Hey, everybody joining us here at the Raymond James TMT and Consumer Conference. My name is Simon Leopold, Data Infrastructure Analyst. We've got a fireside this morning from Cisco. We've got Bill Gartner. So I've prepared a group of questions that we'll go through, informal chat of sorts. So maybe let's set the stage, Bill. Tell people a little bit about your background and your current responsibilities at Cisco.
First of all, thank you for having me. Let me just start by saying I'll be making some forward-looking statements that will probably depart from our actual results. All of the risks are outlined in our 10-Q and 10-K. Please refer to those for our normal disclosures. I started my career at Bell Labs and worked in the optical industry most of my career. I ran the optical business at Lucent. I launched the first DWDM system that happened to be 20 gig of total capacity on the fiber. You can imagine when that was. When I left Lucent, I joined a startup that built the industry's first ROADM, reconfigurable optical add-drop MUX. After that, I joined Cisco. I have responsibility at Cisco for really three different businesses.
One is the optical systems business, which is a traditional DWDM system that competes with guys like Ciena, Infinera, Nokia, Huawei, and others. I have responsibility for the optics business, which are the transceivers that we sell for routers and switches across all market segments, whether it's campus, enterprise, service provider, public sector, or hyperscale. That's a very multi-billion-dollar business for us. We sell everything from one gig to 800 gig optics in every flavor and form factor. And I have responsibility for Acacia. We acquired Acacia about three and a half years ago. It was a $4.5 billion acquisition, pretty substantial acquisition for Cisco. And we can talk more about why that's relevant for Cisco. But those are the three businesses I have responsibility for.
I'd say one aspect of the optical market is that it's constantly evolving. There's always new stuff coming. So what are sort of the innovations that are occurring in optics that you find most interesting and why?
Let me start with the innovation that really motivated us to acquire Acacia. That is that in the traditional optical system, the optical systems market has evolved pretty significantly. I mentioned that the first DWDM system that I worked on was 20 gig of total capacity. Now we put tens of terabits of capacity on a fiber pretty routinely. Just magnificent increase in capacity. In fact, now we have. I get to carry around some of my portfolio in my pocket. Now we have 400 gig coherent and a pluggable form factor. This is much, much more capacity than we had in the entire fiber in early stage DWDM systems. Not too long ago, the total capacity on a fiber was 400 gig. Now we put one wavelength of 400 gig and put it into a pluggable.
But the key shift that's taken place is that in the traditional optical system, we'd have transponders. And the job of a transponder, it's a line card that takes a wavelength coming from a router, takes an optical signal coming from a router that's designed to go maybe 500 meters or two kilometers. And it does two things. One is it gives it a color of light. So it gives it a wavelength. And then secondly, it basically produces a signal that can travel over a very long distance, hundreds or thousands of kilometers. That's the job of a transponder. And it sits in a line card that goes in a chassis. And the unique Acacia technology that motivated our acquisition was they took that line card and put it into a pluggable form factor. And this, I think, is really shifting the industry now.
Because when we first did that acquisition, we said this is absolutely going to have a role to play in DCI, Data Center Interconnect, but it's also going to, over time, play in metro networks and play in long-haul networks, and all of our competitors said, no, no, no, that's not going to happen, it's just going to be in this little niche, and lo and behold, here we are three years later, and this is being deployed in the top five hyperscalers for DCI applications, but it's also being deployed in over 200 service providers in metro applications, and we just announced a long-haul version of this as well, so I think the shift in the industry is going to be a result of packing this transponder capability into a pluggable form factor and then watching that technology evolution over time.
So that form factor, but more traffic, more speed over something that small.
Exactly.
So the basic, I'd say, driver for everything optics is traffic growth. So Cisco used to publish these traffic forecasts. I think you guys have stopped that. But in terms of your forecasting, your business, what kind of assumptions are you making about the dynamics of traffic growth?
I will say that all of our assumptions have been basically under, we've underestimated the traffic growth that we've seen over the last, I'd say, two to three quarters pretty dramatically. You're probably going to hear the AI terms used ridiculously by everybody that's talking. AI is driving that in the hyperscalers today. When we talk with some of the hyperscalers, they talk about building an AI WAN for inter-data center traffic that's being driven by their AI training. We're certainly seeing that growth inside the data center as well. Today, the traffic demand is being driven entirely by AI. That's what's really driving traffic growth. We've seen dramatic traffic growth in the hyperscaler space. We've shipped over 350,000 of these 400 gig units. Most of that has gone into hyperscalers.
But to contrast that, I would say the service providers have been pretty non-growth over the last couple of years. And had it not been for the hyperscalers, we'd be having a very different conversation. So all of the growth that we're seeing right now is coming out of hyperscalers.
Now, it's interesting because I hear this question about, is AI driving growth? And your statement was unambiguous. But if we think about it, the industry's really in a training phase. AI has barely been launched. So how do you think about what's kind of the five- and ten-year vision of what happens with traffic because of AI when we're all actually using it day to day?
Yeah. I think, first of all, I would say there's a lot of uncertainty in terms of how AI applications are going to evolve over time and where they're going to reside and what that means for traffic outside the AI infrastructure. There's clearly going to be a dramatic need for traffic within an AI pod, if you will, or an AI in a box. But how much of that has to leave the customer prem or how much of that has to leave the service provider remains a question. I think the question really is, how is the industry model going to evolve here? Like, are service providers going to be hosting AI on behalf of enterprises? Are enterprises going to host their own AI? That will determine a lot about what happens with traffic.
If every enterprise decides they're going to have a private AI solution, and by the way, Cisco is assuming that's going to happen as well, the traffic patterns are going to look very different than if the answer is hyperscalers and/or service providers are effectively creating a hosting service for AI. I think there will be both. But today, all of the traffic, as you said, Simon, is being driven by training by the hyperscalers. Very little that we're seeing coming from anywhere else. Very little growth.
You've also talked about this concept of Routed Optical Networking, which is a different approach on how to build the wide area network. Can you talk a little bit about what's made it possible and why you think it's the way to go?
Yeah. So Routed Optical Networking is a Cisco term that really takes advantage of a couple of insights about costs in the network and technology as well, this being one of the technologies. And Routed Optical Networking really is, for Cisco, really four things. One is it's a recognition that for the last 20 years, we've built networks the same way. We've basically built router-based networks and assumed that the router was the most expensive thing in the network. And when you assume the router is the most expensive thing in the network, that causes you to bypass the router whenever you can. You don't want to go through the router because it's too expensive a resource to go through. So that motivated an entire industry, an entire optical industry of building very smart reconfigurable switches, ROADMs, that would allow us to bypass routers, go around them.
So if you had to go A, B, C, D, E, and you were going from A to E, you would basically go from A to E directly on an express optical path and not go through B, C, and D. The reason was those routers were too expensive to go through. But now, if you think about what's happened with the cost of switch ports, router ports, and silicon, now we're delivering 25 terabit routers and 50 terabit routers. The cost per bit of those router ports has come down dramatically. And if you look at that compared with the cost of the optics, proportionately, the optics are a much bigger part of that cost now.
And so it really causes you to step back and say, does it make sense for us to effectively put expensive optics to go around these routers, maybe over very long distances, rather than just going through them? If I got a 25 terabit switch, I've probably got lots of excess capacity. It's cheaper in many cases for me to go through it. So that's one insight about how we think about building networks is that in some cases, rather than building very complex optical infrastructure to bypass routers, we should build a simpler infrastructure and just go through them. That's one insight. The other is that there are still private line services out there that customers, especially governments, but a lot of customers, procure: OTN services, DS1, DS3 services.
Those traditionally have been served with custom products, like a custom multiplexer or custom optical product that could effectively aggregate that traffic, but now we can actually emulate those private line services at the IP layer. We call that private line emulation, so you can effectively move that private line, what was TDM-like traffic, to the IP layer and do that in exactly the same way that it would have been done with an optical layer or an OTN layer. The third element is taking advantage of new technology, and in this case, it's pluggable technology, and saying this is a much cheaper from an OpEx and a CapEx perspective. This is a much cheaper way to deliver a coherent signal into the network than using a transponder sitting in an optical chassis.
The fourth is if we got all those things coming together, new network models, new traffic patterns that we think about for a network model, private line emulation, pluggable optics. We have to manage that for customers and automate a lot of that management. Automation becomes a much bigger part of our story. RON is really routed optical networking. Routed optical networking is really all those four things that have come together. If I could just say one more point on that. When we first announced routed optical networking, all of our competitors were like, no, this is a tiny niche. This is never going to play in the market beyond maybe a tiny, tiny niche of the market. Now Ciena, Nokia, Juniper have all announced things that sound a lot like routed optical networking.
It might be coherent routing, or it might be coherent optical routing, but it sounds a lot like Routed Optical Networking. I think they've acknowledged this is a train that's left the station. Cisco is going to continue investing to expand the market for these capabilities. It's not just a niche market, but it includes DCI. It includes metro and even long-haul.
Now, I think you've shared with me some metrics in the past that when somebody buys a switch, it's basically a bunch of empty sockets and you put the optics in. What were the sort of proportions of a fully equipped switch that were optics, let's say, five years ago versus today versus what it might be in two years?
Yeah. So I think a lot of it depends on which market we're serving. Because a service provider or an enterprise buys a switch or a router and maybe puts in one optic or two optics. And they grow over time. And part of the rationale for having this being a pluggable, whether it's a short-reach optic that's inside a data center or a campus environment, or whether it's a coherent optic that can travel over a very long distance, the argument for being a pluggable is that it's pay as you grow. You buy the switch with 32 ports, and then you add the pluggable optic over time as your capacity needs dictate. That's the way the service provider market and enterprise market grow. And historically, we run at about 30%-40% fill rate on those switches and routers. The hyperscalers have a very different model.
The hyperscalers basically buy the switch and fully equip it day one, whether it's short-reach inside the data center or whether it's coherent. They are populating that switch fully day one, so what we call the attach rate changes very dramatically whether we're selling into a traditional service provider or enterprise application versus a hyperscaler application, and part of the hyperscaler argument is we're going to have the traffic. We know we're going to have the traffic. We don't want to go touch this thing once it gets deployed. We want to install it and then basically just never have to touch it again.
So I want to touch on AI, of course, because we have to. So how do you think about the TAM for Cisco? What are your opportunities? What products are you selling? And I know some of this may go beyond the scope of your business, but how do you size that?
So let me just start by saying we're selling this pluggable 400G ZR, ZR+, the coherent pluggable into the top five hyperscalers today and into about over 200 service providers. The hyperscalers, I would say, are driven by AI. Service provider is more traditional traffic, looking more for an operationally simple model. We also sell our optical systems, which leverage Acacia technology, but traditional transponders into three of those hyperscalers. And so hyperscalers will use optical systems for some applications, like a long-haul application or even a metro application. But generally speaking, they're using pluggable technology for data center interconnects. So both optical systems and the DCO pluggables, we play on kind of both sides of that. We are trying to shift the market to pluggables wherever we can. But for customers who want to continue deploying transponders, we're doing that.
I would say the pluggables portion of that is being driven by AI. The transponder-based solutions are more traditional hyperscaler traffic. The TAM for Cisco, I think the big places that we play are in Ethernet switching and in optical and optics. The Ethernet switching TAM in 2024 between front end and back end was about $3 billion. We see that growing to about $16 billion in 2028. And the optics that go with that are growing from about $1 billion to about $6 billion. So like $22-$23 billion worth of TAM is what we see driven largely by AI. Now, that's not exclusively AI, but that would be significantly less if there was no AI in our future.
You've shared, or Chuck's shared on the call, some metrics around the AI wins and the pipeline. What portion of that is coming from optics versus switching?
Yeah, we're not breaking that out, Simon. I know you'd like to hear that, but we're not breaking that. But I would say a significant portion of that is optics and optical. And switching is obviously a big, big part of our business here. And with Silicon One, we expect to be capturing more and more of that market over time. But optics and optical represent a pretty significant portion of that $1 billion that we've committed to the street.
Now, one of the other questions that's come up a lot is related to the availability of utility power, electricity, and so AI, I think, is what's catalyzed this topic, so it's not brand new, but from a sort of Cisco perspective, how are you seeing the availability of electricity to power data centers affecting your business?
I think it's a really significant, I mean, power is probably the one thing that we have the most difficult conversations with hyperscalers over. Like power drives decisions. We've had one hyperscaler tell us, like, we will pay more for lower power. Find a way to reduce power. A couple of things I would say, just anecdotally. We've been told. I'm in San Jose. We've been told there's no more power for Cisco in San Jose. Figure out how to move your labs out of San Jose because you're not getting any more power. We're experiencing a little bit of what hyperscalers deal with every day. I think there's a couple of things that are going on architecturally, though, related to power. One is that hyperscalers are now looking at architectures where they can effectively distribute the AI workloads across data centers.
Some are already doing this in a local campus environment, and they're using our optics technology, like the DCO pluggables or the optical systems, to carry those workloads across a campus. But now they're exploring maybe it's not just a campus. Maybe I have to move it across a city or even between cities because I can't get power all in one place. And I think that is going to be a very significant dynamic as we go forward that we're going to see these AI workloads actually moving across a region. And that will put additional demand on things like coherent pluggables as well as optical systems. And then finally, I'd say I may be jumping here a bit, but I think inside the data center, we also have to look at ways to reduce power.
There's technologies like co-packaged optics and linear optics that are likely to play a role there.
Yeah, I wanted to go down that route in that I feel like for the past maybe five-plus years, we've been hearing about the imminent availability of co-packaged optics and then LPO, LRO. The acronyms, they're going to run out of letters soon, maybe. What's your take on all of these opportunities?
Ok, so let me just define a little bit there. Co-packaged optics is the idea first of all, the idea is to reduce power. That's behind all these things. Co-packaged optics, linear optics, it's all about power. It's not cost. It's not space. It's not performance. It's power. And the idea with co-packaged optics is that as the SerDes's bit rate on the silicon increases, that's the IO on the silicon. As that increases from 50 gig to 100 gig to 200 gig, what we know, it's a principle of physics, whether you're traveling over a piece of copper 12 inches or whether you're sending the signal on a fiber 1,000 kilometers. As the bit rate increases, the distance you can send that signal shrinks. That's a fact of physics.
And so the way you compensate for that is you basically re-time the signal as it's moving along, effectively clean it up. Even if you're sending a signal from a piece of silicon to the faceplate of a router, as that bit rate goes from 50 to 100 to 200, you have to figure out how to clean it up along the way. Every time you clean it up, it costs more power. Ok, so that's like the fundamental problem is the bit rates are going to go up so that we can go from 400 gig per port to 800 gig to 1.6T to 3.2. Those bit rates have to go up on the IO. And as the IO goes up like that, it's going to cost more power to get the signal across the board just to the faceplate.
At some point, we're just like, we can't do it anymore. No muss. The idea is take the optic instead of having a pluggable optic on the faceplate that's 12 inches away, put the optic right next to the silicon and eliminate all the need for that re-timing. That's the idea of co-packaging. You ask who in the industry can do that. Well, who has both silicon and optics in the industry? Well, Broadcom has that. Hock will come tell you that they're doing it. Cisco has that. Anybody else have that? No. If the industry goes to that model, it's basically going to look like a duopoly. Part of the reason we've invested in both silicon and optics is because we think at some point the industry is going to go there. We want to be there to play in that market.
But guys like Andy at Arista basically looked at that and said, when we demonstrated that in 2023, we demonstrated a co-packaged solution. Guys like Andy said, wait, wait, wait, wait. I don't think I like that model because that pretty much boxes us out. And we're moving from a model where I've got 10 optics suppliers and not too many silicon suppliers, but 10 optics suppliers and a silicon supplier to one where I've only got two suppliers in the whole industry. I want to find a different way to do this, a different way to skin this cat. And Andy advocated he was the primary advocate, though I think the industry has embraced this idea, a solution called LPO, linear optics, which basically says, pull the DSP out of the optic.
The DSP's job in the optic is to do a lot of cleanup of the signal and condition it so they can be sent not 12 inches, but 500 meters or two kilometers. Pull that out and let the SerDes in the silicon just work a little harder to, instead of moving the signal just 12 inches, move it 500 meters or two kilometers. And so the silicon has to get a little smarter. And the optics that's left in the optic still has to be a little bit better. But you could take the DSP out, and that's a big part of the power. Ok, and the question is, can that be done in a way that still preserves a multi-vendor environment? And there's an MSA that's been formed. We are supportive of that MSA. In fact, the guy that works for me, Mark Nowell, leads that MSA.
I would say there's some good things that have been demonstrated there. There's a lot of challenges ahead of us as an industry to see, can that work in a way that allows multi-vendor? And then the next big thing is, does it work only at 100 gig, or will that continue to work at 200 gig when the SerDes's rate goes up to 200 gig? Because the problem gets a lot harder then. So at some point, we're going to be left with co-packaged optics as the answer. I just don't think that point is within the next couple of years. It's probably further out.
Can I paraphrase the way I interpret you as being willing but skeptical?
Yes, yes, yes. Yeah, and we're prepared to enter that. But I think we're going to look for every other solution possible first. In part, by the way, an ecosystem has to form that can do that manufacturing and do that in a reliable manner as well.
Is it fair to say, in a sort of cynical way, you don't care because you can do either?
I'm sort of agnostic, yeah.
The other area where I think your view would be interesting because you can do either is this idea of coherent technology, which is the technology of choice in the wide area network being used inside data centers. You've got Acacia, which is coherent, and then Luxtera, which is not. How do you sort of think about the prospects of coherent technology being applicable inside data centers?
So there's at least one of my optical system competitors, Ciena, that's, I think, advocating that coherent's going to come inside the data center. And that claim arises from the fact that what I just talked about, like as the bit rates go up, the distance that you can send a signal reliably shrinks. And so inside the data center, we use pretty simple optics technology. It's not really I mean, I don't mean to undermine it, but it's simple relative to coherent, which is much, much more complex sending a signal over a very long distance. But as that bit rate goes up from 400 gig to 800 gig to 1.6T to 3.2, every time you do that, the distance shrinks.
So at some point, you're going to have to use some coherent technology even inside the data center in order to send the signal two kilometers or certainly 10 kilometers. And the question is, when is that point? So the claim is a fair claim to say at some point you need coherent. But the challenge is coherent is more power hungry, and it's more costly. So there's going to be a lot of resistance to bringing it into the data center where we've got effectively much, much cheaper optics. And there's going to be a lot of people innovating in very creative ways to avoid bringing that coherent in. So I would say not going to happen at 400 gig, not going to happen at 800 gig. I don't think it's going to happen at 1.6T, maybe 3.2T.
I don't know how at 3.2T we get to two kilometers yet, but we've got a lot of clever people in the industry that are going to be working hard to avoid bringing coherent in. If at 3.2T we kind of hit the wall and we can't move the signal inside the data center as far as we need to, we'll bring coherent in or some variant of coherent, and we have that technology if we need to do it.
What's the sort of time frame when that might happen? Is it like 2027-ish?
Yeah, I think so, yeah.
And then, in terms of the other technology that might sort of change some of these curves, is the introduction of liquid cooling related primarily to keep the GPUs cool. But how does that affect your industry, your side of business?
So there's a couple of variants of liquid cooling. There's liquid immersion, which you've seen at trade shows with a bubble chamber and everything. The liquid is actually boiling, and you drop the Ethernet switch in there. And then there's liquid cooling that is basically a cold plate. We have liquid passing through pipes effectively cooling the equipment. We are actually supporting both of those right now in various applications for optics. Let me say that. It's a much, much harder problem for optics because there's always some portion of the optic that has basically free space, and you can't put liquid into free space and still send the signal. So it's a much more challenging problem for the optic. I think the industry seems to be leading more towards non-immersion but liquid cooling with cold plate. Other than servers, servers are an immersion environment.
But we're kind of keeping our eye on that right now and working with the hyperscalers to see what they think the right answer is.
The other dimension of the market is around the telecom, wide area networks. Hasn't been as big a market for you, but you've got a couple of projects there. How's that market been in terms of it had a lot of excess inventory, tends to be too small to show up in your overall results. But what are your observations of telco trends?
Yeah, so I mean, I think the telco market in general has been a sluggish market over the last couple of years. A lot of these guys built out infrastructure in anticipation of 5G four or five years ago. 5G never really materialized in terms of a big demand. And so I think we had that coupled with the fact that there was a lot of inventory in the pipeline that needed to be consumed. Some of that is still getting consumed, although I think we're at sort of the long end of that tail now. I sort of expect that by like second half of 2025, calendar year 2025, we'd be in for a refresh cycle for many of these guys. I don't see a trigger out there like 5G. And as I said before, I'm not sure how AI is going to play for the service providers.
That could be a trigger. But I don't see a big trigger like 5G that's going to cause a massive pickup on the service provider side. But I think they'll all go through a refresh cycle. They're all dealing with like four- to five-year-old networks right now. So we'll see sort of a natural refresh. And I expect that in maybe second half 2025.
So believe it or not, we've run out of time. But I always like to close with sort of a similar type question. But based on sort of your interaction with the investment community, what do you think of as the least appreciated aspect of the optical business at Cisco?
Two things. First of all, thank you for asking that. Two things. One is our optics business, which I didn't say too much about, is probably one of the biggest optics businesses on Earth. We don't talk about that. We don't break that out. But we have a very substantial multi-billion-dollar optics business that we sell to all of our customers who are buying routers and switches. But one change that we've made in that business over the last couple of years is we said we've really got the gold standard in terms of quality, in terms of the certification that we do for an optic to make sure it works in all environments and all applications.
And we now go to customers and say, look, because optics becomes a bigger portion of your spend as the cost per port on a switch or router decreases, we want to be considered as your optics supplier. So if God forbid we lose the business to Arista or Nokia or whomever else, we want to supply optics into that application. So we're supplying optics into third-party hosts now as well as Cisco applications. And the other thing I would say is Acacia has been a phenomenal acquisition for us, helping not only on the optical side, but really, really transforming the entire market on the pluggable optics side.
Great. Bill, thank you very much for joining us. Appreciate it. Thank you, everybody.