Great. Why don't we go ahead and get started? I'm Aaron Rakers, the IT hardware and semiconductor analyst here at Wells Fargo. First of all, thank you all for joining us to discuss Cisco's Silicon One strategy with Martin Lund. Before we get into that, I'd like to kick it over to Sami Badri with the investor relations team to go through a quick safe harbor, and then we'll jump right into questions. Sami?
Great. Thank you, Aaron, and thank you again for hosting us. I'm going to read a quick safe harbor, and we'll get rolling. Before we begin, please note that we may make forward-looking statements which are subject to the risks and uncertainties outlined in our filings, in particular our Forms 10-K and 10-Q. Actual results may differ materially from those forward-looking statements. With that, Aaron, I'll hand it over back to you, and we'll get started.
Perfect. Thanks, Sami. You nailed it. That was great. Happy Friday, everybody. I'm extremely excited to host this conference call with Cisco's Executive Vice President of the Common Hardware Platform Group at the company, Martin Lund. Martin, you've got a very deep background. I think you joined Cisco a couple of years or so ago. You were previously at Microsoft Azure. You spent time at Cadence. You spent time at Broadcom. I think it was 10 or 12 years or a long duration there. To put it simply, you've got a tremendous background around silicon and the development of what's going on in the switching market. I'm extremely excited to host this discussion. Before I ask my first question, I'll point out to the audience, we will pause at some point and field any questions that the audience might have.
If you do have a question, feel free to email me at aaron.rakers@wellsfargo.com. You can also Bloomberg me as well. If you'd like as well, you could email Sami at sambadri, B-A-D-R-I, @ cisco.com if you have any questions, and we can definitely try and field those. I have a list of, I do not know, 40 or 50 questions for you, Martin. We are going to jump right in here and get going. Martin, I think just to level set, the first thing I want to kind of bring up is, can you just give an overview of Cisco's Silicon One strategy? What is the current strategy? How long have you been pursuing this strategy? How prevalent is Silicon One currently across the product portfolio for the company? Certainly, we will dive deeper.
Just to level set the audience here, I'd like to maybe start there.
Yeah, I'd be happy to do that. The Silicon One strategy essentially is a continuation of an ASIC strategy that Cisco has been building ASICs for probably four decades. What's new is that we acquired a company called Libra that was founded by some people that had another company called Dune Networks that you may have heard of. They were acquired in 2017, and we introduced the first sort of came-out party in 2019. Since then, we have been working kind of quite intensely on broadening the portfolio and be able to deliver solutions that are customized or optimized for different roles in our portfolio and in different roles in our hyperscaler customers' portfolio. I would say that we're still in the now we announced the I think we have eight distinct platforms that are using our internal solutions to date.
We're in the first quarter of that transition. You should see us over the next, I'd say, three-ish years, maybe three to five years to have completely adopted the Silicon One architecture across our estate portfolio. I think that speaks to the uniqueness of the architecture, us that with a single switching and routing architecture can cover this very, very broad space that we have in our portfolio. Obviously, the one thing I want to point out, a lot of people get a little confused. It's not like one chip. It's not like you have one chip that fits all. It's a range of families that meets low-end needs, high-end needs, roles in the service provider networks, the enterprise, and very importantly, also in the hyperscaler space.
Yeah. I think that kind of segues into my second question. I think from a device perspective, when we as investors think about that proliferation, correct me if I'm wrong, I think Silicon One's been deployed across 17 devices for the company. You recently launched Silicon One into the campus market. What are the you mentioned next three to five years and kind of across the portfolio. What's kind of how do we think about the what's next? You launch into campus. You've done, obviously, some of the core data center and routing functions. Walk us through maybe that proliferation from 17 devices to what does that look like?
Yeah. I mean, as I said, this is a long journey that we've been on. We are rolling them out. Most recently, we introduced them in our campus portfolio. I think we will continue to put it out there. I would say we probably will be done, as I said, in about maybe three to five years. It kind of comes in waves. We have a very significant and comprehensive portfolio. As we're rolling them out, we do not like to kind of pre-announce stuff too much. You can expect that we will be fully adopted in that time frame.
Yeah. That's perfect. Part of the Silicon One strategy has also been kind of being agnostic, right? Not just offering Silicon One on internal platforms, but supporting S1 deployments into third-party devices. Where are you seeing traction from that perspective? Can you help us appreciate that side of selling Silicon One silicon only, if you will?
Yeah. In fact, we have the publicly stated, and part of the strategy is to, and this is what I believe you have to do, you have to meet the customers where they are and what they want. We offer the Silicon One solutions as part of our branded portfolio. For hyperscalers and others, we offer also they can buy the silicon, just the raw silicon from us. We do the in-between, which is sort of the white box model where we use the very strong systems engineering capabilities we have to build these systems. In many cases, hyperscalers would want to put their own software on top of it. That is really the flexibility that I think sets us apart. I also think that is very much needed in this space, especially hyperscalers. They like the control a lot.
That is what we are seeing. We have announced that we have adoption in five out of six hyperscalers. That is in different ways of consumption. We have all three models at play.
Yep. I am going to definitely get to the hyperscale discussion because I do not know that I have been on any call where that does not come up or AI comes up, obviously, in this context. Before we go there, just kind of sticking with the portfolio and the proliferation, I think when we think about this, you look at the market, you have got Broadcom, which obviously you well know. You have got Marvell. You have got Merchant Silicon, if you will. Internalizing that silicon strategy, how do we think about that from a margin perspective? I would assume that Silicon One has not just portfolio advantages, but model advantages from a margin accretion perspective.
I think that's a reasonably fair assumption. When I was at Broadcom, and I was there for 12 years at the beginning of basically 2000, 2012. When I left, we had a very healthy business. It was very, very strong gross margins and a very significant market share. Last time I checked, Ted has and Broadcom has some pretty good margins too. I think they've gone up from there. That, of course, means that when we make our own chips, and we do not have to pay that margin, we have to invest the R&D dollars in, but the margin for us is obviously better.
Yeah.
Maybe significantly so.
Yep. And so that just naturally accretes to the model. So just to kind of paraphrase what you said, we're a quarter of the way through that full extent of the proliferation, which should be kind of over the next three to five years, that rest of the 75% we should see across campus, data center switching, routing, across the portfolio effectively.
Yeah. I would say I would be careful on the percentages. That's sort of the conceptual level. I think of it. I don't know if it's a don't take that to the bank.
Yeah. No. Yeah. Perfect. When we think about describing Silicon One, I think in the past we talked about a unifying architecture or unified architecture approach. That implies that the prior landscape was fairly fragmented. Can you elaborate a little bit on that in terms of the context of historical divisions or trade-offs that customers faced in networking and how Silicon One's fundamentally kind of changed that perception or interaction with customers?
Yeah. I mean, it still is the fact that it's a fragmented solution space. If you go to merchant silicon vendors, they will have one architecture family for sort of the routing class devices. They'll have another one for sort of enterprise-class devices and yet another architecture for the more high hyperscale/AI networking. What we have is an architecture that is flexible enough that it can serve, as I said, there are different devices, but the underlying architecture is flexible enough and has the programmability capabilities that allow us to serve all the use cases with a fundamental thing. It is a benefit for us internally in terms of we have less architecture investment. We can invest into one and then get lots of coverage. From our customers' perspective, they do not get these fragmented behaviors in their network, which can be very, very difficult to manage.
Because it's like you get some difference of behaviors in your network that basically creates problems operationally.
Yeah. I think we're going to maybe I'll touch on that a little bit later around the programmability and some of the different underlying architectural considerations of Silicon One. Before we go there, again, I want to make sure I get this out in front of people. We have to talk on any conversation about hyperscale. Obviously, Cisco has performed well this year. You've outpaced the initial expectations as far as hyperscale orders. That $1.3 billion of web scale orders, and again, emphasis on backend by the company. Last quarter, you talked about 60/40 split between switching and optics. Maybe first, can you talk a little bit about how we should think about the order to deployment or revenue conversion cycle of those opportunities and how maybe that's changing or could change as we look forward?
Maybe this is one of those questions that Sami would be greatly equipped to cover.
Yeah. Happy to jump in here on this one. So really kind of trekking to plan, the second half of fiscal year 2025 was really when we expected to start seeing the conversion of backlog to revenue from the AI networking orders. In fiscal Q3, we did see a sharp step-up increase in rev rec from the AI networking activity contribute to the quarter. Like we've said before, we expect Q4 to also be comparable or more of the same. We should be able to tighten up orders to rev rec from here. We just had to ramp it up to a certain level. From here, I think there's going to be less of a lag from order receipt to rev rec conversion.
Sami, just to kind of put a finer point on it, six to nine months, and maybe it compresses a little bit from there. Is that fair?
Yeah. It's fair to say that it tightens up.
Yeah. Okay. I think, Sami, or Martin, whichever, there's been some clear confusion around those AI orders and what they maybe encompass. Can you clear that up for us? Backend, frontend, optics, transceivers, maybe I'll throw in there top of rack versus spine. Just help us level set us or the audience on unpacking those numbers a little bit more to the extent you can.
Yeah. I'll go. And then, Martin, you can kind of chime in if you have anything else to add. When we give the street AI networking orders, they have to follow a very specific criteria. One, it has to be a backend infrastructure order, right? This is all purely non-frontend. This is essentially completely new TAM for Cisco, right? It has to start with AI backend first being the requirement. From there, it can be one of three product areas. It's Silicon One, it is Series AK, or it is the optics and optical portfolio that are sold alongside those deployments. The other filter that we're using to classify our AI networking orders is it needs to be one of the top hyperscaler web scalers to fit into that criteria. We do not put any enterprise numbers in there.
We don't put any additional kind of items from any other category of customers in there. It's a pretty strict definition. We did this mainly because we wanted to show the street examples of how our product portfolio was relevant to where the industry was going. I think up to this point, we've done that. We may not disclose AI networking orders as often, maybe starting in FY2026, but we'll see where we are on that point. Hopefully, that gives everyone an idea on the filters that we're using to classify those orders.
Yeah. Just to be very succinctly clear, like the HUMAIN deal , some of the things that we've seen on the sovereign side, not inclusive of those numbers. You'll still break out like web scale, but those would be additive, obviously, to those dynamics as we move forward.
Yeah. The biggest, I think, dynamic to remember, short answer is all sovereign AI is not included in those numbers. All the new sovereign AI projects that we have announced are not contributing in any way to the AI networking orders we have already announced. Maybe just to kind of fine-tune this a little bit more, a lot of our backend activity, we are essentially competitors with everyone who tries to sell a solution into the AI infrastructure backend with U.S. hyperscale or web scale. Once you leave the U.S. web scale and hyperscale, we have some pretty solid collaborations or partnerships with other technology companies where we think the Cisco portfolio will be incredibly relevant, and customers are very familiar with the Cisco quality base as well as the overall solution set. That is kind of a different vector than working with U.S. web scale.
We really don't expect product orders from sovereign AI really to start ramping up until the end of fiscal year 2026 because most of those projects are in design phase right now. We're giving the Middle Eastern entities some time to complete designs, get the powered shells up and running, and then they're going to submit orders. We think that window is towards the end of FY2026. They tend to build data centers faster in the Middle East than they do in the U.S., and they don't have power shortages the way we have power shortages in the U.S. Those are some of the key differences and comparing and contrasting points on how sovereign AI is different than the U.S. web scale opportunity.
Some of those opportunities could be as sizable, right? Quicker, maybe deployments, and even as sizable as some of these hyperscalers. Fair?
There is one of them, which we believe HUMAIN AI could be comparable to a U.S. web scaler in the magnitude of spend. This is just measuring the quantum of gigawatt aspirations that they want to bring online. It is looking very comparable to what a web scaler would spend at.
Yep. Sami, I really appreciate it. Yeah. Go ahead. Go ahead.
Absolutely. Martin, anything to add to some of the points we just discussed?
No. I mean, I think the only thing I would say as we're going forward architecturally, we see certain signs that this frontend backend delineation is starting to blur a bit. There will be sort of historically, it's been very much there's a backend AI infrastructure, and then there's sort of a web frontend. We're starting to see that these things are blending a bit. As you move more inference out there into the network, you will see that these over time probably blend, and it all becomes, maybe it all becomes AI at the end of the day.
I completely agree with what Martin just said. I think for Cisco, it's almost like a different equation completely because when people think of the backend versus the frontend, they always just talk about what's essentially in the data center, right? When we think about backend to frontend and then the downstream connectivity all the way to the end users, that starts hitting all of our product lines all the way from Silicon One in the backend all the way to Silicon One in the frontend or some other chip and system in the frontend. Then those packets start communicating and translating all the way through down to the end users. For Cisco especially, the backend to the frontend kind of ratio is not very clear, right? I think it might be easier for other companies to claim something like that.
For us, we're in a stage of more discovery and outline than it is a very kind of scientific equation.
Yeah. That's perfect. I think some of that conversation kind of segues, Martin, to the competitive landscape a little bit. Oftentimes, in my discussions, you talk about Silicon One, talk about with Chuck. One of the comments that comes up is, "Hey, we're winning because we're not Broadcom," right? Broadcom dominates the merchant silicon side of the market. Is there other, what other attributes of Silicon One would you say have been competitive insertions for the company when you look at it, either be it relative to Broadcom or obviously NVIDIA comes up a bit in the discussion with their Spectrum X product? How would you frame that, the differentiators at the architectural level?
Yeah. No, I definitely was smiling here because it's not enough not to be the other vendor. Obviously, supply chain diversification is an important element. It's something that, sort of, there is a good driver for it. At the end of the day, you have to have good technology. In a way, every opportunity is heavily contested. What our customer tells us is that the value in our architecture is that we have a unique flexibility with our programmability in our packet processing. You can think of that as we have more intelligence, or we have more flexibility, or we have the ability to adapt with our silicon when it's made and shipping. We have the ability to change the forwarding behavior to match unique requirements. That can be load balancing requirements that are very important in the network.
It can be certain security capabilities that they see. That is one of the key things. There is another key element. There are basically three. There is the programmability. There are the packet buffering capabilities that we have. We have a fully shared buffer architecture, which is unique. Others do not have that. We also have more capabilities or more buffering in it, or memory, if you like. That means, especially for AI, that allows us to basically absorb larger bursts. AI traffic is a very bursty thing. Therefore, we do not drop packets, so we do not have to issue flow controls. That leads to basically shorter job completion times. We did recently show a benchmark online that showed that we were better than the other vendor that has a lot of market share, the Tomahawk device, and very comparable to InfiniBand. This is the packet buffer architecture.
Finally, we have some advanced telemetric capabilities in the devices that are being highlighted by a customer. Flexibility or programmability, the packet buffer architecture, and the telemetry. Those are the capabilities. That is essentially what makes a switch. That is the thing. This is not just something we say. It is what we get from our key customers.
I'm going to jump over. I got Sami, definitely jump in if you had any email questions into you. One I've gotten emailed in is that, "Martin, specifically you, how is your approach in leading Silicon One different?" The person asked relative to y'all. I think what he's referring to is basically NVIDIA, right? You guys obviously have a relationship with NVIDIA. How does that, I guess, paraphrasing the question, how does that relationship work? What's your approach in leading Silicon One relative to having a relationship with NVIDIA that the company's established here?
It could be a question also about Eyal Dagan, which was my predecessor. We can take it both ways. It's hard for me to, the way I operate and the way I've always been operating is sort of in the Silicon switch base. This is what we did at Broadcom. This is what we're doing here. It's basically, I like to say, it's 5% strategy and 95% execution. With anything that we do in silicon, it comes down to execution. If you miss and you have to redo a chip, you lose half a year or a year. If you execute with predictability and with quality, you can get to market faster. It's an execution game. There's a lot that goes into that, relentless focus. I think also just have a really top team. It's not about the army of people as much.
It's about really, really having an environment that attracts and allows top engineers to thrive. That's sort of the combination. I think we spend a lot of time anticipating our customer needs. There's this thing about going out to the customer and saying, "What do you want?" They may not know what they want. You have to actually think two, three, four years ahead about what are the things that are required in that time frame. That's sort of my approach to it. It's an execution game. With that, it's just focus every day.
That's perfect. Sami, I don't know if you had any email questions in or else, I'll just keep going down my list.
I haven't received anything yet. I wanted to add a little bit more kind of answers to the question regarding us versus NVIDIA, right? Here's what NVIDIA sees. It sees its addressable opportunity with enterprises to adopt GPUs as very significant. What they want is they want to accelerate or enable that distribution channel. They've really identified us as that key partner to do that. As part of that, we are doing reference architecture integration to make this as resistance-free as possible for enterprise customers. If this comes at the expense of us competing against them in some of the web scale backend activity, I think that's a trade-off that they're willing to live with. We get it. They probably get it because at the end of the day, they want to sell GPUs. They want that to proliferate dramatically.
That is a much bigger piece of the wallet than splitting hairs on the networking part, at least for enterprise. I think big picture, that's what they see. That's actually big picture what we see as well.
I would maybe add to that because there's actually very little overlap in the portfolio if you compare our portfolio of switching solution versus NVIDIA. Now, they have their Spectrum X portfolio. We have our G-class devices. We have deep packet, deep buffered solutions. We have big routing solutions. We do WAN interconnect. A lot of these data centers or AI training clusters are so large, they don't fit into one data center. You really need to have very advanced, high-performance data center interconnect solutions. We offer that. Our portfolio scales, it goes down into a much smaller set into the enterprises where a lot of inference is going to be happening. I think for sure, we're competing a little bit. In the end of the day, most of these networks will need a dual vendor solution for the networking needs.
It's a good practice. It's a risk mitigation. Another way to think of it, you don't want to have your $25 billion AI GPU cluster be bottlenecked because of a switch problem, right? You really don't want to have that. Having multiple, at least two vendors in your network is a good practice. It's proven out again and again to be the way that.
Yeah. That makes a lot of sense. I think maybe going a little bit deeper there. Again, I think I mentioned earlier, I think you said 60/40 split between systems and optics this last quarter as far as the order momentum. If I were to double-click on the switch side, can you talk about where you're at? Meaning, I get a lot of discussions around top of rack. I think a couple of quarters ago, Cisco announced that they had won a Super- Spine design win at a hyperscaler. So where are we at in terms of seeing inside that backend network fabric opportunities for Cisco top of rack and increase opportunities maybe at the spine and maybe leaf layers of the network? Any thoughts there?
Yeah. At the highest level, actually, we have solutions, and we have wins in all of the roles in the network. It's different. Maybe different customers have picked us for different roles. By the way, a lot of this is sort of a journey. You do not just show up with a new chip, and then people are going to design you into the core of the network and say, "Hooray," right? There is lots of you got to get the software. Their software needs to be ported on. You need to operationalize and prove that you can work with them, that you are predictable, that you have the quality. It is a little bit like first you have to get in, get known, get the right to play. Then you can get another piece, and then you can get another piece.
If you keep executing, you end up with getting your fair share. In terms of technology, we have the technology solutions that put us in play for all of these roles.
Yep. How much does backend pull frontend? Have you seen that? I know that we talk a lot about, obviously, Sami, you just went through it. It's backend-centric. They have to define where the deployments are taking place. Have you seen evidence where winning the backend actually is pulling frontend opportunities for Cisco?
I think yes. The answer is yes. Because once you're in a broad definition of a given, if you're first in, the software is on. You can get repurposed into other parts of the network more easily. You have the credibility to play there. Maybe they say, "Scale is a little bit different." Do you have a version of this chip that's maybe half the speed or half the capacity? In most cases, the answer is yes. I think there is a but it can also go the other way. You can get in the front, and then it's more of the notion of being a proven vendor. Do you have the right technology? This is very much a game of being ready with the technology and being proven when the technology choices are there.
I like to say it's like in the hyperscale space, it's like going to the final exam every day, except you can't pass. You can only fail. It's like you just got to show up every day. The other thing about hyperscalers is that it's not like they just sit and wait. You have a problem, and then they send you an email. They expect you to, in minutes, to reply. It doesn't matter if it's 3:00 o'clock at night. It's a very, the reason I don't, I mean, if you really think of it, hyperscalers, it means hyper. It means it's very, very large and different than other functions. The expectations are equivalently high.
Yep. Sami, I have one other question emailed in here, and it kind of speaks to the competitive dynamics. Maybe I'll paraphrase it. The pace of innovation is remarkable, right? We've dramatically changed over the last couple of years, arguably since mid-part of 2023 with this whole AI insertion. How do we think, I think you recently launched the G200. You're shipping 51.2T Silicon. We've seen the biggest competitor announce or introduce their Tomahawk 6 platform moving to 102. How do we think about the competitiveness of Cisco in keeping pace with that dynamic in the market? I have a follow-on to that. Why don't we start there? Interested in your thoughts.
Actually, I think that would be a great one for Martin to answer, right? Yeah.
Yeah. Yeah.
Yeah. I think, I mean, it's true that Broadcom have been first to market with their new devices. And most recently, they announced the Tomahawk 6. It was a little bit of an odd announcement because it said it was shipping. But normally, when you talk about these devices, you say you're sampling. And then you say you're in production. So this notion of shipping, it's a fun one. It's like the volume shipping or are you sample shipping or did you ship one or did you ship five? And it's hard to tell because there are no customer quotes in that release. So it's a little bit hard to judge that. But obviously, there's a lot of pressure on it. And Broadcom have executed well for many years.
I would say that for our 51T, our volume kind of shipment where we're starting to kick in was a bit somewhat later than when Broadcom did. However, the gap is shrinking. I think we'll be continuing to execute. I said it's an execution game every day. I think we'll be catching up. Hopefully, at some point, we'll be out there with solutions at the same time, maybe even beat them once in a while.
That's perfect.
It is at the end of the day, it really comes down to you can sample. That's fine. What really matters is when can you take the device to production because that's when the networks can get turned on.
Right. Right. Makes a lot of sense. One of the other interesting things, and this is kind of the second part of the question, was with that introduction, we're starting to see we talk a ton about Ethernet on the scale outside. But there's this idea of scale-up Ethernet or UAL ink. And when you look at NVIDIA's rack scale architecture, you've got nine switches of ND link sitting inside that architecture connecting those compute nodes. How do you envision Ethernet as far as playing in the scale upside? Is that an opportunity that we or myself and investors should be thinking about for the traditional switch vendors or the switch ecosystem on Ethernet?
I've been an Ethernet guy since the '80s. I've always been betting on Ethernet. The reason why is there was also ATM and token ring and other technologies that came around. I've always been sort of on the Ethernet and IP side because I like to say it's simpler. It's easier for me to understand it. It has also proven again and again and again to be able to meet the use cases. We made it work for cars. It's an infotainment network for cars. We can make it work for high-fidelity audio. You can make it work anywhere. I think it works for AI. It's not only a year or two ago where it's like, "Oh, InfiniBand rules the world, and you can never get the performance." That debate is over, right? We already know that that's fine.
Ethernet is perfectly actually better suited because it's scaled for these super large clusters. For scale-up, it's a compute interconnect. And Ethernet, I believe, will be able to adapt to those requirements. But it will go in and replace NVLink? No, I don't necessarily think so because NVLink is a walled garden that has its own, there's no really reason why you would open up and put Ethernet in there if you're NVIDIA. Maybe they don't do that. For other classes of GPU that benefits, I think there's definitely a play there. It might be an optimized version of Ethernet that has some better, that looks more like what we're doing with UAL ink and other things. UAL ink is one of those technologies that are out there. There's also UAC, Broadcom announced, Su, and so forth. Absolutely. We're probably on the team Ethernet here.
Whatever is the best choice. I think most importantly, it has to be open for innovation to really thrive.
We're not wrong in thinking that Ethernet, there is an opportunity on the scale-up. We'll see how that evolves. UA link or whatever that version might be, it's fair to assume that that could present an incremental opportunity for the.
I think over time, definitely, I believe that there will be GPU architectures that direct connect to an Ethernet variant that is used for scale-up.
I'm going to kind of move past the hyperscale dynamic a little bit. I'd be remiss if I didn't ask this question. When you think about Cisco and the success that the company's reported here over these last couple of quarters, how do you size the market? Or Sami, jump in here. How do you think, how does Cisco think about sizing the AI networking market? I have to ask. We'll see where it goes. Do you have any kind of thoughts on share aspirations in that market opportunity?
I can say about share aspirations. I think in general, we are, I don't think we have a very large share today. I think the market is quite significant. More is better when it comes to market share. I think we're hopeful that we'll be able to execute and deliver market-leading technology that, and then with the trust and our open approach to the industry, that we'll get our fair share. It's hard to do these predictions because it really comes down to a lot of factors. We can control our execution. We can control what we deliver. We can't control what others do.
Yep. Sami, any thoughts on sizing?
Yeah. So here's a common theme. The dollar TAM for backend networking continues to increase. It's also starting to include a lot more categories of products. It doesn't look like it's just switching anymore. It's including switching, NICs, optics. In some cases, it's including cables. So it's kind of blurry what the exact TAM really is. But there's kind of two things we're comfortable saying. We think we're taking share in this market. And we're also growing from a very low base of revenue relative to the actual TAM size. If we just go back to our investor day in mid-year 2024, it was something around south of $20 billion TAM a year. So I think by now, everyone realizes and knows that it's much larger than that.
If you put that in the context of what we've already reported year to date in FY2025 from an AI networking orders perspective using our filtering with web scale and products, there's a large runway for us. Like Martin said, it's an execution game. It's an exam you have to show up for and can't pass, but you can definitely fail, which is a very realistic quote.
Yep. Yep. And just to be fair, Sami, I mean, a lot of the things that we've seen develop for the company on this front was not going back to that analyst day. When you look at that networking business growth profile, a lot of the things that we've seen have been additive since back then to what you were contemplating.
Yeah. I mean, cloud traction was contemplated in there. We did outline AI use cases and status and use case wins at the 2024 investor day. I think what is additive is things like the NVIDIA partnership is additive. Getting Silicon One into more of the product families is additive. Sovereign AI is also additive, right? Those aspects are additive. The AI networking traction was factored in, right, like we were asked at the time.
I think there's one thing I wanted to kind of bring out, which is maybe not fully understood. The technology innovation that's happening right now is moving fast, as you said. At the technical level, it means that it's not enough to build a nice chip. You have to build advanced systems with it. That means that the systems and silicon optimization is becoming more and more important because of cooling, because of signal integrity challenge, because the systems that we're building are really, really, really difficult. You can see it. NVIDIA is probably the other company that has the same chops in building systems as we do. Obviously, they're doing very well with rack scale architectures.
This co-optimization comes like co-packaged optics leading into a systems development with cooling, with the whole nine yards, and then with the ability to manage a supply chain and ramp production in a very, very steep way and manage quality, etc. That is where it starts to play into the strength of Cisco. I am very optimistic about that.
Yeah. That's a great point. That co-optimization, I think, to your point, gets more and more increasingly complex, which plays into the Cisco strategy. You mentioned there, Martin, the supply chain side, right? As you see these opportunities present, Sami, earlier, we talked about the order to revenue rec progression. How much of that has been a challenge for the company? Are you able to keep up with the demand in terms of getting your allocation of, if it's wafer starts from your manufacturing partners? How have you managed that? How has that been progressing?
I would say that we have, I mean, we are subject to the same laws of physics that everybody else are. There can be limitations and so forth. We have a very long-standing relationship with all the key players in the ecosystem. We are happy that we are able to get the support that we need when we need it. I would add that we have also a very resilient supply chain that allows us to bring on capacity or move things from different locations to other locations. This is sort of the big machine, is very much an advantage for us in terms of managing supply chains, not just the semiconductor piece, but also just the whole, there is a whole terrorist complications to the world and so forth.
I think we're able to have a very flexible and adaptable supply chain is an advantage for us.
So I'm going to maybe rapid-fire here last couple of questions unless, Sami, you have any that came into you that you want to make sure we get to.
No, go ahead. Go ahead.
The quick last couple of questions, and I want to be appreciative of your time, Martin. And again, I really do appreciate you letting us host this discussion. One of the things on the technology side that I read a little bit about, but I'm honestly a little bit confused of where and who's what, is the radix scaling, right? I think Cisco's talked about G200 having, I think it's 512 radix scaling in some things I've read as a competitive advantage. Can you help level set that? Is that a competitive advantage? And how would we think about that as far as important when you're in dialogue with these hyperscalers?
Yeah. It's sort of at the end of the day, it comes down to how do you build a network? Can you build a complex super spine cluster, if you like, so you can distribute traffic around it? It is an advantage, yes. It is an advantage that we have another part of our architecture choices that we kind of foresaw that as a key requirement. We were able to scale that. In many cases, the competitors, they see it too, but they may have an architecture that does not allow them to scale it that effectively. The radix essentially comes down to can you address every port in the network? We have 512 100 gig ports in our 51.2T. We can address every single one of them.
That means we can put a very sophisticated network together that allows for very good load balancing in a cloud fabric. That is where it comes down to. I think most of these advantages are transient because you see a need and can then spin a chip to meet the need. Obviously, it takes about two years from when you get the requirement until you have a chip that you can maybe sell.
Yep. Final two questions for me, Martin. I'll let you get on with your day. We can, hopefully. The one question I want to ask is just we talked about NVIDIA a little bit earlier. Where do you think, how would you characterize the progression that we've seen of traditional enterprise AI adoption at this point? When do you think maybe that starts to inflect more materially as you look at the relationship with NVIDIA or just more broadly for Cisco? I'll throw out the second question. It will end the discussion. I'd be interested in how the team around Silicon One has expanded, developed, anything you want to share on that front of just Cisco's investment here, right, and this being core to the focus or the strategy of the company. I'd be curious if any metrics you might be able to offer us there.
Yeah. I can start with the last one first. I would say that we have been increasing our investment in the past. You can go online and you can see that we probably have hundreds of open positions in our networking. But it's not that they're just sitting there being open. They were filling them. And then we open some new ones because there's a good turn on that. I think we have added probably 30%-40% capacity over the past 18 months into this. It's rough numbers. But again, it's not just a numbers skill. It's about the talent and the quality of the people and the ingenuity of those folks. But yeah, it's a scale game. And I'm happy to get a lot of support from corporate to continue this game. As I like to say, it's the deep end of the pool.
There's not that many players here. We got to show up with intensity. It's not for the faint of heart. The field is narrowing. I think there's probably just a handful of teams in the world that can do this kind of analysis.
Yep. And then on the enterprise side, I guess, I guess at a high level, how do you envision that becoming more pervasive? I'm starting to hear a little bit about the $20 million, $50 million kind of projects when in traditional enterprise starting to show up on AI deployments. But be curious of what you're seeing.
Yeah. I think, I mean, there's the kind of the so AI-specific stuff. But then there's a lot of modernization that our customer sees that they need to have a modern infrastructure across their estate, not just inside the data center, but across their estate. So there's sort of a pent-up need for an upgrade on modernization. And the enterprise AI deployments are coming in. I think you're right. I've seen similar things during that chunk. But it's like a long tail. Hyperscale is at the very top. And then you have Neo Clouds and Sovereign Clouds. And then you move down into smaller enterprises. And we'll see, and we have technologies in the pipe for that, that we'll see true edge applications that are the rocketized inference at the edge doing work out there. And I think that's where the huge opportunity lies.
I mean, at the end of the day, training is good, but a lot of this stuff needs to actually do productive work. That is at the edge of the network where a lot of that will reside. It will be both cloud. It will be geo clouds and so forth. There will be a lot at the enterprise edge, I believe so.
Yeah. I think Cisco and that NVIDIA relationship is well positioned for that. I want to appreciate your time. Sami, I do not know if there is anything you want to throw out there that maybe we did not discuss that you would leave investors or the audience to kind of think about here. Or if not, I am happy to end it at that.
No. I think we covered a lot. Thank you, Martin. And Aaron, thank you for hosting us. I'll hand it over to you. Anything closing, and we can wrap up for the day. Yeah.
Yeah. Perfect. Again, I appreciate your time and allowing us to host the discussion. I do not know, Martin, if you had anything left to say, but if so.
I would say that it's, I'll just say, watch this space. We're kind of just getting started.
Perfect. That is a good way to leave us to think about things. I appreciate the time and again, letting us host this call. Hopefully, you guys have a great weekend. Thanks so much.
Thank you.
All right. Thank you.
Cheers.
Thank you, everybody.