Hi, good morning, everyone. Thanks for joining us. Thanks for joining the Rosenblatt fourth annual Age of AI Conference. I'm Mike Genovese. I'm the Cloud and Communications Equipment Analyst here at Rosenblatt, and this is the group fireside session for AAOI, an optical transceiver company, and a cable component and amplifier company. I'm very pleased to be joined by the Chief Financial Officer, Chief Strategy Officer, Stefan Murry. Nice to see you, Stefan.
Thank you very much, Mike. I appreciate your invitation to be here this morning. It's great to have the opportunity to present our company to some new investors.
Perfect. So this is just gonna be a fireside chat. We're not gonna do slides. We're gonna just have a conversation here, and I welcome questions from the audience. You should all have a widget on your viewing screen there, where questions that you type in will come directly to me, and I will certainly ask questions. So we already have questions coming in, which is great. So I'll get back to those in a moment, you know. 'Cause I wanna. Let's start with just some bigger, you know, bigger picture questions.
I guess, you know, we see the 800G transceiver market taking off, you know, Fabrinet reported $259 million in 800G revenue last night. I think Coherent had about $250 million. Lumentum had some good news in the market. We know InnoLight is a player here. So I guess my first question is how many players do you think... I mean, you've been in this market a long time, but, so you should have good perspective on this, but, you know, how many players do you think, the AI, you know, 800G-plus transceiver market can support going forward here?
I mean, I think the market size for 800 G and 1.6 terabits, which will be the likely follow-on to 800 G, you know, is enormous. It's being driven by this shift from, you know, traditional compute workflows to generative AI. And a lot of the volume is around training the neural networks for this new AI. So they need a very high performance, you know, low latency, high speed data connection to be able to facilitate that.
And so, you know, to answer your question directly, I think, you know, the size of the market would indicate that, you know, it should be able to support, you know, four or five solid players with reasonable and, you know, relatively fast-growing revenue. I would say, at least over the next couple of years.
Okay. So in this kind of most recent cycle of news flow, you know, related to, you know, whether it's 400 or 800 G, you know, AI adjacent or direct AI product, the excitement really started with Microsoft several quarters ago for you guys. And, you know, there's been a lot of talk about that, I think, second technology agreement with Microsoft, and so far, you know, I guess, you know, it's fair to say the revenue ramp has been a little bit, you know, disappointing. It's certainly been slower than expected. So, you know, I'm just wondering, what do you see as the reasons for it being slower than expected?
And then what gives you the confidence that, you know, it's still that the size of the contract is still what it originally was, and that it's, you know, it hasn't gotten smaller because just because it's been slower?
Sure. Well, I can't go into the details of the reasons behind the slowdown in any particular way, because as you can imagine, Microsoft is, you know, we're under a non-disclosure agreement with them, and so we can't talk about the particulars of their network evolution. What I can say is, as I've said for the last year or so, that this represents the evolution of their next generation architecture, which is designed to more efficiently support generative AI. So, that has not changed. It's still very much their network architecture for the AI future, at least, you know, again, for the next several years.
The fact that it got off to a slower start, again, without going into too many particulars, I think is not uncommon when you have a new technology and a new architecture. This is not a simple drop-in, you know, component replacement. It's an entirely new architecture, that they're deploying. And as such, as you can imagine, it requires not just one supplier to have, you know, the appropriate products fully qualified and ready to go, but all of the suppliers that are necessary to build out that network. And they have to have enough operational experience, in the lab and in field trials to be able to confidently deploy that new architecture. And so, you know, I liken it to, if you've ever done any sort of home construction, right?
I mean, it's a big project, it's got a lot of moving parts, and, if you've ever experienced it, they don't always go according to plan, right? And that's not necessarily because the contractor doesn't wanna get it done, or the suppliers don't wanna get it done, and the homeowner certainly doesn't wanna get it, you know... Everybody wants to get it done on time, but there are steps that have to be followed and things that have to be done in an appropriate order. And, you know, as with a lot of things in life, those don't always go according to the most optimal schedule. You asked about, you know, the size of the opportunity, and I would say, again, this is the architecture that they are rolling out for their future of AI.
I don't believe, based on the conversations that I've had with all of our large hyperscale customers, and in particular with Microsoft, and certainly what I know of the industry and the general backdrop behind this shift to generative AI, I don't believe that the emphasis that our customers are placing on evolving their networks to better support generative AI, I don't think that that has diminished. If anything, I think it's grown, and so, you know, that would also point me to the belief that, you know, this program is certainly not diminished in its importance.
Great, I like that answer. So let's move on to another big, before we get back to transceivers, you know, cable is, you know, potentially here, or seemingly, I mean, a major driver. You know, through the first half of the year, the numbers were low, right? We hadn't really seen the transition yet to DOCSIS 4.0. So now we're kind of, you know, in the middle-ish of the, well, still the third quarter, the second half of the year. You know, starting in the second half of the year. Are you seeing, you know, are we seeing the projects?
Are they happening now or are we still waiting for them, or do we, you know, do we have more visibility that they're actually, the orders are coming in at this point?
Yeah, the projects are starting. You know, the technology is being. We've talked about on the call that, you know, for us, there's initially four different products that are cable TV amplifier-type products that are used to support higher frequencies in the outside plant, which are required to deliver higher bandwidth services. In particular, by the way, the bandwidth services that are being targeted are upstream bandwidth. That is, the bandwidth that you know goes from the home up to the cable TV operator into their internet connection there. So it's not...
And I, I wanna kind of emphasize this, particularly for some of your investors who may be a little newer to the story, the cable TV play here is not about, you know, enabling more video channels or anything like that. It's purely around better aligning their bandwidth delivery needs for internet connections with the demands that we're seeing. And this became particularly acute during the pandemic when, you know, if you think back, almost difficult to think back to pre-pandemic times, but if you remember, you know, bandwidth was highly asymmetrical. There wasn't - we weren't doing the kind of video conferencing and working from home and, you know, being highly interactive in to the extent, anywhere near to the extent that we are doing it now.
And as a result, that placed great demands on the cable TV operators' networks, in particular, that upstream bandwidth, right? We can get a lot of bandwidth in the downstream direction, but not necessarily in the upstream direction. So the technology that AOI has developed, which is encapsulated in the DOCSIS 4.0 standard, is designed specifically to improve cable operators' ability to deliver bandwidth in the upstream direction, while also not diminishing their bandwidth offerings. In fact, it's increasing the bandwidth offerings in the downstream direction as well, but the capacity expansion in the upstream is what they're really targeting. That's why the operators want to expand.
I think they recognize, you know, the limitations of their network, and they now, I mean, really just this quarter, basically, have products from vendors that are able to support these new standards that would enable them to, you know, increase their service offerings. So I give you that background to try to illustrate that, you know, there is a reason why the carriers want to, wanna do this transition. It's not simply, you know, just an idle thing that they're thinking about. It's something they've been planning, along with us and other vendors, for many, many years.
Charter, depending on how you look at it, the largest or second largest MSO in terms of broadband subscribers, announced just about a year ago a major capital expenditure associated with upgrading their network to this new DOCSIS 4.0 standard. That work has you know begun but really slowly, and we anticipate that, you know, they and other MSOs will you know will be rolling out these new technologies you know very very soon. So to put that more clearly, we said on the call that, you know, we expected cable TV to grow sequentially from Q2 to Q3. We do not expect that to be coming from older you know prior generation DOCSIS 3.1 products. It's gonna be...
The growth is gonna be fueled by, you know, the first deliveries of DOCSIS 4.0 products, which are, you know, being produced by us right now, and we'll be shipping this quarter.
Okay, great. Well, look, I mean, I'm looking at the presenter dashboard here, and this is a very well-attended session, and there's a bunch of questions coming in. So, I'm gonna put my questions aside for a moment and just go through some of these audience questions, and I'm just gonna ask them in the order that they've come in. So some are, you know, a little bit more strategic, some are a little bit more tactical, you know, or factual. So for instance, the first question, which is, "Does AOI produce four-inch wafers?
The answer is yes, we do. It depends. I mean, we can produce a variety of different wafer sizes, starting from two inch to four inch. Historically, most of our volume has been on two inch and three inch, but we do have the capability to produce four inch as well, and for certain types of devices, we do produce four inch.
Okay, and now next question: "In the past few days, Coherent, Lumentum, and Fabrinet all talked about strong demand and capacity constraints for the next few years. Does AOI see the same thing?
Yes. Yeah, I mean, I wouldn't say we are personally capacity constrained at this point. As we talked about on the earnings call, we've been adding capacity in anticipation of this ramp, and so, you know, we think we're doing a good job of resourcing the ramp that we see, but as an industry, I do believe that we're gonna see some capacity constraints in the ability of the industry as a whole to deliver enough 800G to satisfy the demands coming from AI.
I mean, that theoretically should help you, right? If you have capacity and other people are out of capacity, then you-
Yeah, I believe it will be an asset to us, and that's, you know, one of the big reasons why we've been investing in adding that capacity. Yeah. You know, for us, I'll say a few words about, you know, adding capacity. One of the things that we think is very different about our production process for the transceiver, specifically for the 400G and 800G and, you know, eventually the 1.6 terabit products, is that we've specifically engineered those products to be amenable to production in a highly automated fashion. That is, we have, you know, relative to our industry peers, and we believe we have a much smaller labor content involved in the manufacturing, because we've spent a long time developing automated techniques and a platform for the module itself that's amenable to being produced in an automated way.
And that gives us a couple of advantages. One, when we talk about adding capacity, you know, the lead time for us to add new equipment and, you know, add machinery to our production process is typically less than the lead time it would take to hire and train the types of skilled operators that are needed to do the manual processes, that are used by most of our competitors.
In addition to that, typically when you do ramp up, in a manual type of process, and we've seen this in our history, it's one of the big reasons why we went to an automated manufacturing process is because we noticed that whenever we try to ramp up, you know, the, the initial quality and reliability metrics, it took a while for the new operators to get skilled enough to produce the products that we, that we felt comfortable, you know, producing the products and, and shipping them to customers, and not having to worry about, you know, quality or reliability issues.
So, so not only is there a time associated with the ramp, but there's also, you know, a risk associated with, you know, rapidly increasing production, which doesn't exist with, you know, an automated process. Obviously, a robot, you know, putting a second robot next to the first robot, as long as they're running the same program and the same, you know, configured in the same fashion, they're gonna do the same job from day one. The other advantage that it gives us is the ability to do manufacturing virtually anywhere in the world that we would like, but, and more importantly, that our customers would like. It is not the case that we're relying specifically on a low-cost labor force to be able to be, you know, competitive in the market.
Because we have this automated process, we can, you know, put production lines, again, where we would like. We're hearing from a lot of our customers that they would like production to be done, at least outside of China, if not, you know, many of them seem to have a very strong preference for production in North America, and even in the United States in some cases. Our ability to enable that with our automated production capacity without dramatically increasing the cost associated with the production is something that we think our customers find very attractive.
You know, I mean, I've been hearing recently that the hyperscalers are pretty happy with Thailand production. And I'm kind of wondering why, you know, I haven't personally been hearing more of the preference for the U.S. Is that any thoughts on that?
I guess I don't know who you've been talking to, so I can't answer why you haven't been hearing it. The guys that I'm talking to are very interested in it. I would say, you know, right now, what you're hearing is a lot of the companies that formerly did production in China, setting up an alternate or in some cases, replacing their production factory in China with one in Thailand. So you're hearing a lot about it because that's the only option that there is right now. There aren't viable options in the United States.
So, you know, you're not gonna hear a bunch about that from the hyperscale operators because you know, absent AOI in our automated manufacturing capacity, I'm not sure that there is much capacity or much plans by our competitors to build capacity here in the U.S. It's not surprising you wouldn't hear much about it. Nevertheless, we think it's a very attractive proposition for you know, for the data center operators. And it kind of makes sense too, if you think about it. I mean, you know, there's been a lot written in the popular press you know, just around sort of the angst about AI, right?
There's a lot of security concerns, there's a lot of, you know, concerns about, you know, what the future would look like as AI begins to become more and more prominent. I think layering on to that, some of the concerns around security, for Chinese, and, you know, Chinese-manufactured IT infrastructure, et cetera. You know, it makes sense to me that the data center operators would be, you know, would have a preference for a product that's manufactured, at least, you know, as close as possible to the United States, if not in the United States.
So kind of, you know, in the broad sense, it makes sense, and when you actually talk to the customers, they're saying, "Yeah, that would be, that would be really nice." But they're not seeing that from the ability to do that or the willingness to do that from many of our competitors, which is probably why you're not hearing that much about it.
... Right. Well, you know, I don't think I'm gonna get a chance to ask any of my questions for the remainder of this, 'cause they're coming in so fast and furiously. I've got at least 10 here from the audience. And again, I'm just going in order that they came in-
Right
... rather than grouping them. So, on the one Q earnings call, the notion of an $800 million revenue year in calendar 2025 was presented as doable with large 800G input. Is this still a possibility or on the table?
Yeah, absolutely. You know, we don't want investors to take away from that this is somehow a slam dunk, right? That it's you know, all in the bag and ready to go. What we're trying to say is that and this is kind of how we opened the call, right? You know, you had asked, you know, "Is there room for multiple players to be very successful in ramping at 800G?" and the answer I gave you was, "Yeah, I think there's room for four or five, maybe even more viable players at scale," which would imply that those players are gonna have revenues, you know, of $500 million or more just in 800G, you know, for AI annually, and so I think that's very much in play.
Now there's a lot that we have to do, and the industry has to do, as I mentioned from the very beginning. You know, as an industry, I'm not sure we're there yet, but certainly AOI has some work to do, and our competitors have some work to do, but if we can do the job that we've set out for ourselves, that's very much in play, so you layer in some revenue from 800G with our existing 400G revenue, which also is gonna grow because that's adjacent to AI, as you mentioned, and then you know some strong growth in our cable TV, which we're expecting coming from the DOCSIS 4.0 transition.
You know, you can very quickly see numbers that are getting up into the, you know, upper hundreds of millions, closer to a billion in revenue.
Okay. Great. Next question. Following up on the Microsoft comments regarding building a house, can you level set the latest timing in general terms, or is this still unknown?
No, I mean, we talked about it on the earnings call, so, you know, we're seeing orders now. Those orders are growing, and we are told that those orders will expand further in Q3. That's what we're expecting. We have various, you know, supply agreements and things like that with Microsoft that would indicate that they're committed to, you know, expanding that business with us. So can't go into a lot of details on those, but they're fairly recent. You know, we're seeing a lot of recent activity, new orders, ramping. Pretty much what we would expect to see given the timeframe that I outlined on the call, which is Q3, beginning of the ramp, basically.
Okay. Great. Moving along here. Can you update us on the datacom component supply chain? Are you seeing an improving supply components such as 100G SerDes, PAM4 DSPs? Let's start with that, and then,
Yeah. So, well, we haven't had, you know, a tremendous problem with supply shortages so far, certainly not at 100G. You know, there's always some spot shortages here and there. Anybody who's done manufacturing, you know, you know that there are some parts every now and again, but nothing that we would consider to be, you know, would hamper our ability to deliver on our revenue projections.
Okay. Well, there's a couple of questions here about, you know, confidence and timing in winning 800G orders. Somebody else put it, "Can you update on the 800G pipeline, where you are on the timing of your prospective awards and deployments?
Right, so you know, we also talked about that on the call, and I'll kind of echo those comments. You know, we expect to be delivering 800G beginning in Q4, maybe as early as the end of Q3. You know, that's coming up pretty quickly, so we'll see, but certainly in Q4, we expect to be delivering 800G. I don't expect it to be, you know, really large volumes. I think I characterized it on the call in the Q&A session as, you know, kind of single-digit millions type of revenue, but we expect it to grow quickly into 2025, and again, you know, we talked about what the trajectory could look like throughout 2025 in terms of aggregate purchasing.
Would your first 800G orders, you know, later this year, be with the you know customers that are now your 400G customers? Is that? Or would they be new customers?
It's most likely gonna be customers that are 400G customers already. Yeah.
Okay. Like, like, for example, when you won the new 400G customer last quarter, right? I think the news was: you had one grow into a 10% customer, and then you had a new 400G customer. Like, what? Did that award for 400G contemplate that in the future, they would be buying 800G? Is, was that kind of part of the deal, or not really?
I would say for at least one of those two customers that you mentioned, 800G was the primary driver. They have a strong interest in 800G, and I think we made a reasonably convincing case that, you know, the... They also have, you know, a strong demand at 400G, and I think we made the case that, you know, giving us some of that 400 G business to get us into the pipeline and, you know, kind of get us delivering and get all the initial, you know, bring up of a new supplier. You know, get those hoops, you know, get us through those hoops at 400 G, get us into the pipeline, so that when we transition to 800 G with them, it'll be a smoother transition.
And I think that, you know, they've largely agreed with that philosophy, so that's where we're, you know, that's where we're going. So to summarize, you know, 400G is a stepping stone to 800G, with a lot of these customers. And that's not to say that 400G is going away, it's just 400G is what we have today, and then we'll add 400-800G, you know, later this year and into next year.
Mm-hmm. For the one that became the 10% customer, that was kind of a long time lower than that, you know, maybe 4% customer turned into a 12% customer, is that higher level of revenue, you know, sustainable? Are they expected to remain, you know, about 10% of revenues or should it be, or could it be, you know, very lumpy?
I would expect it to be pretty sustained, and you know, obviously, when you mention 10% of revenues, as our revenues grow, that would imply that this company would grow as well, which I think is reasonable to expect. You know, if we get to the levels that we're talking about for next year, would they still be a 10% customer? Maybe, but certainly there'd be some dollar growth, I think, in that revenue stream, and I don't expect a diminution in the revenue.
Okay. So I'm gonna get better here at grouping some of these questions, because we have a supply... You know, questions on, whether the, you know, 800 G, supply could be an issue and could hinder, growth if there's any components or bottlenecks that would limit, you know, growth at 800 G. Somebody else says, "Is EML supply becoming an issue? Could this be a headwind to your transceiver ramp?
Right. So, I'll just say in general terms, it's a little early to say what the supply situation is gonna look like at 800G as we start to ramp. We have, as I mentioned earlier, you know, we haven't seen any shortages. We're not, you know, anticipating any significant shortages at this point. We are communicating very actively with our vendors, that, you know, what our expectations are in terms of the ramp and what we see, you know, coming. They've indicated that they do have an ability to support that, and at this point, I have no evidence that would suggest that they're not being truthful with those statements. So, you know, time will tell. We do have multiple sources for most of the key components.
The other thing that I would say, particularly on the 800G, is we're working very closely with our end customers, our hyperscale customers, and, you know, they're helping us with the supply chain partners as well by putting, you know, kind of their, you know, mark of approval, if you will, on the arrangements that we're making with these, you know, key suppliers. And I think that's helping give the suppliers confidence in building the capacity that they need as well. It's one thing to hear it from AOI, it's another thing to hear it from, you know, a large hyperscale operator on behalf of AOI.
So I think the hyperscalers have recognized that there's a potential for shortages, and they're doing everything they can to help push the industry to ensure that those shortages don't come about. The second question had to do specifically with EMLs. That is an electro-absorption modulated laser, by the way, for those who may not be familiar with the arcane acronyms that we use in this industry. So, the EML is a relatively advanced type of semiconductor laser. We actually make our own EMLs in-house. We also do source those EMLs as... I get this question a lot, you know, "Do you source externally?" And the answer is yes. I mean, we buy product from other vendors, and we also oftentimes make it in-house.
The reason for that is, well, there could be two reasons. Number one, in some cases, you know, a particular vendor may have a better price or a better performing product for whatever reason, and so we'll buy it for that reason. More often than that, it's that the customer wants or requires, in many cases, us to have two suppliers. Even if one of them is our in-house supply, they still want a backup, second source. And in any case, it's prudent to have that, for us. So I think one of the aspects, you know, in addition to the automated production capacity, which I talked about earlier, one of the other aspects that customers like about AOI is that we do have a fairly differentiated supply chain.
If you can imagine, you know, if you have four different suppliers of transceivers, but they're all buying key components from the same company, EMLs, for example, or, you know, DSPs or whatever, if those are all coming from the same company, then you still have one company that represents a significant risk to your delivery. With AOI, because of our internal manufacturing capacity, you know, we have a more diversified supply chain than many of our competitors, and that's also attractive to customers. Now, that's not to say that we can't, you know, experience key component shortages. There are certain components that we buy externally, that, you know, we only have a couple possible sources. And so, you know, we're watching those very carefully.
But generally speaking, our ability to manufacture stuff in-house gives us some differential advantage when it comes to these supply chain shortages.
Great. We've got a question here on the gross margin. So, you know, confidence in a rebounding after the, you know, the blip, the last, I guess, couple of quarters. And I guess I would add to that, you know, what do you see as the timing for getting back to at least the high 30s%, if not, if not 40%? You know, would that, could that be mid-2025 or, or, you know, is it, is it end of 2025, early 2026? What do you think?
Yeah. So, I guess, first of all, on my confidence on the gross margin improvement, I mean, I do have a fair amount of confidence in it. The thing to understand about what will drive that gross margin improvement, and then you'll understand the basis for my confidence. So, the first thing is, as we talked about from the very beginning, our cable TV business has been really, really low, historically. I mean, we have had, at its peak, which, by the way, wasn't even in a strong upgrade cycle, we were doing $35 million in a quarter in cable a few years ago. And that was on, you know, again, not in an upgrade cycle, it was on an older generation product.
So, as we roll in these newer generation products, and as they contribute a greater amount to revenue, we would expect margins. You know, so the cable segment for us is meaningfully higher margin than the data center business. So by layering in a greater contribution from cable, that will help us improve the gross margins. And as I mentioned earlier, those rollouts are happening. MSOs are beginning to roll out the new DOCSIS 4.0 technology, and so as that happens, that will help us improve gross margins. The other thing is the 400G business that we're seeing coming in from some of these new opportunities that we've talked about is largely single-mode lasers. Those are typically used over longer distance.
And the fact that both the 400G and 800G we talked about. By the way, I mean, again, I probably should mention this for those of you who may not be technical experts, but the 800G, the EML that was asked about earlier, this electro-absorption modulated laser, is a single- mode laser. So when we talk about single- mode in the context of 800G, we're generally talking about an EML-type laser. So the new opportunities that we're seeing, both at 800G and 400G, are largely around those single- mode type devices, and those generally are higher margin as well. So that will help improve the blended margin on the data center business.
So you see, you know, data center margins improving by a greater predominance of, single- mode versus multi-mode, and at the same time, later on, more cable TV revenue at a higher margin there. Both of those things will be additive to gross margin. So as far as your point about, you know, when do we get back to, you know, margins in the, mid-thirties or upper thirties, and we haven't given a specific timeframe on that. It's very difficult to predict. I mean, as we talked about earlier, we got started on the Microsoft thing, and the timing of things is always the most difficult thing for us to predict. We know the general direction that things are going, but timing is a little bit hard to predict at this point.
I think we'll see meaningful improvements sequentially after Q3. As you can see in the guide on Q3, the margin's not, you know, dramatically improving. It's hopefully gonna tick up a little bit from Q2. But we do see a stronger growth in Q4 as the revenue contribution from cable, you know, comes in at a higher margin, and also some of these single- mode data center products. Then beyond that, I think we'll see some incremental improvement throughout twenty twenty-five and certainly into the thirties. Where exactly that levels out for the year, we'll see.
Okay, great. I'm gonna combine four questions here, so I'm gonna ask them, all four of them, and-
Wow!
You'll get the theme here.
Should I be writing this down? Do I need a pencil and such?
Yeah, I-
I believe him.
I think you'll get the theme here, which is, you know, basically, what revenue level does the current capacity support? What is added capacity going to support in terms of revenue? But then, you know, the other side of that, the balance sheet, cash needed for CapEx if you win new business. Given your cash constraints, how you manage automated line capacity expansion? Will your customers pay for the capacity expansion? And then finally, how much capital do you need to raise to support the 800G capacity? This is all about, you know, kind of the current capacity you have, the capacity you're planning to add, and where you're going to get that cash.
Right. Okay. So, broadly speaking, you know, we've said before that we, you know, we have about $120 million a quarter production capacity today. Now, that makes some assumptions about the product blend and all that kind of thing, so you have to do some you know, a little bit of modeling there to get to that number, but that's basically what we can produce. Today, we're adding capacity, as I mentioned, so we'll close Q4 with a higher capacity than that. We haven't indicated exactly what that capacity is, but you know, there'll be some expansion from there. And then moving into twenty twenty-five, a lot depends on, you know, kind of what we're hearing from our customers about their forecasts.
I mean, we don't have the intention of building a lot of additional capacity until we have a clear commitment from the customers that they're gonna purchase the products that would utilize that capacity. So the clearest way for us to get that commitment would be to get some sort of, you know, a guaranteed revenue or an investment associated with the CapEx. That question was asked as well, and that's something that we're certainly discussing with various customers about their, you know, the nature of the commitment that they're willing to make, that would enable us to feel comfortable making those CapEx commitments, and that's very much a detailed conversation that we're having with customers today.
And so the answer to, you know, how much capacity do you end in 2024, and how do we expand in 2025? Really, kind of, is not certain at this point. In aggregate, to get to the levels where we think we need to be based on what customers are telling us, would require an incremental CapEx of about somewhere between $40 million and $60 million. And again, how much of that we would pay for directly versus some manner of support from customers, that's an open question as well.
Okay. So that's that kind of high hundreds of millions, approaching a billion type of investment needed to get there, roughly?
Yeah. Yeah, I mean, it would support that, that type of number, yeah.
Yeah. Okay, we've got more questions from the audience, and we're not done yet.
Awesome.
Given what you said about EMLs, do you see the opportunity to sell 100G EMLs to other transceiver manufacturers? You know, do you see a laser component or other component business, or are you gonna stick to transceivers?
I mean, we do have some business in selling laser components. It's been a part of our business for well, really ever since the company started. In fact, we were founded as a laser component company, and we expanded later on into the module business, including transceivers. So it's certainly something we have done and will do. Right now, it seems like the capacity that we have available for EMLs and other types of laser devices that are used in data center. It seems like we will utilize all of, or virtually all of that capacity for our own internal use. And we're not likely to add capacity specifically for you know, servicing you know, laser-only type customers. So I would say, you know, it's a possibility.
It's certainly something we've talked about, and we've done it in the past, but it doesn't seem likely that that's gonna be a big revenue stream for us, certainly in the next year or so.
Okay. Will micro-scale cloud service providers like CoreWeave and Lambda, as a category, adopt AOI as well as hyperscalers? Are there more new customers out there, potential customers out there?
Yeah, that's a good question. You know, the answer is likely yes, there are probably new AI customers. I mean, certainly, there's new AI customers. What we've seen typically, and the reason why we talk mostly about the hyperscale customers, is the hyperscale customers tend to work directly with companies like AOI to design, specify, and procure key components for their data centers. Smaller operators tend to purchase, you know, their optics along with, for example, their switches and servers and the other network elements that go into a data center. So we wouldn't necessarily see that business directly.
We do have business with various network equipment manufacturers that make switches and servers and other things, and they will buy our optics and package it with their components and sell it to some of these smaller operators. So the answer to the question is, yes, I think it's likely that some of the smaller operators will adopt AOI transceivers, but it likely will not be sold directly from AOI. It'll be through one of these other channels.
Great. And here we go. Just a question about, can we expect $50 million a quarter at some point in cable TV revenue? Can that happen in 2025?
Yes, I think it's reasonable to expect. As I mentioned earlier, we in relatively recent history, a couple of years ago, we had sales of $35 million or so in the older, you know, DOCSIS 3.1 technology. It's basically the same type of products, very similar type of products to what we're selling for DOCSIS 4.0, but just the older generation of those. So, the newer generation of these amplifiers are what are called smart amplifiers, so they have some intelligence built in there. Part of that is to enable AI to be used to better predict network outages by the cable operators, so that, you know, we can do more proactive maintenance. So there's some additional functionality, some smarts embedded in the amplifiers that was not present in the older generation products.
and that implies a higher price point for the DOCSIS 4.0 products, and we're seeing prices that are, you know, 20-25% higher. So if you think about it, you know, if we were able to produce 35 million of the older generation product, and assuming we can produce at the same rate, which is a reasonable assumption, you know, that would imply at this, you know, at this higher price point, somewhere close to $50 million a quarter, you know, in revenue. And that's without any, you know, capacity expansion. And, you know, we believe, based on what we're hearing from the MSOs about the, the nature and extent and duration of their network upgrades, that there probably will be a need at the peak for, more than $50 million of deliveries per quarter.
Will we see that by the end of twenty twenty-five? You know, that's a really... Again, I keep coming back to this, how, you know, how difficult it is to predict the timing. I think it's likely that we'll see it by the end of twenty twenty-five. But I would also note that, you know, there's a lot of work that the MSOs have to do to be able to deploy that scale of infrastructure every quarter. I mean, this is not like a data center where, you know, it's kind of all contained in a you know, a building or a couple of closely adjacent buildings. This is infrastructure that's spread out across a geographic area in the outside plant of these companies' networks. And so it's not a simple matter to deploy some of this equipment.
You got to get, you know, bucket trucks and trained technicians and all that. And I think as an industry, you know, there's work that needs to be done to be able to have enough trained technicians and equipment to be able to do installations at a rate that would imply $50 million-plus a quarter of business for AOI. But I think they could get there. Will it happen in twenty twenty-five? We'll see.
Okay. Well, I really do wanna thank the audience for doing my job for me, and just supplying me with a tremendous amount of questions. We just have a couple minutes left, and so I'll ask a question. It's kinda, you know, just a general high-level industry kinda shift. I mean, are we seeing a shift where the early days of 800 G was multi-mode, and then, you know, now the demand is kinda shifting to single- mode and longer distances 'cause there's more switches getting involved? And, you know, earlier it was GPUs, and now it's switches. Is that a fair way to look at what's going on?
I wouldn't necessarily say it's exactly that way, and I have to be careful here because, again, you know, we have certain knowledge, you know, that's covered under non-disclosure agreements by our customers. So you're correct in that the early days of 800 G were largely dominated by multi-mode, and we see 800 G single- mode becoming more and more important. Let's just say that the reason for that has to do with the difference between how a hyperscaler would optimally architect an 800 G network versus how a smaller scale, AI network might be architected, okay? So it's not necessarily about more switches or less switches or whatever, less GPUs. I mean, certainly, the GPUs are still absolutely critical.
But it has to do with the difference between how a really large, massive AI deployment would look versus a smaller deployment, and that's about as far as I can go.
So the idea that the deployments are getting bigger and they're getting more massive, is that...
Absolutely, yeah. I mean, well, it's not so much that they're... Yes, they are getting bigger, clearly, but more importantly, the larger hyperscalers have figured out, you know, they have a better sense of what their optimal network configuration would look like, and they're rolling out those architectures to support that. If you go way back to what I said earlier about, you know, the Microsoft and the delay in the ramp, and I mentioned that this is an architecture shift. It's not just, you know, kind of a replacement part or something like that. That's what I'm talking about. Like, the 400 G and the 800 G in the future all goes together. It's not that they can't make an AI network.
Clearly, they can, 'cause they've been deploying those services, but there's a more efficient and better scalable way for them to do that, and they need some new components from the optics industry. 800 gig single- mode is one of those, and the 400 gig optics that we've been talking about is part of that as well.
Okay, perfect. Well, thanks so much. Thanks, everybody, for participating. Thank you, Stefan, for that conversation. Really enjoyed it.
I appreciate the opportunity to talk and have a great day.
Talk to you soon. Bye.