Welcome back to this afternoon's session at the BofA Global Tech Conference. I'm Vivek Arya from BofA's Semiconductor & Semi Cap Equipment research team. Really delighted and honored to have Wupen Yuen, the Head of Business Unit at Lumentum, join us this afternoon. I'll go through my questions, but please feel free to raise your hand if you would like to bring something up during the discussion. With that, very warm welcome.
Thank you.
Really appreciate you-
Thank you.
You being here with us.
Thank you.
Maybe let's start at the high level if it's okay. I see Lumentum kind of involved in so many interesting high growth areas, right. Whether it is lasers, right, whether it is optical circuit switches or transceivers, right, just silicon photonics, right, multiple different areas. What are the key kind of operational and technical areas that you are spending your time on to make sure, because there is no lack of demand, there's no lack of growth, but how are you kind of optimizing? What are the top two or three operational things that you are keeping your eye on?
Right. First of all is the fab, right? I think the fab is the most critical part of it. Our value is in the fab, primarily. We're really making sure that we have the right teams, right? One of the things that we are doing today with Lumentum, I'm seeing Chuck here, right? These fab are actually like a temple. There's a lot of very experienced people running the temple. We actually have make sure that our fab people are fully supported, right, with the right tools, the right automation to ensure they can actually maximize their efficiency. Number one is the fab. We spend a lot of effort on the fab. Second thing really is thinking about ramping all these kind of assembly and test operations is not easy, right? OCS is a complex product, and the modules is rapid cycle products.
These ramping are challenging, right? What we do here really is say, "Let's put our best resources, best engineering talent on ramping these complicated products." Frankly, we're actually moving our legacy products out of our own factories, moving to the CMs for the CMs to help us manage those kind of products, right?
Right.
The third piece really is supply chain, right. As now AI demand going through the roof, everything from PCB to garnet to lasers and PDs and DSPs, everything's kind of short. We're also spending a lot of time managing the supply chain piece. Make sure that this do not become the bottleneck. Really the fab, the test and assembly piece, and also the supply chain piece.
All right. Makes sense. Wupen, you have described almost kind of a 30% imbalance, right, between?
Right
supply and demand on your EML lasers. When do you think that will close? Like what needs to happen? Is there a certain time frame at which you see that closing? As you look outside, Wupen, do you see competition catching up and filling, right, some of that gap?
Yeah, we're seeing now, let's think about where we are in the AI space, right? Today, all of our so-called the supply-demand imbalance is really still in a scale-out phase, right? The scale-out's already taking all of us, all of our wherewithal, and we're still short, right? What we're seeing also is that optical scale-up is happening in 2028. When the optical scale-up happens, the optical intensity increase by at least 3x, if not more, right? Frankly speaking, if you look at, we're struggling as an industry to even deal with the scale-out demand for this couple years, and then before you can really fully get your head above the water, the scale-up demand actually happens, right? Frankly, we're not too concerned about there's more capacity, more competitors joining the game. We're actually focused a lot more on ramping our own capacity, right?
What we see is that the CPO demand is as strong and on time or even stronger as before. We see the scale-out EML demand stronger or equal or stronger than it's before. There's more demand for CW lasers. We're seeing this tsunami of scale-up happening in a couple of years, right? We're not overly concerned, frankly, by competition, but we're actually much more focused on ramping our own capacity and driving as much efficiency as we can.
Got it. You made an interesting point that the specifications of these laser, right, I mean, it's a similar laser.
Right
Same technology, right, for these EMLs can vary as you go from scale-out to scale-up. Could you expand on that? Like, what has to change?
Yeah. I think there's also a very important point on the competition I want to emphasize a little bit, right? Not every laser is built the same. Not every one is the same, right? For example, right, let's take the UHP laser or CPO laser, for example, right? It has to go through a very special modulator in silicon photonics in order to get the density of the CPO the customer's looking for. That requires a very narrow linewidth of the laser, very low noise. That's hard to make. Right? That as yourself is a very high bar to overcome. For the CW laser, for example, right? When you go to a higher frequency CW laser, you actually are sacrificing certain features of the laser from design point of view, that makes yielding the laser a lot harder. Right?
Therefore, we are a user of different CW lasers too. We see a drastic difference between brand A to brand B to our own, right?
Right
in manufacturing at the module level is a really big deal that people do not talk about. Right? That's also really important. There's a major differentiation there, and this is frankly where Lumentum's pricing power actually exists. Because our customer telling us when they use our lasers, they actually get a higher module yield. Now, at the module level, a few hundred dollars per module, a couple of percentage difference in yield is a lot of money.
Right.
That's also really important, right? Third, at the EML situation, right, when you go to 100G 200G EML, it becomes harder to make. You also have to go to the structures that will reduce the interference between different channels. That requires a different feature on the EML, right? That also make it harder to make, right? As we progress toward scale out and then scale up, the challenges, because the higher density of packaging and the higher density of bandwidth makes the design features that much more important, and then the yield consideration that much more important. Therefore, not all lasers are created equal.
Right
at the final solution level.
Got it. That's a good point. Does that show up in pricing? Because people have always thought of the optics industry as kind of even more deflationary than the broader semiconductor industry. Do you think it is helping from a pricing perspective also?
Yeah. We hate to say this, right, but in the.
I mean, you like to say this.
We like to say that we're over that phase. Right? Previously, I've been in the industry for 30 years, and Chuck has been there for 30 years too, that was a kind of zero-sum game market, right? Today, the market demand is growing so fast, right? If you have a good quality you can supply, the pricing power does exist, right? We are seeing, certainly, there's a structural level pricing. We're also seeing opportunities for up charging when there's a last-minute hot demand that when competitors fail to deliver, and we see opportunity there for us to also up charge based on the local supply and demand mismatch as well.
Got it.
Yeah.
I wanted to get your perspective Wupen on the rollout of CPO. Right? Maybe first start with what is kind of the most important milestones that you're looking at, how will it evolve? Then the question that I had, right, following that would be, there's a lot of talk of CPO from the suppliers, not yet from hyperscalers quite yet. Maybe first start with, how do you see the timeline of CPO deployment across your customers?
First of all, on the CPO side, right, we're seeing our ramp of CPO lasers to be on schedule, the pressure from the customers are the same or increasing. Right? That means that they are all in on the CPOs. In fact, we heard some of their direct customers, the neoclouds are saying that definitely CPOs is on schedule. I want to kind of bring the stage out a little bit. What is driving CPO? Why is it very important, right? The CPO is never a scale out story. Right? The CPO really is a scale up story. Right? If you look at the AI hardware racks, right, the scale up capacity or bandwidth is 10 times that of optical scale out. When you go to 10 times the bandwidth, it's hard to say, "I'm going to use pluggable modules.
Right. Definitely.
The power consumption, the cost, the size is just crazy, right? That's one thing, right? Because of the inference, the AI inference traffic now growing so fast, I just made a statement that we're seeing now the inference induced optical scale up ramping happen is kind of 2028. To meet that timeline, that's what's fundamentally driving the CPO because you need the density of CPO, the power consumption of the CPO to really enable, right.
Right
optical scale up kind of timeline, right? That really fundamentally is driving it. Therefore, we're seeing more urgency in the CPO adoption. We're seeing more urgency on the high density optical interconnect adoption for that particular reason, right? Therefore, again, from the laser supplier perspective, we see that trend continues, right? For the next couple of years, we're still, it's mostly in a scale out domain. In 2028 and beyond, once the scale up happens, there's inflection of the optical intensity, right, in the AI racks. That's going to be another huge opportunity that we've been talking about, right, since the OFC time.
Got it.
Yeah.
Does that transition leave a hole that because CPO is implemented in scale out first, but that's a very temporary part, does that then leave a hole that scale up then has to fill? Do you think, as you described, scale up is such a massively larger opportunity that it shouldn't be a problem?
Yeah. I think the eventual state will be the following, right? The scale up is a very proprietary setup. Right? Hyperscalar A to B to C, or GPU A to B to C, they have a different kind of design for the rack level, right? Therefore, scale up CPO is very straightforward. It's my own design. I own the racks. I do CPO, right? The scale out is different. Scale out, you're connecting different racks together. That's where you may not have to use CPO. You could use DR optics, FR optics.
Right
LR optics, right? That's where you can actually use pluggable. It's okay.
Right.
Right? Again, I would just say that the scale out CPO is a first step, is to really trailblaze and mature the ecosystem. Right? That's not the final destination. The real reason for scale up for CPO is in the scale out. You will see, I think, a mixture of optics used for scale out, including CPO.
Right.
You're going to see a lot of CPOs in the scale up applications because they self-contain.
Got it.
That's how we're seeing the world developing.
Got it.
Yeah.
Where does NPO fit in this?
Yeah.
Is that a temporary solution? Does it have legs? What do you think about it?
Yeah. NPO is a great question, right? Here's our view. Still coming back to the scale up story, right? When people realize that scale up has to happen because of the inference impact, people figured out that our view is that they're not ready for CPO yet. You cannot use pluggable, front panel pluggable. What do you do? Right? Okay. Well, let's use an NPO, which is a socket-based.
Right
which is more like linear optics, close enough, not quite as good as the CPO inter-power consumption, but it's much better than front panel pluggable.
Right.
It is really an intermediate first step, to really for time to market reasons. What we're seeing as a general conclusion is that the next generation after the NPO will be CPO. Everybody's focusing on CPO.
So just a temporary-
Temporary thing. We really think, okay, there's going to be two group people in the first generation of the scale up implementation. It will be CPO + NPO, but in the next phase, the Generation 2, everybody will be focused on CPO. This is really, in our opinion, NPO is intermediate step.
Got it. On the CPO, the question I really wanted to ask was, I think it's clear NVIDIA has described their roadmap. They have kind of very clear timeline of when they want to get to CPO. If you're not deploying the NVIDIA solution, are you still seeing enough of a pull? Yes, I think there is a Marvell, Celestial solution with Amazon as an example. Are those really kind of the two alternatives, or do you see a more diverse set of customers? We don't hear yet from the hyperscalers, because this is brand-new technology. They're already putting in so much complexity. What do you think? What do you hear from the end customers about the pace of adoption? What are their practical challenges that need to be addressed before CPO becomes a bigger deal for them?
That's a very complicated question. Every hyperscaler is very different, right?
Right.
One hyperscaler is very much, because of unique architecture, is very much a pluggable play. It is OCS pluggable.
Right.
That's one group. That customer, they're probably not so hot on CPO. There's definitely NVIDIA enables every single hyperscaler to use CPO, so that's one thing. We also see another hyperscaler that's all in on NPO. That's again because optical scale up. We see a couple more customers who are more looking at, okay, I want to go to CPO, but I'm going to use a merchant solution first. There's merchant solution on CPO, including the switch. That's also where they're going to go with. I would say that, for next two years, two to three years, you're going to see the pluggable customer with their own architecture. You're going to see the NVIDIA with the CPO architecture, and you're going to see some other hyperscalers.
Right
say, hey, have our own, whether it's a high density XPOs or whatever, or solutions. It's going to be a mixed bag. That will change. After 2028, it will change to be more on, look at this OCI MSA. Who's on the OCI MSA? Every single hyperscaler on the OCI MSA.
Right.
The OCI MSA has to be done by silicon photonics, by ring modulators, by slow and wide optics, DWDM. That is the ultimate solution. That's where everything going to converge. It just, everybody has a different starting point.
Right.
We're going to actually gradually converge onto CPO from kind of different angles.
Got it.
The final solution is pretty clear.
Does it make your job of forecasting tougher? You think that, yes, all these things sound very fragmented, but you have very good line of sight, good visibility for each of these implementations that you can roll up for your factory forecasts?
Yeah, I think we have a good visibility. We have a good view on which hyperscaler is doing what. We're engaged in, from the laser piece, we're actually engaged very well in all these different opportunities. We also engage on the module side in some of the major application of modules. I feel we have a really good understanding, and also there's multiple LTAs where we have with the customers. That also give us the visibility of the product types, product mix, and the volume.
Got it.
I think we're pretty well-positioned. That's why we're pretty aggressive in ramping our capacity, because we have these kind of visibilities, and in most cases, the commercial assurances as well.
Got it. Makes sense. Next thing on optical circuit switch. Again, from the outside, it seems like there is one big hyperscaler who has adopted it, but Lumentum has described engagement with several hyperscalers. First, the use case, is this still kind of limited more to high density kind of training clusters, or do you see the role of OCS expanding in inference also? How is it doing that? Then are you seeing engagements with multiple customers, or is it still kind of dominated by one?
That's a really good question. Today, we have been engaged with two kind of opportunities. One is more of a spine level replacement, whether it's really data center spine or campus level spine or scale across spine. Really got spine level replacement. That's a smaller opportunity. The primary opportunity here really is the optical scale up in the particular torus kind of implementation, right?
I see.
That's majority of the volume, right? What we see actually going forward is that when the CPO or optical scale up happens, then there is a interesting phenomenon we're thinking really is that that's the time when you are scaling the inference clusters from one to eight or 10 or whatever, go to 1,000 XPU clusters. That's where OCS becomes a really interesting opportunity because then you can have different instances, you can have different models, you can actually serve, you can connect, and then should leverage the whole clusters and not wasting any resources, right? That's when we see that the OCS group will become a really good complementary technologies to go with optical scale up. Right?
Got it.
we think there is an interesting potential convergence.
Got it.
of optical scale up and OCS, right? Not right away, right? The world takes time to really learn how to operate in that domain, but I think there's a multi-year trend that this could become another major growth opportunity for OCS on top of the current application and on top of the scale across and the scale out applications of OCS.
Got it. I know with CPO, right, you have a major system developer, right, really kind of promoting it. Do you see some kind of similar trend? Yes, OCS has been deployed by one hyperscaler, but for it to be widely deployed, do you think some system vendor has to really start pushing it, or do you see the organic pull from end customers to deploy it also?
That's a good question. I would say that now I think it's pretty clear that the inference requires clusters, racks, right?
Right.
Racks and clusters. We see hyperscalers thinking about how they can build their own clusters, right? Really for high efficiency inferencing. I think the XPU guys, they're all thinking about it.
Right.
To learn how to use it will take time. Google spent many years perfecting the use of OCS, right? The software piece is pretty complicated, right? I think that whoever want to use OCS going forward needs to go through that learning cycle. I don't think it's going to be overnight kind of thing. Again, once the signal goes to optical in the scale up domain, it makes sense to process in the scale up domain.
That's right.
Right?
Why make the OE convert, right?
Why go back, right?
Right.
I think that's a natural kind of migration for the world to go in that direction.
Got it. No, that's a very good point. Last question on OCS. Do you think there is still industry debate on the specific technology, right? MEMS versus liquid crystal. I imagine you have your own perspective, but do you think that debate has been sorted out, both technologies work well, or you think MEMS has a better shot at being more successful?
No, I'm highly biased, right?
Yes.
Let me just have to say that, right? I'm highly biased. If I look at just the pool outside, look at technology, right? The way I look at it is that the MEMS is an engine you build in parallel, you build at the wafer level. You build the engine and you put optics around it, you put a fiber alignment to it, and you're done, right? The liquid crystal is more of a serial approach. You have those layers and layers and layers of a liquid crystal, right? It becomes more a.
Right
manufacturing process, thereby you introduce a little more insertion loss and so on and so forth. Right? I think manufacturing process itself is different and resulting in potentially different performance optically or potentially different manufacturability, right? We don't have liquid crystals, so we don't really know. Third piece really that some people are using is the piezoelectric, right? That's probably even more serialized, right?
Right.
Therefore, our view is that we know MEMS the best. We know it meets all the specification. We know we can manufacture it, right? We know we're ramping it. I think jury is out, right? I think if the market is big enough, it needs more than one supplier that can do this. Right? Again, if the whole market goes to the inference-based, optical, scaled up domain, big enough, the market will need multiple technologies to support it, right? I think we know our technology is pretty well suited for the application, and we can actually scale it.
Got it.
Yeah.
Next topic, Wupen Yuen, I would like to move to is optical transceivers, right? pluggables.
Right.
Lumentum, I find, has been a little reluctant initially to invest, right, too much, right, into the business. You got Cloud Light, and then you got an important customer. How strategic of a market is optical transceivers for Lumentum? Is it a lot more strategic or is it still more tactical that you have one or two customers, you will do what they need, but you don't want to make this a big business?
Yeah. I think optical transceiver is an interesting piece, right? First of all, there's a lot of questions in the conference saying, "Hey, when CPO happens, what going to happen to your module business?" Luckily, I would say that our primary concentration of the module business was customer who don't like to use CPO.
Right.
There's no cannibalization, right? We're kind of enjoying the benefit of module ramping along with OCS ramping, along with the CPO ramping. We're kind of lucky in that sense. The module piece is pretty important to us because through modules, we actually get to see a lot of the roadmap piece from our customers, right? Whether it's high density like XPO or next generation architectures where things are moving, right? Then there's a lot of evolution happening now, technology space around the modules. I would say that's strategic in that way, but we do admit, right, the margin is definitely below our corporate gross margin, right? Our goal is say, let's serve our current customer well, right? Let's not go crazy to expand our customer base, but they just focus on serving our couple of hyperscaler customers really well.
Right along with their roadmap, enabling them, and continue to develop the technology. I would say it's strategic, but it is kind of a confined strategic product line for us.
Got it. Is there a scenario where, because of various geopolitical reasons, U.S. hyperscalers decide to buy more design in the U.S.? Everyone is kind of making it in Asia, so I don't know whether that makes a difference.
Right.
At least packaging it there. Do you think there's a scenario where you get some advantage and that might make transceiver a bigger business?
That's a really good question. I think that there is a general concern and hesitation on the hyperscalers of the dependence on the China-based module vendors, even though they're all moving the production outside, right? Different hyperscalers have a different level of concerns. Some are very concerned, some are less concerned, right? Very hard to say what they're going to do, right? Frankly speaking, the China-based vendors are doing a really good job.
Right.
Serving the needs of the hyperscalers. Without them, probably our AI, it's not going to be totally that well either. I think it's a complicated question, but we do see concerns. We do see concerns about the IPs, the manufacturing location, and where are things coming from, and so on and so forth. I cannot predict the future, but I think the concerns do exist. Yeah.
Have customers spoken to you? Let's say if they wanted to, for whatever reason, do you think you have the desire, the capability, the capacity to actually become a bigger transceiver vendor?
Well, we don't have the capacity, for sure.
Okay. Well, that makes good-
Given time, anything happen, right? Given two years, can you build capacity? Yeah, you can build it in two years. It depends on what is the commercial condition, right? Would that be worth our capital allocation to make that a much, much bigger business?
Right.
Technology, we have it. We have some of the key components, but whether we do that or not, depending on the commercial conditions, I think by that time. Yeah.
Makes sense. Then finally, on Long-Term Agreements. It seems like this year everything is in short supply, right? Customers, I imagine, are willing to sign up for whatever. How durable, how enforceable, right? How are these, obviously not proprietary agreements, but how in general are you structuring them? Are you holding customers to a certain unit commitment? Are you holding them to a certain pricing commitment, percentage of volume commitment? What is the general nature of these LTAs?
Yeah. General nature of that is basically a take or pay kind of thing. Right?
It is take or pay.
Take or pay. There's a very strong commitment from the customer side, and in some cases, there's a prepay element as well. It's very much structured like that. Again, the duration depends. Typically, two or three years or some are actually longer. We're also thinking strategic, like we don't want to max our capacity with LTAs because we still do believe that the AI optics usage will continue to rise. Therefore, we don't want to just lock up all capacity we have.
Right.
We want to preserve some capacity for engaging more customers or opportunistically spot price of the market and so on and so forth. That's kind of our overall strategy, but by and large, it's a take or pay kind of architecture of the framework of the LTAs.
Got it. How long do these run till, some of these LTAs?
It's three years, five years.
Three years to five.
Two to three or even five years in some cases.
Got it.
Yeah.
If, let's say theoretically, cloud CapEx slowed for whatever reason in the next two years-
Right
you feel that these are enforceable?
We feel it's enforceable, are we going to sue a hyperscaler? Who knows, right? I think that these are important relationships.
Right.
I don't believe AI's going to crash. It may slow down, but you're going to grow again. The reason we're building with hyperscalers are truly strategic, and therefore, I hope that they view it that way, too.
Nice. Of course. Makes sense.
We don't see that being a transactional relationship.
Got it. Okay. Last question. NVIDIA made an investment in Lumentum. From what has been said publicly, what is the nature of that relationship? What kind of assurances, what kind of engagements do you have because of that relationship? What does Lumentum need to do? What is your part of the relationship?
Yeah, that's a great question. Basically, there are two pieces. The piece number one is the purchase agreement contract. That's a multi-year LTA for the ultra-high power laser. That is actually separate from the investment piece. Investment piece is completely independent from the purchase agreement. Investment really signifies the strategic relationship between the two companies.
Right.
That's an important thing because both parties will say, "Okay, we are strategic partners," but there's no string attached to what we can use that funding for. They just become a shareholder of Lumentum. The ultra-high power laser for CPO, that's something we're ramping to. There's a commitment on both sides. There's a volume, there's pricing, there's take or pay.
Will you be exclusive, you think, in that opportunity?
No, I don't think that's a exclusive relationship.
Okay.
I think also because we don't want to be exclusive.
Right.
They don't want to be exclusive. There's going to be at least two companies sharing that share.
Right.
Which I think is good for both companies.
Right.
Yeah.
Makes sense. With that, Wupen, thank you so much. Really appreciate the time.
Thank you very much.
Okay.
Thanks for the opportunity.