter again, and that should give us a double-digit operating margin. And then I mentioned right at the very beginning of my presentation, we're expecting to see step function improvements in revenue. With the activity that we've got going on for -- with new customers and new products, each one of those opportunities are sizable in nature. And so it won't be as linear as is the growth that we might see from where we are now to the north of $400 million. We should all expect it to be much more stepped in nature. And that's why we didn't put a year on here. We just said, okay, once we get to that level of quarterly revenue, we should be able to comfortably get back into a 17% to 20% operating margin cadence. That allows us to continue to invest in capital and the depreciation expenses that are associated with that capital, and it allows us to invest in R&D as well. As you can probably tell from Wupen's presentation, he's got a lot of opportunities that he can invest his R&D in. And so we want to make sure that we've got significant space for that. And so that's why we're communicating a 17% to 20% operating margin even at a $600 million quarterly revenue. So the business model. So again, you've got fiscal year '24 consensus estimates. And then we've taken the last bridge chart and basically put it into 2 columns so that you can take a look at what our expectation is on operating expenses and on gross margins. You saw on the very first slide that I showed, we're expecting a significant shift in mix between our 2 different platforms. And it's really that shift in mix and the types of new products that we'll be shipping that's causing a shift in the gross margin model for the company. You can see that we've got to continue to keep stringent controls on operating expenses while balancing the different R&D projects that we have in order to maintain operating margins that we've got listed on this sheet. Again, this will be on the web at 8:30. So I think I've talked about all of these things. We've got some structural cost savings that the team is working through between the NeoPhotonics synergies as well as some of the fixed cost reductions that we're actually making this quarter and next quarter. Those should start flowing through into our P&L. That will really allow us to give us some opportunity to invest in the R&D projects that we've got to fund the road map that Dennis and Wupen talked about. And we've got the balance sheet in order to be able to not only leverage our operating cash flow to support the CapEx that we need. But given the fact that it will probably be a little bit elevated for the next 12 to 18 months, we have a sufficient balance sheet to be able to do that and remain comfortable from a CapEx standpoint as well. Okay. So with that, I'll move it over into Q&A.

So I've asked that if you have a question, please make your way to the microphone, the standing mic in the center. Say your name and your affiliation. And we'll take you in the order of people lining up.

Perfect. so Alex Henderson, Needham. I was hoping, Wajid, you could talk about when the step functions are likely to happen. It seems probable that telco is, at the earliest, 1Q '25 calendar year. But the optical piece is -- there's the chip piece for the optical elements. And then there's the ramp in capacity produced AI products, whether that's 800 gig or 1.6 terabits. When are those 3 likely to kick in? My guess is that it's probably December '24 for the chips and for the AI products ramping into the March '25 quarter. Is that kind of the window?

Probably Wupen is better to start on that, and then I can back them up.

So can you repeat again, a question about -- the specific questions?

So basically, when do the step functions happen on the cloud-like revenue based on our...

And the 200-gig chips.

And the 200-gig chips.

Yes. The 200-gig ramp, let's start with that first, right, Alex. It's going to be, I would say, second half of next year in volume, right? So calendar '25, right? So today, it's all about sampling and getting qualification ready. But I would say the real volume from 1.6T is going to be second calendar -- second half calendar year 2025. The 400 or 800G or kind of current generation products, I think it's a little bit uncertain at this point. I wouldn't put a date on there yet. A lot of activities going on right now. But I would -- I think we'll probably need for more updates from us a little bit later.

Yes. I mean I think sampling of the chips are happening now.

Now.

And sampling of the transceivers will happen this summer, and then it will take some time to qual and ramp into production. Telco, I think that it's a matter of the carrier spending, right? It's slow and the inventory is there. So I still think that's probably more of at least the first half of fiscal '25, so the September or December quarter before we have normalization. you want to translate that into money?

No, I think you covered it.

Tri Lam from Ardsley Partners. A quick question on the gross margins going from 33% to 37%. You talked about product mix lifting that up. What about the impact of customer mix, right? Customer mix, deal size probably drives more margin than products, right?

Yes. I was trying not -- you picked it up very well. Yes, I was trying not to highlight that. But yes, the product mix is for new customers, and new customer margins are better than prior customer margins, at least from a deal activity that we're seeing right now. And so that in combination with the fact that 200G EMLs are a chip business. And so that will give us more of a tailwind on margin and the 800G ZR+ products will also give us a tailwind in margins. But yes, on the kind of classic 800G and 1.6T products, we're seeing better margin opportunities than what we've historically seen.

Due to customer mix as well.?

Correct.

And then second question is, you talked about increasing in the numbers of transceivers. But I think you're kind of assuming that the spine remains 64 ports on the spine. When you look at NVIDIA, they just released 144 ports, which means you can now collapse like a 9,000 or so GPU cluster that used to de required 3 networking tier. Now you can collapse it to 2. If more and more of these switches -- or the spine switches coming at 144 or even higher node, wouldn't the number of transceivers you need to build address the same number of clusters because of the way that we did NVLink? It doesn't have to follow that natural pathology that you have these in these spines. So therefore, you can actually get less transceivers if you build more dense spine, right? So isn't that...

So basically, I agree with that, right? So today, it's like 3:1 ratio between transceivers versus GPU, 3 to 3.5, right? And with that condensing, 3 or 2 layers beyond like 2.5 to 3. Okay. So yes, but still a big jump, right? GPU will continue to be deployed. And remember, that's only for one GPU, right? GPU vendor. There's actually a bunch of others with a call -- that in aggregate, the volume is still huge, right? There's still a step-up of the so-called front-end network, traditional data center networking, right? So there's a step-up. And people are trying to use less transceivers, less links, but the step-up is still there, right?

One thing to add to that the objective in doing that is to be able to build more, right? The goal isn't to save money or to reduce. The goal is to be more efficient, so you can build a much larger cluster. So I think not lose sight of -- it's not a 0 sum. It's actually the intention is to lessen or minimize what is a limiter so that more can be added overall.

Okay. So more dense but bigger networks?

I think that's the key point.

So Alan, I saw you at the event yesterday as well at Optica, and I think you heard on stage an NVIDIA representative talk about the cost differential between copper and optical. And I think Wupen showed a slide about 800 going to 1.6T and kind of showing single-mode fiber coming into both the high-speed [ radic ] side and also on the front end side from server switch. What makes you feel like you can get the cost profile down? I think yesterday, it was like $0.50 a gig to $0.05. Like that differential is pretty large. How can the cost profile come down enough for you guys to take those sockets where copper exists today in such volume?

Okay. So I think that's the biggest barrier, right? So today, I talk about these 2 steps, right? Step number 1 is optical transceiver, we talk about it already, right? To bring it down to $0.50 per gig to like $0.05 per gig and having like 2 gigajoule per bit of power consumption. It's challenging, right? And therefore, it has to come down to optical integration intimately, right, with the silicon, right? That's where silicon photonics really does come in, right? But how to achieve that? I don't think anybody knows it yet, right? So that's one area of evolution or innovation that's really sorely needed by the industry, by us. We're definitely part of that equation. Another approach to doing that, we talk about this VCSEL-based interconnect, right? That's the way. The VCSEL today already has scale. It has the sensing-based scale, right? It has a pretty good power consumption compared on the power per bit basis, [indiscernible] per bit basis. And therefore, that's a potentially attractive solution to achieve it. Can we get to like $0.05? I don't know, right? But it's at least in the right direction. So that is where the optical innovation will have to happen to scale beyond the limitation we talk about, the Moore's Law, the [ electrical ] bandwidth. very good question.

The other one is just for Wajid. So like you showed various scenarios for fiscal year '24 that had a pretty wide range, like Street $450 million the run rate, $600 million run rate. Does some of that need to happen to get to the $450 million or the $600 million? Or can you just describe what are the variations between kind of your outlook?

Yes. So there wasn't any year put on it. And a lot of it is because of the timing of the telecom recovery. So the way to think about getting us back to north of $400 million a quarter is recovery in our telecom business, some normalization just between the product transitions we're seeing in the former Cloud Light business and kind of normalized recovery in our lasers business. So we certainly expect that to be a fiscal year '25 mode. Is it Q1? Is it Q2? That's the part that we really don't know. And then the step function from $425 million up to the $600 million is really the opportunities that we see from a customer standpoint that we're currently bidding on and taking -- just taking a portion of those. If we took more of those, then it wouldn't be as balanced a financial model. But even kind of one of the customer activities can comfortably double our datacom business even if you take a look at some conservative numbers. And so that really needs to happen in order for the $600 million a quarter to happen.

Maybe just to give you a little bit more color. A lot of the discussions I'm having with the hyperscalers today is how do I reserve that capacity you're putting in place today, both at the EML and in the back end. And so that's a good discussion to have. Now they have to sign up, and we have to execute our product development, and then the business will come. So I'm confident it will come. The question is, is it the middle of '25 or the beginning of calendar '25 or the end, right? So that's kind of the question of when are they going to be ready for 1.6T because that's really what's going to ramp up strongly in Thailand.

Ananda Baruah with Loop Capital. Two, if I could. On the new customer quals for Cloud Light, what are the meaningful hurdles that you guys are working through the thresholds? And is there any reason -- I guess really what I'm wondering is kind of probability weighting, the opportunity to get qual-ed at the hyperscalers, the remaining ones and way to think about probability weighting like legitimate inclusion at volume once the quals occur. I'll stop there. I have a quick follow-up as well.

I'll give you my perspective, and then maybe Dennis can chime in as well. I mean the gating item for 1.6 terabit is the DSP. So today, we don't have the DSP. We're going to get it soon. Now it needs to work. I think on the optics side, the silicon photonics, CW laser and the EMLs, that's -- we're ready. So it's a matter of putting it all together and letting Dennis do his magic.

I will echo that opinion. I think it's an industry-wide challenge that we are all gated by the readiness of DSP. Optics front end, I think we are there. And I think the 200G per lane product, be it 4x200 or 8x200, I think we are making very good progress in terms of getting customers to qualify it.

Great. And then on the differentiation, how would you guys describe the Cloud Light differentiation? It sounds like the manufacturing process is part of that. Is there anything else? And how well -- like what -- do you consider yourselves having like any sort of like meaningful moat on the manufacturing differentiation as well? It can't be replicated in any near-term period of time?

Well, I'd say most of the transceiver manufacturers today buy their lasers, right? Many of them -- most of them buy them from us. And so that vertical integration of EMLs, VCSELs, CW lasers, silicon photonics, all within our house, gives us a -- certainly a cost differentiation but also a time to market radiation. So I think that's a moat. Is it sustainable for the long term? I think cost, it is. Time to market, I think eventually, they will come along because we're going to continue to enable them with our laser chips. So I'd say that's -- and what Cloud Light has done with respect to the manufacturing infrastructure and capability is quite amazing. So I think it's a combination, but it really comes down to how you can design something for lower power and ramp it up as fast as you possibly can.

I'm not the technical guy but...

Yes, you're right.

But the capacity investments that our customers see us making and Alan mentioned earlier that the discussions he's having with leading hyperscaler customers are around capacity reservation. I think that, that is giving us an edge from a competitive advantage standpoint because we are investing and we've got something in place. And so the time to market for our customers and security of supply is really important, too. So I'm sorry, you were going to answer that.

Yes. So people start to talk about scaling up the volume of silicon photonics. The platform I show, we actually start to build that 5, 6 years ago. And it is, in our opinion, a proven platform. I talk about shipping volume of 800G optics is coming out of that platform. And we are still perfecting it, making it more efficient and cost competitive.

Chris Rolland, Susquehanna. I have a few. Kathy, you can cut me off if I go over. I guess first of all as we think about 1.6, what do you see Light's share, let's say, in the optics market, the laser market? And then, Wupen, you had some interesting things to say regarding kind of the single-mode, multi-mode VCSEL versus EML debate there. I think 200-gig VCSEL is still a viable lower end market. Maybe talk about how that shifts and put that into answering that question. Where do you see Lumentum's share in the laser market at 1.6T?

Yes, I'd hate to -- I would say our share is going to increase given the challenges in going from 100 gig per lane to 200 gig per lane at the optical component level. We have, as Alan alluded to, a very healthy market share at current lane speeds. So I think it's only going to increase. I don't know if you want to add to that.

Yes. So on the, call it, single-mode laser side, right, the 200G per lane is actually very challenging. It's when you go to higher data rate, all these RF problems becomes exacerbated. So we're already solving those problems with customers and with solutions solve them at the chip level. So we think our share is going to grow in the single-mode space. And then on the multi-mode side, I think 200G VCSEL, like you said, we've heard announcement this week, right? I think that connection between the server to the switch, for example, like you can still use multi-mode devices, right? But I think it's just the way that now the GPUs are scaling, right, and the distance limitation. I think corresponding share of the single-mode device is going to increase, and that will be at the expense of the multi-mode devices, right, in the current architecture of connectivity.

And I think maybe to add to that point that in these AI clusters, right, I mean, you've heard about copper and lots of excitement about announcements last week. Reality is you're only talking about a meter, 2 meters. So even at these higher speeds, copper cabling can scale. The issue is then, as you go more than that, a few more meters, and all of a sudden, the window where multi-mode for a single mode crossover versus multi-mode is incapable of going at 200-gig line rates. All of a sudden, it pushes towards single mode. So sort of this window of where it's applicable, but you're talking about an AI cluster, which is a little bit different than a broader data center, really collapses. And so we expect to see a lot more single-mode cabling. Now I think a point that's come up that I want to reiterate is we're talking about cabling. But as Wupen also alluded to, down at the chip level, being able to communicate a 200-gig per lane electrical, going even from chip to chip or within a circuit board for those probably not attending the technical sessions over the last few days. But this is sort of the industry big deal of why pushing optics closer to the chips isn't to replace that 1 meter of cabling within the rack. It's the fact that to replace centimeters of circuit board at 200 gig per lane is where optics is starting to push much more closer to solve that problem. And as it solves that problem, it's already in the optical domain. They'll probably stay in the optical domain.

I mean there are 3 big players in the laser market. Do you hazard a guess size to your share at 1.6?

I don't think we want to speculate on share. So I certainly think we have more than 30%, 40% market share.

Okay. For my next question, perhaps, Wupen, you mentioned the WSS opportunity. I know last year, Google was showing off an optical switch. Can you talk about optical switching opportunities with WSS and if Light has some revenue opportunities there over time?

Yes. WSS is still early, right? We're not at the routing at the [ wavelengths ] range, right? Right now, it's too early. I'm just giving you the kind of the road map, right, in the future. I think eventually, that DWDM level connectivities end point to end point will probably become needed within data centers, right? But that's years out though. I wouldn't speculate on when that's going to become the case. But I do think that the optical circa switching at the spine level, right, a few hundred kind of [ ratings ], I think that could happen in the next couple of years, right, because that's a good way of scaling AI and machine learning without really -- I mean, with all the -- if it's the traffic pattern and it also lowers the cost, lowers the power consumption, I do see that happening in the next couple of years. But WSS is probably years away.

So switching in the entire fiber is what we're working on that goes into the data center, correct, within the next 12 months. And I'd say that wavelength selective switches when coherent comes into the data center, I think that's when wavelength will potentially play a bigger role, but that's a ways out there.

Okay. Understood. And lastly, vertical integration of your various products into the Cloud Light transceiver, if you could talk about those opportunities. And then inorganically, would there be an opportunity to acquire some products to put into that?

Well, I'd say November -- middle of November, we started shipping them samples. And so we're building EML-based transceivers where most of their products have been silicon photonics or multi-mode. So that's happening. And I think that as we introduce 1.6 terabit, both the team collaboration on the silicon photonics PIC and the EMLs based on different customers' requirements will be going into those products in calendar '25 in a more meaningful way. So I think we'll get a pickup of margin as a result of that as opposed to having Cloud Light go on the open market and buy those products. VCSELs, evaluating the VCSELs as well for multi-mode applications, but [ PDs ] as well. I mean that's all happening. I'd say probably more of a -- I think Wajid had it on his between $425 million and $600 million. There was a product in-feed or vertical integration, and that's probably in the same time frame that those products would come in and give us a bump in margin.

Yes. So I'd add to that. When it comes to whether it's silicon photonics, indium phosphide components, gallium arsenide components, that's capabilities we already have within the company. So the time to -- it's really a time factor of developing the needed components if we don't have them or qualifying them with customers if we have those components for that level of vertical integration. Obviously, you heard about Thailand from a factory and manufacturing. The missing piece, yes, there may be other little optical widgets that we don't have today because to be in that business from a component standpoint wasn't necessarily attractive. Now that we have an in-house customer, if you will, those are kinds of things we're looking at.

This is [indiscernible] from [indiscernible]. So just a clarification. A couple of clarifications on the GPU to optical ratio. You mentioned previously it's 3 to 3.5 to -- the current structure is 2 to 2.5. So I just want to clarify whether the 3 to 3.5 is referring to the H100 HGX previously. Version now is more like NVLink GB200. Architecture is going to be 2 to 2.5.

I want to be very careful about those kinds of numbers because they depend incredibly on the size of the GPU cluster. You're mentioning specific customers. I mean there are other customers. But I think the answer is that the transceiver -- individual transceiver bandwidth is less than the bandwidth of the I/O on a GPU. So there's going to be multiple period. Now it depends on the level of networking levels, if you will. There was a gentleman that asked a question about recent advancement in switching, being able to collapse the number of layers. Yes, that reduces. But then the whole goal of reducing the number of layers in a cluster is to be able to then stack more clusters together. So I think you're going to see numbers that continue to be multiples, and they're going to go up and down depending on the size of the cluster, the relative transceiver line rate relative to the GPU.

Just for the 2 to 2.5 roughly, that number, is it referring to 800 gig or 1.6T?

It's really the -- like the gentleman talked about, right, when you have the -- it doesn't matter what the data rate is, right? We have the network architecture on 3 layers to 2 layers, right? So I think that's the ratio. Whether it's 400G, 800G, 1.6T, it doesn't matter. It is how you actually put them all together, right? It's more of a structure question rather than data rate question.

Okay. Okay. Got it. so like the ratio could vary a lot if it's 1 -- like the cost of the GPU if it's 5,000...

The larger the cluster, the more links, which means more transceivers per GPU. And the goal is to build bigger clusters, right? Don't lose sight of that key point.

Karan Juvekar from Morgan Stanley. So on the slides, you mentioned sort of a pretty great sequential ramp on 800 gig sequentially. Just in terms of 1.6 terabyte, do you sort of see the ramp in production time line similar to 800 gig? And I guess revenue opportunity, do you still see a similar step function as you saw with the 800 gig? Or do you see any differences between the 2?

Yes, we expect a very similar ramp signature, if you will. And I think this is also AI and machine learning signature. It comes in a burst. And I think we have shown our ability to cope with that, switching from 400 to 800, and we expect to do the same thing at 1.6T.

Okay. That's helpful. And then on the CapEx investments that you mentioned, I guess, is there a lag to the time line of how that turns into revenue? Is there anything where the CapEx is sort of prioritized that we should pick up and how quickly that can turn into revenue?

I mean if we order it, it comes in 4 to 6 months later, and then we put it in service. So that's really the time that it kind of impacts us. So it can be anywhere from 6 to 12 months.

Okay. So the CapEx investments getting to that $425 million or $600 million quarterly, those aren't baked into those, are they?

Those are already baked in, in the depreciation calculations we have.

I have a clarification on the time line that you guys saw for 1.6T. I think you said on the last earnings call, expected to have 1.6T transceivers later this calendar year. In a prior answer, I think you guys said kind of first half of calendar '25. Was that the 1.6T products you see ramping initially, those are silicon photonics based and EML based? You don't see until the first half of calendar '25? Or can you level set that difference?

So I think based on specific customer needs, we'll be doing both sampling this summer of 1.6-terabit silicon photonics and EML-based transceivers. Question comes how fast does that turn into qualified transceivers that are ready to ramp, and that's really probably more a calendar '25 thing as we iterate and have the customer qualified and be ready for that 1.6T, which is not trivial.

Yes, remember the context of the question was when is that step function in revenue, and the point is that it does -- there is a -- while it -- once you have the chart and it's done, it looks like almost a light switch being turned on. There is a delay between when we launch and get into the customer and when that light switch turns out.

And by the way, we're getting ready to turn on the light switch at the end of this calendar year. So we'll be ready for those customers if and when they're ready to go in January.

And there's at least 2 different major compute platforms that will be utilizing 200 gig per lane by the second half of this year. The kind of bottlenecks to scaling the Lumentum 1.6T trajectory, like is that more a function of kind of Cloud Light specific qualifications at customers? Or is there another industry ecosystem bottleneck, whether it's switching or network interface cards that kind of pushes the time line from compute to optics, high-volume production?

Yes, it's the latter. It's the latter. The 100G transition to 200G is a very, very big deal, right? So everything has to work. And then I think the optical portion of it is probably relatively deterministic because we already have been working on it for a long time, right? But the transition when you integrate everything together, switch and everything together, that can be pretty challenging. So therefore, I guess, the turnaround -- everybody is increasing the turnaround timing. But I think there's uncertainty there because the whole ecosystem has to be ready. And that's the big one. So we have to watch very closely.

Let's take George as our last question.

George Notter from Jefferies. I guess I'm curious about what's realistic in terms of market share with Cloud Light. You've got a major customer right now. I think you're the third supplier, if I remember there. I think you guys are doing some customization of the product rather than shipping maybe more standards -- more standard off-the-shelf type product. Can you just talk about -- are we competing for third sources and fourth sources among these other cloud provider customers, a la the Google experience? Or is it -- do you have an opportunity to kind of displace some of the other transceiver manufacturers that are historical players in the space?

I think it comes down to execution, right? And the doors are open, and it's up to Dennis to execute. And if we execute, we'll be first player. And we're not striving to be the third person in the door where prices are not as good as the first one in the door. So that's certainly our goal, and I think we're lining up to be able to do that.

I will not describe ourselves as a third supplier. I think it depends on the product. I tend to think that in certain products, we are actually doing well. Yes, and the rest of it to broaden further customer portfolio is up to our execution. And I think we have all the right tools to the that.

And maybe, George, one other way to kind of come at your question is today, let's put us -- or today or last year, there's a very limited number of transceiver vendors that are shipping the kinds of volumes that we're talking about. So now you've asked the question of, can we steal share from those folks? And I think that's the right way to look at it. Do they have the right profile? And as we've highlighted between being vertically integrated, the manufacturing footprint outside of China and the -- changing the nature of who Cloud Light is by being one company and changing the nature of who Lumentum is by being one company, I think those are factors that play into our success, whereas I don't think any of the other top 3, let's say, transceiver vendors have any substantive change in who they are or what they're doing. And in fact, the technological challenges and the geopolitical challenges probably put those competitors in a much weaker position moving forward.

Is there anything different about this market that kind of speaks to higher margin structures? Like I get the vertical integration discussion, but as I think historically about the transceiver business, it's been tougher margin business. You guys exited at one point. Like I get that the scale here is different. AI is here. It's a big driver of all this. But I think embedded in that 17% to 20% operating margin kind of profile that you put out there at $600 million a year, I mean, you're implying this business is significantly higher margin than maybe it has been historically. I guess I'm just curious in your comments on that.

Yes. I mean I think it's not as high as our old gross margins used to be, but the operating margin should be fine, especially with the scale, right? So we spread our R&D and SG&A over much more volume and much more scale. So I think that helps us at the operating margin level. And I do think that there's a big benefit because our -- if you remember right, our EML chip business is very high margin, right? And so that's a product in feed that now Dennis gets at cost. So I think from that perspective, that should help significantly on the transceiver margins.

Okay. Great. Thank you, George. And thanks, everyone, for attending. If you would like to come to our booth, we're going to have demos at noon today and tomorrow at 10:00 a.m. So we'll be open for investor booth tours at that time. So thanks, everyone, for coming. This concludes our webcast.