Broadcom Inc. (AVGO) Earnings Call Transcript & Summary

Good afternoon, and thank you for attending JPMorgan's 19th Annual Technology Investor Forum. My name is Harlan Sur, semiconductor and semiconductor capital equipment analyst. This year, we have something very special. Broadcom is kicking off a series of mini-Analyst Days to provide a deep dive into the company's key semiconductor and infrastructure software franchises. We're pleased to be hosting the first of these events with a focus on Broadcom's networking franchise. As you know, we have written many reports over the past few years on Broadcom's leadership within networking, especially in cloud and data center, where we estimate that Broadcom has greater than 80% share of switching and routing silicon market. Broadcom also has a very strong position in providing optical connectivity solutions alongside its switching and routing products. With us from Broadcom, we have Hock Tan, President and Chief Executive Officer; along with Senior Vice Presidents, Ram Velaga and Alexis Bjorlin. Ram is General Manager of Broadcom's Switching Products Division, where he's responsible for the company's extensive Ethernet switch portfolio. Prior to joining Broadcom in 2012, Ram spent 10 years at Cisco in a variety of product management roles. Alexis is General Manager of the Optical Systems Division or OSD at Broadcom, where she is responsible for developing and manufacturing devices used in optical communications as well as developing Broadcom's silicon photonics-based platform for interconnect and co-packaging optical IO. Prior to joining Broadcom in 2019, Alexis spent 4 years at Intel, where she led the Data Center Connectivity Group and commercialized Intel's silicon photonics technology. So today's agenda will include introductory remarks from Hock. Ram will then take us through Broadcom's switching and routing opportunities and market position, including the flagship Trident, Tomahawk and Jericho family of switching and routing products. Alexis will then discuss Broadcom's opportunities in silicon photonics. Following the presentations, we'll have a lot of time for Q&A. And so with that, I want to thank all of the Broadcom team for joining us today. And with that, let me go ahead and turn it over to Hock.

Thank you, Harlan. It's a pleasure here -- to be here today to kick off our first analyst event of 2021. I'll start with a few words about Broadcom and then lead into a discussion of what you did call the networking franchises. So next slide, please. So on this, Broadcom today, to put it in context, is a leading provider of technology to the global IT ecosystem. So how did we get here? As you'll see on the left on this slide, at the time of our IPO in 2009, Broadcom had revenue of just around $1.5 billion came from -- and this $1.5 billion comes from -- came, I should say, and it still continue to come from 8 product franchises in semiconductors. So let me pause to take a minute to explain what we mean by those franchises. These are category-leading market verticals with 3 attributes in common, which, by the way, applies across all of our product lines today. First, has to be a mission-critical technology, which ensures a sustainable end market. Second, we are the technology leader in that vertical. And finally, thirdly, we have typically the #1 market share. And so we're starting with these 8 franchises. We embarked over the last many years on a series of very disciplined and selective acquisitions of key technology franchises in semiconductors and software -- infrastructure software, nurturing those acquisitions and operating all of them synergistically on a very robust Broadcom platform. So the business model, I have to emphasize, it's not a consolidation play as people may think nor a simple roll-up of the industry. It is, in fact, about monetizing on a sustained basis, returns on technology innovation where technology, the underlying thesis, is an extremely deep profit pool. So looking back on this slide. By 2020, if you look at the left, we drove our revenue to approximately $24 billion or 16x growth from 2009. Now of course, this came both through acquisitions but also organic growth. What's interesting is our R&D spend over this period increased even more, something like 25x. But lo and behold, pointing to the strength of the business model we have, operating profit over this period of time grew 84x. All right? And today, we're now 25 franchises in semiconductors and software and a critical and strategic partner to the world's largest enterprises. Next slide, please. Now we compete, as you well know, on innovation, technology and for good reason. We trace our roots in fundamental technology to icons, icons such as Hewlett-Packard, Bell Labs, AT&T, Brocade, even Broadcom Classic Digital QAM. Now we gained with these acquisitions exceptional engineering talent and continue to invest for growth. Today, Broadcom is one of the largest, broadest, I would say, IP portfolio in the semiconductor industry, with over 23,000 core patents. So where do we focus all this technology and talent? Next slide, please. And this shows in a fairly succinct manner where we are. We are not in general-purpose compute nor are we in memory. But we are very laser-focused on everything that's important in connectivity -- connected communication, another word for it, throughout this IT ecosystem represented here. So if you look at the left, our technologies address growing need here in these networks for increased bandwidth and at the same time, improved latency. So looking at the left we have devices, and we connect them wirelessly and through wired devices in the home. And in -- through broadband, in subscriber access and gateways and in -- through routing, through carrier networks and in switching, enabling cloud and enterprise data centers. And even here for chip-to-chip interconnects within storage and service. In fact, we are such a critical part of the ecosystem, we believe 99% of all Internet traffic today has to cross at least one Broadcom chip. But for today, turning on the next slide, please. As highlighted here now on this slide, we will just talk about our products and technologies represented in switching and routing, as we call it, our networking franchises. Other areas like wireless, broadband, we will do in following Analyst Day events throughout 2021. So moving on, as the next slide shows, networking represents over 25% of our consolidated revenue. And in fact, it represents over 35% of our semiconductor revenue, shown here in red. And it covers products such as switching, routing, physical layers and collectively the compute offload products. What is particularly significant is that collectively, networking has been one of the fastest-growing markets for Broadcom over the last 5 years. So at this point, Ram and Alexis will provide more color on diverse and broad portfolio of products and exciting opportunities we see ahead in networking for Broadcom. So with that, I'll turn it over to Ram on networking.

Thank you very much, Hock. Good afternoon. My name is Ram Velaga. I'm the General Manager of Broadcom's Core Switching group. During my section today, I'll give you an overview of our switching and routing silicon business. Next slide, please. Our business is based on a couple of fundamental tenets. Number one, the need for bandwidth is ever increasing, right? Whether it is binge-watching videos, online learning, gaming, security cameras, image sharing and just the ever-increasing number of connected devices are all contributing to this. The second tenet for us is that the physics of building silicon is not getting any easier. All of you probably heard this comment about Moore's Law is flattening. And that's very true. Moore's Law is flattening with diminishing returns between successive process nodes. The benefits that you got from going from 60-nanometer to 40-nanometer to 28-nanometer now to 16 and 7, every one of these successive nodes, it's diminishing returns. And secondly, maybe this is something that is less heard of, is the maximum size of a chip that can be built has not changed. Over the last 10-plus years, the biggest chip that you could build is still 800-millimeter square, okay? And so these 2 are very important things to keep in mind, which is: one, laws of diminishing return on Moore's Law; and the other, is the largest chip that you can build is still 800-millimeter square. So that's the second element of our fundamental tenets. And then the third is that the Internet connectivity that we enjoy today is really kind of a collection of many subnetworks. And each of these actually have very different attributes, right? And one way to think about this is it's kind of no different than how, let's say, our interstate highways that then connect to state highways and then which eventually feed into the local streets, right? If you think about it, they're all very different and they have their own attributes. And you cannot build a local street like a highway and vice versa. So that's kind of a very important concept. And when you think of that concept and then think about, okay, how is this whole networks or the Internet as you think about it, segmented or grouped together? Do you think of this as essentially the enterprise network, the service provider network and then the hyperscale data center, right? So the enterprise network, what is it? That's kind of pretty much what you experience when you walk into your office, a traditional enterprise, which might have a large campus or a branch office or something along those lines. And then the whole idea there is to be able to provide a workforce, whether it is remote workforce or in an on-premise workforce, but generally a mobile workforce, access to the enterprise resources. The second segment here is the service provider networks, right? And when you think about the service provider networks, basically, what are they doing? They are connecting the homes, the offices and the campuses. And these service provider networks sometimes connect to other service provider networks, which is -- think of it as an AT&T connecting or pairing to Verizon, pairing to Comcast or somebody else. And then eventually, all of these service providers take these clients and then they connect to some kind of a data center, whether that's an enterprise data center or a branch data center or it's a large-scale, mega-scale data center. Now when you think about these connections for a service provider, they kind of traverse to central offices, right? Then they go into this metro network. And then they go into this very long-distance network sometimes between cities and countries and then even between continents, sometime -- over the subsea, right? Now the service provider networks are really dominated by fiber connections. And they're very shared with many customers on them. You have many enterprise customers on them. You have many consumers on top of them, and they're all kind of competing for bandwidth. Now the service providers kind of thrive on essentially providing a service-level agreement, right, which is, if you subscribe to my network, I guarantee you get certain kinds of bandwidth. And so it's kind of an important aspect of a service provider network. Now if you look on the other hand, the networks for the hyperscale data center, right? And first, let me just talk about what do we call a hyperscale data center. These hyperscale data centers are things which are very large warehouses running a few football fields long and typically, the kinds of one that either an Amazon, an Alibaba, Facebook, Google, Microsoft and Tencent build. Now inside these data centers, the primary purpose of the network is really to kind of interconnect these hundreds and thousands of compute and storage nodes, right? You're trying to interconnect all of these together and then be able to provide users access to these compute and storage nodes. Now these networks tend to be very high bandwidth. Now when I think -- when I talk about bandwidth in the kind of the context of the analogy of the highways I was talking about before, think of this as the bit of the highway, right? And you might have many lanes on that highway, but it's really the bit of the highway. So these networks tend to be very high bandwidth, which is very wide, and they also tend to be low latency. Now when you think about latency, think of latency as, yes, you might have many lanes and you're very wide, but how fast can you travel on a given lane, right? So think of latency as a speed at which you can travel in a given lane, and think of bandwidth as the bit of on it. So across these different segments, the enterprise, the service provider and these mega-scale data centers. Today, when you actually think about the system-level revenue, when I talk about systems, which is people who essentially sell a system, whether it is Arista, Juniper, Cisco, Nokia and so on and so forth, it's about $50-plus billion of Ethernet IP business revenues that these system companies derive every year. So our switching and routing silicon portfolio addresses all of these segments at a silicon level. Next slide, please, which is Slide 11. So what I'd like to do is kind of now that we've talked about networks at a very high level, I want to kind of talk about what is a switch/a router? And then after that, I'll kind of take you into -- deep dive a little bit into our silicon, right? So these networks that we just talked about are really just a collection of many of these interconnected switches and routers that you're seeing in this picture here. This particular one is an example of what is called -- essentially referred to as a fixed configuration pizza box, which is about one rack unit high and one rack unit is about 1.75 inches. And this particular one also has 32 front panel receptacles for high-speed optical modules, right? So you basically are connecting these optical modules to bring light in and out of this pizza box, right? So when you look at this switch router, there are effectively a few main components of this. One is the switching and routing silicon, which is built by Broadcom. Two is the network operating system, right? And the network operating system is available from the likes of Arista, Cisco, Juniper, Nokia and so on and so forth and the likes of H3C and the others in Asia and other places. So that's the network operating system. Then there is the platform hardware, right? And this platform hardware, kind of -- you can always think of it as sheet metal in PCB and a few components and the fan tray and the power supply and so on and so forth. And generally, today, these are built by original design manufacturers in places like Taiwan and so on and so forth. And lastly, the pluggable optics that I referred to before, but you can kind of actually really get it from a variety of module vendors. Now what's happening? And the interesting part here is as the performance and the level of integration of our silicon improves, right, the network operating system and the platform hardware, I'm referring to that by all the sheet metal, the PCB and all of this, really becomes simpler and simpler and easier to develop. And that's actually a very important element, which is -- what I'm saying is as we integrate more and more onto our silicon and as we keep increasing the performance of our silicon, the network operating system and kind of the platform and the hardware actually become simpler, easier to develop and also achieves higher levels of availability and so on and so forth. And that's actually a very important element of our product strategy, which is how do we drive simplification at a system level by integrating as much as we can in our silicon and increasing the performance of our silicon. And if you can go to the next slide, please, which is Slide 12. So when you kind of think about it, we're saying, "Hey, you can simplify the system by integrating more and more onto your silicon." But then we talked about if we're integrating more and more into the silicon, but at some point, you're faced against the challenges of flattening Moore's Law and then the fact that the maximum die size you can build is only 800 millimeters square. So there's like multiple competing forces here. How do you kind of rationalize all of that and build a winning platform? So we kind of made the decision looking outside in -- this was a decision that was almost made 9, 10 years ago, saying, "Look, we understand networks. We understand the systems market." So really, you cannot build the exact same chip to serve the needs of the enterprise market, the service provider market and the hyperscale data center because each of them have very different attributes. And so we said, "Okay. You cannot have the Swiss Army knife and pretend to be this one device that conquers the world." But really, you have to be very surgical and build the right device for the right purpose, right? And especially because of the physics of -- the physical constraints that we talked about, the Moore's Law limitation and the 800-millimeter square die size limitation. So by building unique silicon for each of the markets, we make the difficult choices of what kind of goes into a particular chip to meet the performance target. And just as important, sometimes even more important, is what do we remove from that chip? Because that particular chip is not going to play in a particular segment of the market, right? So I just kind of want to pause for a second and say, this is extremely important, right, which is you cannot have one size fits all because you're constrained by the physics. And hence, you have to make tough choices on what goes where, and more importantly, what doesn't go where. Because then if you get the silicon right, it really simplifies the system. The whole name of the game here is simplify the system, right? So that's really kind of what we are doing. So we have a purpose-built portfolio, and that purpose-built portfolio allows us to actually execute with product leadership. And without this notion of a purpose-built in a portfolio for each segment of the market, there's no way you can have product leadership because it's just not supported by the laws of physics, right, period. And there's just no debate about it. There's no other way to do it. So Slide 13, please. So just based on what we just discussed, right, what we have done then is we said, "Okay. So there's these 3 segments of the market, and these 3 segments of the market have very different needs. And so now let's align our product portfolio to these 3 segments." And again, this is not something we've come up with in the last year or the last 2 years, and this is something that we've kind of laid the groundwork for almost 10 years ago. And we said, there's the enterprise networks, there's a service provider and the hyperscale data centers. And each of these segments, we will serve them with a separate switching routing in a silicon family, okay? So there is a segment that we serve with a separate switching and routing silicon family. Now within each family, we have multiple chips at different performance and cost points. So for example, we could have -- in the service provider segment, we will serve it with a product line called Jericho. And within the Jericho, we have multiple versions of Jericho with different bandwidth points. You might go from 100-gig to 400-gig to 1 terabit to 4 terabits and 14.4 terabits. And because even there within a family, it's not one size fits all, right? So that's what we do. And so the way, essentially, our product families map with these different market segments is we have the Trident silicon that maps to the enterprise segment. Next is the Jericho silicon family, which is purpose-built for the service provider networks. So for example, you might say, "Hey, how is the service provider Jericho family different than Trident? What does it have that the Trident does not have?" So one example is the Jericho family of products have very deep buffering. And what is buffering? Buffering is the ability to store data on the silicon during periods of network congestion, right? So that you're not just dropping the traffic and saying, "Hey, can you retransmit?" But because you're a service provider and you have a lot of different customers who are essentially being given SLAs and all of them are using the same shared network, you will have periods of network congestion. And you've probably seen this, whether you're watching Netflix or something else. You'll sometimes see this circle going on in the TV saying buffering or network congestion. That's what's happening. And so you kind of need an architecture that supports these attributes. And then on the other hand, you have the Tomahawk family products. Here, it's really designed for these hyperscale data centers. These hyperscale data centers, they essentially have these large data centers with a lot of fiber running inside, and they want to be able to send large amounts of traffic at very high capacity and low latency, right? And then our Tomahawk silicon today, because it's purpose-built for this market, offers the highest bandwidth per chip in the industry, right? And not only does it offer the highest bandwidth, we are consistently a year, 18 months and sometimes more, ahead of the rest of the marketplace by the time they kind of get to a similar bandwidth point. Because, again, we're building -- purpose building it and not trying to be one size fits all, in which case, you just can't achieve it. And sometimes you have to go to the next process node to try to even just catch up with what we do. So Slide 14, please. So while we are purpose building our silicon, we provide our customers with a single unified development environment. And that's extremely important. Because by providing this single uniform development environment with a consistent application programming interface, also known as APIs, our customers can really leverage their engineering resources to cover a broad set of markets that they might play in. So you might have an OEM that covers every place from a hyperscale data center to an enterprise, to a service provider. And they can write to a consistent set of APIs. That means they get a lot of engineering team leverage and yet be able to address a lot of markets that our silicon supports, right? And so this is kind of no different than when you think about the APIs that app developers are provided with, whether it is an Apple iOS platform or an Android operating system platform, right? And the benefit of it is, once that app, in this particular case a network operating system, is written to our APIs for a given silicon, this application is easily ported to other family of silicon in our portfolio, right? So something that's written to the Trident product line is easily ported to our Jericho product line or easily ported to our Tomahawk product line. But just as importantly, it's not just between these product lines, but it becomes very simple for them to move from one generation of the product -- one product to the other generation of the same product. So once you ported your network operating system to our APIs on Jericho1, you can go to Jericho2, you can go to Jericho3 and so on and so forth, right? So again, what are we doing is there's different networking segments, different requirements. We distillate down to different product families, but all of these product families are still bound together by a single development environment and a common set of APIs. And if you go to the next slide, please, Slide 15. All of this, why? Because it's an important question, you say, why do you have these different product lines? Why do you go through these efforts of having this common set of APIs? Because remember what I said in -- as I said before, it's all about what we do with regards to integrating and scaling the performance of our chips is all singularly focused on simplifying the system, both at a hardware level as well as the network operating system level and then eventually at the network level, right? And so I want to share with you a few examples that kind of bring this to life and give you a very concrete example of why do we go through the pains of this portfolio that we have and then eventually how the customers see the benefit of this. This particular slide shows 2 options for designing a high-capacity routing line card, right? And on both sides, you've got 32 ports of 400 gig with integrated security. Because this routing is facing the wide area network. You're facing the wide area network. You want to be able to have integrated security so that you're dropping the traffic as it hits the wide area -- as soon as the wide area hits your demarcation point and not necessarily carry this malicious traffic inside, right? So customers want integrated security. So the design on the left, when you kind of design for a general-purpose platform, you have -- don't have integrated security. When you don't have integrated security, then you end up actually requiring 18 external security devices. That brings the total chip count to about 21 devices or 21 chips on this design on the left. Now you look at this design on the right. Because we said, "Hey, look, it's a service provider market. We need to have integrated security, and we're designing a Jericho2c+ chip." So we said, "Let's integrate security like MACSec, even IPSec into it." And these designs are for 32 ports -- 36 ports of 400 -- 32 to 36 ports of 400-gig. We need 14.4 terabits. And let's do it in 2 pieces of silicon. And we were able to achieve that, right? And when you do that, there's fewer parts on the board, the hardware. And when there's fewer parts on the board, the PCB layer actually becomes very simple. Because when you have a lot of parts, you have a lot of crisscrossing signals on these PCB layers. And the way you avoid the signals from interfering with each other is you build a lot of layers for PCP. So what that means is your board becomes more expensive. You're dealing with a lot of signal integrity issues. And then you've had a bunch of rocket scientists to kind of get these systems to be built. And it might be fine and dandy, if you're like vertically integrated and you're trying to make a ton of money out of your customers by selling these things at 10x more than they should be sold for. But that's not the way to simplify the designs, right? So you're dealing with this whole problem of hardware that becomes very complex. Now if your complexity stopped at just the hardware complexity, maybe you can kind of find your way through. But it doesn't stop there. The software problem is even larger. When you have a lot of devices on -- in a hardware, half the time or maybe even more than that, companies employ tens of thousands of software engineers. Because most of the engineers are actually trying to make these different pieces of silicon look like one single piece of silicon and not really focused on building the best protocols and are stable, scalable, whether it's an IPv4, IPv6 protocol, MPLS protocol or something else. And so they're spending a lot of their cycles trying to just make this beast of multiple chips work together, right? And again, that's not the most productive way to build platforms. So not only are you kind of overcomplicating your systems when you don't build these chips right and wasting a lot of resources for your customers who end up using your chips, but also you end up with -- you're kind of in a basket case of device that requires far more power than if you actually built this right. In this particular case, we almost have 50% reduction in power when you compare the design on the left to the design on the right. So kind of I'd like to leave you with, just -- this is one example, and I'll cover a few more, that really the integration and the choices that we make are all around simplifying stuff at a system level, hardware, software and eventually, the networks that get built. Because that's the winning strategy, is that the strategy of saying, "Hey, there is a systems market that's $50-plus billion. How do we enable it and do this right?" So this integrated products like Jericho2c+ make that possible because we make the hard decisions on what features are included and what features are removed when a particular target application does not require them, right? So now this, again, kind of goes back to being able to have the scale to build multiple devices, which are optimized without any damaging compromises, right? And these are very damaging compromises in how you build a silicon and the manifestations operated the system well. Slide 16, please. Now when you look at the slide, it looks for ray -- it looks like you've got a lot of chips, but you actually think about this in the context of what we've been able to achieve in the last 10 years, right? We've gone from chips which were 640 gigs, which essentially means you were able to get about roughly 64 ports of 10 gigs of performance. To today, a device that is a 25.6 terabit, we call it the Tomahawk 4 silicon. And that is 256 100 gigs in performance, right? 256 100 gigs in performance. It can also be used as 64 ports of 400 gig or 32 ports of 800 gig. This Tomahawk 4 switching routing silicon is kind of really targeted for hyperscale data centers. And it's another example of the benefits of the purpose-built in a silicon strategy. This device actually first sampled in 2019 and went into production already in 2020. And to date, no one else in the market has been able to release in production a 25-terabit silicon, right? This is, again, very similar to what we've been able to achieve with Jericho2c+ that I just explained on the previous slide. We made the tough decisions to remove some of the features, which are not necessary because we said, "Hey, look, this is going into the hyperscale data center." When you think about hyperscale data centers, all the customer cares about is moving bits from one compute to another compute, one server to another, one piece of storage to another. And they're not trying to do all these service provider stuff in this device. So let's not confuse ourselves and the customers by having features that we don't need to have in here. And instead use the die size because we are limited, remember, in that 800-millimeter square die size. Let's use that die size to build the fastest performing race car for this market, right? Then you could say, "Okay. That's correct. But why does the hyperscale need this much bandwidth? Why can't they just use the same bandwidth as a service provider has, right?" And that's because, look, it's not just networking that's feeling the effects of the flattening of the Moore's Law, right? The rate of CPU performance has also flattened. But while the rate of the CPU performance improvement has flattened, the application layer is driving a rapid growth in demand for the computational power. And because of the flattening of the CPU performance, this demand cannot be met just by going from one generation of the CPU to the next generation of the CPU, right? So then you say, "Okay. How do you compensate for this?" So kind of compute has moved from this whole notion of a scale out to a distributed architecture that really kind of relies on a scale-out architecture, which is let's just kind of go wide and have a lot of CPUs. But you got to interconnect these CPUs. And when you have to interconnect these CPUs, you have to interconnect them with very high-speed links, right? So ideally, if somebody actually had figured out how to make PCI work at very high speeds and made PCI work across a very large fabric and go distances wide enough once that leaves the rack. Maybe they would have solved a lot of this problem by using something like a PCI, but they could not. Multiple people have tried, but it's not. It's not going to go beyond a few feet, a little more -- a few inches maybe, let alone a few feet, right? So you need something like an Ethernet, but you need Ethernet that is low latency, very high performance to be able to connect all of these CPUs together. And so when you're doing it, you cannot get confused, thinking that you're building something for the service provider, and you can use it here in the mega-scale data center. So to meet this demand for this interconnectivity because the compute architecture is going scale out and it's kind of getting distributed out, we're constantly innovating our switch architecture to really provide a step-function improvement in device bandwidth, right? And this is against the backdrop of diminishing improvements in shrinking geometry and the maximum die size of 800-millimeter square. Okay. So again, this bandwidth where we're able to drive in this device, again, is only possible because of this purpose-built strategy. And so by removing the functionality that's not needed in the data center, and believe me, actually, every successive generation of the Tomahawk class device that we make will go through many hours of what features were there previously that we should be removing so we could use that then trade-off for the higher bandwidth that is needed. And that is -- takes up more of our time than what features will be added. And so by going through this exercise, we are very judicious with the real estate that is very expensive. And we are able to consistently double the bandwidth of our device every 18 to 24 months and keep scaling, right? And I hear this every time. Somebody comes in and says, "Oh, there's 5 other players who have -- who say they have something like this and the like." Right? We will see and we will talk when things actually show up in the marketplace. And then number two is, it's not about one device, one time. It's about being able to execute consistently with investment protection because, remember, this is the infrastructure business. You don't build silicon now, throw it away, 3 years from now, say, "Oops, you know what? The team that did it didn't do it right. I'm going to go hire another team and try to make it come." And then go to the customers and say, "Replace -- rip and replace everything you bought previously. Let's start all over again." That's not how the infrastructure business works. It's about predictability. It's about steadfast execution. It's about investment protection and so on and so forth and which is why it kind of all ties back to a strategy in what we've been executing. Now if you can go to the next slide, please, Slide 17. So when you think about doubling of the bandwidth, right? The first thing that most people will logically think about, okay, great. Doubling of the bandwidth. So I can replace 2 devices of the previous generation with one device of the next generation. Is that it? No. That is not the case. And this is a beautiful thing in networking. This is every time you double the bandwidth of a device, you end up taking 6 devices and collapsing it to one. It's very counterintuitive, but it's true. And kind of this picture shows you why. So when you take a device that's half the bandwidth of today's available state of art, right, so today's state of art, we're already shipping it in production, the 25-terabit device. And we've sampled it over a year ago. If you said, "How do you build a hardware that's 25 terabit, but I've only been given building blocks of 12.8 terabit, how do you do it?" You actually need 6 of them. You use 4 of these 12.8 terabits facing the network or facing the front panel of the switch, and you use 2 more of this in the back connecting the front 4. And so what ends up happening is you burn a lot of the bandwidth of these 12.8-terabit devices just to interconnect amongst themselves so that those 6 of them together will act like one 25-terabit device. Okay? So that's actually not very intuitive to somebody who's not living and breathing networking. And then they'll say, "Oh, I'm only at 12.8 terabit. But don't worry. In 18 months, I'll get to 25 terabits or 24 months. Hold your breath. But in the meantime, let's just take 6 pieces of silicon and try to hold them together." That doesn't work. It's too power-hungry, and it's slow because you have multiple switches that your traffic has to pass through, and that means the latency gets very high. And that thing is going to consume way more power than you need to. And it just complicates the systems. And there's no better way to show it to you than if you kind of go to Slide 18, which is the next slide. I'm not poking fun at others on this. So I'll actually show you what was possible with the state of art or with our own silicon one generation ago which to build a platform which was 12.8 terabit, the image on the left actually had multiple silicons in there. We had at least 6 of the half-bandwidth devices. Now when that moved from the previous generation platform to the next-generation silicon that we currently have, you're able to collapse the system from something that is greater than 8 RU in height. And remember, each RU is about 1.75 inches to something that's about 1 RU in height. And you're able to take something that has at least 6 pieces of silicon devices to one. And in that process, if you just kind of even look at this image of the hardware on the left, it has a lot of slots in it. Each slot is a PCB layer. Each PCB player has one piece of silicon on there. And it has a lot of traces running on top of it. And all of these PCB layers are then kind of connecting to some kind of fabric in the back. And then you're trying to power this thing up and you're trying to cool this thing up and you're trying to have this air moved from the front to the back. And you take up a lot of rack space, which is very expensive real estate, that the cohost -- the hosting sites and all of them would like to charge you for. And it's an expensive proposition. And as I talked to you about it before, it's an expensive hardware to build. It's a complex hardware to build. On top of that, even the software that runs on it, it's very complex because it's trying to keep all these different pieces of silicon, working as if they're one piece of silicon. So by doubling the bandwidth because we have dedicated product lines and integrating more and more for a specific market, we've really simplified the system from what it looks on the left, what it looks like on the right. And in the process, effectively what the customers enjoy is at least a 75% reduction in system, power and cost and a similar 75% plus reduction in the height of the chassis, right? And then obviously, in terms of volume, that's a lot of volume that's also saved. So then you say, "Okay. Great. Really simplifying the system." So what does it really mean to that ecosystem of whole networking business, this $50-plus billion of networking system business that we talked about? And we kind of understand that, if you can kind of go to the next slide, please. Really, what we are able to do here is, I call it the democratization of networking, right? So when silicon is not built optimally, you end up having a lot of engineers you have to hire to build this very complex hardware, like the picture on the left that you saw in the previous slide. And then you have to have signal integrity engineers. You have PCB engineers. You have airflow and thermal management and power engineers and so on and so forth, and you kind of have a big engineering team. Then you go and say, "Well, I also have to have a very large software team because I have to build this distributed software infrastructure before I can even write my first networking protocol or before I even write my first -- a software feature on top of it, just to make all of these different pieces of silicon kind of company." Right? And so what that means is the barriers to entry into the networking industry were very, very high because this silicon was not done right. So now that we've kind of steadfastly focused on thinking about the end markets and then our customers in between and making the tough decisions on investing very heavily, obviously, in this process and making the tough decisions in what's in our silicon, how many pieces of silicon we build, we've really kind of democratized that ecosystem that helps build this overall networking platforms. We build a silicon, multiple pieces of silicon, and then we put it out there with a common APIs and our software development kit around it. And then we have ODMs and OEMs that build hardware on it, whether it is a traditional OEM or ODMs from Taiwan and so on and so forth. And then we have software partners. That software partner could be an OEM, whether it is Arista, Juniper, CNS, Cisco somebody else or it could be software companies that are purely built only to build software. And they don't build the hardware, and they let you just procure the hardware from anybody else in the ecosystem. And then on top of those, there are those who provide support and service and provide management capabilities and so on and so forth and top. So net-net, really, what we are achieving here is simplifying this ecosystem so more players can participate in it. Everybody can do what they're really, really good at and in the process, accelerating the pace of innovation, being able to get the technology in the hands of the end customers a lot faster. So you're not kind of hamstrung in the hands of one OEM, hoping to get it all right. And then there's this kind of a virtuous cycle that kind of feeds on itself. There's so many people developing on top of our platform. We get a lot of great feedback on what we might be missing, what needs to be added or something that we are not doing right. And we get to kind of innovate and get the product out fast. And then the nice thing about it is also, when you kind of think about this ecosystem, our switches go from everything from kind of a base station to a switch in an automotive Ethernet in a car to something that is an industry Ethernet in time-sensitive networking, in the service providers and hyperscale data centers and so on and so forth, and everybody kind of features this capabilities. And so we've developed a set of IPs that can be used across these markets, but they're so fundamental that these IPs can actually be used across these different chips even though we build dedicated chips. So if we go to the next slide, please. So while we have these different product lines, the Trident, the Jericho and the Tomahawk class of products I was mentioning to you, we also build multiple chips within each of these product lines. And so every year, we put out 6 to 8 new chips. If you just kind of step back and think about it, 6 to 8 new chips is roughly a chip every month -- every 6 weeks to every 8 weeks. Now it doesn't necessarily always fall that way, which is a chip comes out every 6 to 8 weeks. But over the course of a year, we come out with 6 to 8 chips per year. And this is just in my division. I'm not counting what Alexis does or the remaining 18 organizations in Hock's franchise -- collection of franchises, just my team, right? So when you look at it and say, "Okay. Well, how do you achieve the 6 to 8 chips per year?" It's because we don't have this one team that's kind of going from one chip to another chip to another chip. You kind of have these multiple teams are running in parallel across the globe. And each one of these sites is capable of producing multiple chips a year. So we have a team of about 1,200 people in my organization, over 98% of whom are engineers. So that's a good thing, which is over 98% are engineers. So there's very little overhead doing a lot of PowerPoint slides and selling to each other, people who are focused on what they need to do, who are very good at what they do, and they methodically crank out silicon, right? And this team has been developing, the switching and routing silicon for, at this point, close to 20 years. And that's all we do, just build silicon, right? We're not trying to do building the software and building the hardware, and they're kind of trying to go and sell an entire system, get confused in between. No, just focused on building silicon. And then -- but how do we do it? Because building these many pieces of silicon is not cheap. When you think about this, sometimes some of these chips cost between $50 million to $100 million, if you're starting from scratch. There, we leverage the Broadcom's corporate silicon platform. This corporate silicon platform gives us access to the most advanced process nodes, right, which is 16, 7, 5, soon to be 3 and so on and so forth. And these process nodes -- every time you go into this new process node, you have to develop this extensive libraries for memories. And we do our own custom memories, we do our own custom memories because if you just took memories off the shelf, chances are, you will never be able to get the kind of memory density to build a switch chip with the kind of bandwidth that you need because we may want to have to do multiple region rides into the memory at very high speeds, which is not what you get off the shelf, right? And then we have to develop our SerDes. And when we develop our SerDes, it's not just about developing 50-gig SerDes and 100 gig SerDes or 25 gig. We have to be able to operate all of these SerDes, not one SerDes at a time, but 256 or 512 of this at one time and then be able to operate a different lens, right? Sometimes be able to reach BSR distances or MR distances or LR distances because there's all these different use cases we have to sell. And so we could not do it if we were not actually built on top of Broadcom silicon technology platform. And then on top of this, obviously, we have a very comprehensive development methodology. Look, in the software world, you put a few bugs out there. And then you come back and say, "Oops, I'll fix it again. I'll throw a few more engineers." And the next thing you know, there's a blob of course that continues to build. And it's okay because you know what, I can just put the next patch and the next patch and the next patch, and life goes on. But that doesn't work that way in the silicon world, right? As a matter of fact, if you think about it, building silicon is far more complex than building software. The way we build silicon is we start with the customers' needs. And then we kind of actually end up designing what the product spec is. We take the product spec, and then we actually then write the C code for that. And then we do an architectural black box validation of that C code. Which is a bunch of guys who say, "Hey, I don't know what the C code is, but I've read the IEEE specs. And then the IETS specs, and I'm going to build this black box testing environment. I'm going to test the code." Then the C code gets passed, and then we take that C code and then actually then give it to the guy who writes the RTL. Then there is another guy who kind of doesn't know what the RTL that was written it, but if they go develop another black box to test it. And then we didn't emulate it before we actually first turned it into silicon, which costs tens of millions of dollars of CapEx to build the simulation environment. So there's a very complex, comprehensive development methodology that we continuously improve, right? Every silicon we put out, if there was a mistake in it, we kind of go back and say, "What failed in the process?" And we keep improving the process. So the end result of this is that we have a very reliable development cycle that we can deliver 6 to 8 pieces of silicon every year on a cadence of almost a chip every 6 to 8 weeks that, generally, we go to production with a zero version of the silicon, okay? Or a zero a versions of the silicon. So that's a very, very well-ingrained process. And which we couldn't do without the Broadcom technology silicon platform underneath it. Next slide, please. So because of what we do and how we do it, it's a classic -- here's an example of what is possible with this in a world-class team that we have and the process we have in place and the access to the Broadcom technology platform we have. This is a picture of a 12.8-terabit Tomahawk -- sorry, 25.6-terabit Tomahawk 4. And this has 30 billion transistors. That's more than 4 transistors per person for every person on this planet, right? And it is the largest monolithic networking device ever to go into commercial production. So this is already in production. We're not talking about something that's going to come in the future or something we just telling our customers saying, "Hey, hang in there. Wait up. Something is coming," right? It's sampled in 2019. It's in production in 2020. And by the way, just an important concept. When you go from samples to production, we take -- it goes through a year. And generally, what happens during that year is when we first come out with the samples, we distribute it to all of our OEM customers, our end customers, and they start their software development on our silicon when we provide them the samples. And generally, they are done hardening their software and ready to go to production over that period of about 12 months from the time we sampled. And over that same period of time, we go through all the testing that we have to do in kind of fine-tuning our manufacturing processes with our partners and everything else, our testing and everything else. So we kind of just essentially hit time to production at the same time that our system partners are ready to go to production. So -- and again, in this particular silicon, we made a decision in 2018 -- actually 2017, 2018, we kind of looked at the market and said, "Okay. We're going to build this 25-terabit device." But majority of the optics ecosystem is still on this 50-gig PAM-4-based platform. So if we went and built something with 100-gig PAM-4, we might be coming to the market too early and our Tomahawk 4 might not ramp as quickly as we would like to do. So we said, "Hey, why don't we actually build it with 56-gig PAM-4s, but which meant you need 512 56-gig PAM-4 SerDes in a single package, okay?" So think about this. You've got 512, I actually think you cannot even cut flashing thread into 512 basis and put it around something that is a 2 plus square inch wide and do it well, right? But we have 512 56-gig PAM-4 SerDes coming out of this package without signals interfering with each other and then being able to actually make this 25-terabit device work. We say, "Okay. Can we do it?" We work with our packaging engineers built on, again, our Broadcom silicon technology platform, and we built it. And lo and behold, it actually went to production. It works, things are in production, and people love it because they don't have to wait to make this 100-gig SerDes work, right? Because the optics ecosystem already has this 50-gig PAM-4. So what it means is our time to market is very, very fast. So what we have already done with this now is, well, we've said, "Okay. The first version of it, we came out with 512 56-gig PAM-4s. The second version of it, we're going to come out with 256 of the 100-gig PAM-4 SerDes. And that is also already sampling to that. So kind of goes to say that the scale in which we operate, the IPs that we build, all of these technology platforms that we have allows us to really come out with industry-leading platforms that we are able to take to -- take them to time -- take them to market quickly with the time to market advantage and really enable a lot of our systems partners, systems vendors as well as our end customers. So if you can go to the next slide, please, Slide 22. This is a picture that really kind of shows what an end-to-end portfolio looks like, right? And this particular picture kind of show everything from -- there is a base station and there is a fronthaul on the base station between the antenna and where the base station processing happens. And from there, it kind of connects into the backhaul of a network for the wireless traffic. And all of that also feeds into the wired traffic that's coming in either from their enterprise or it's kind of coming in from the homes being connected or branch offices which are connected. They all kind of go with some kind of a metro network and then kind of eventually go over a larger core and to hit the data centers, whether they're enterprise data centers, branch data centers, edge data centers or large-scale mega data centers, right? And so we have this end-to-end portfolio that is essentially supported by our silicon strategy of purpose-built silicon and multiple devices in between any one of these families. So I'd really like to kind of, at this point, wrap up my segment. As I wrap it up, I'd like to leave you with a few thoughts. One is, we're really focused on building the industry's best silicon for networking, absolutely no compromises, with separate silicon families for each market segment, okay? But together, they provide a holistic end-to-end platform, all the way from the base station into the data centers, connecting the enterprises to homes and everything in between. Two is that our development program is really built on a disciplined and a mature methodology. And we have the luxury of building this on Broadcom's silicon technology platform of the advanced process nodes and all the associated IP that goes with them. And our engineering team is world-class, operating across multiple sites. And really, it's all about -- it all comes down to product execution. One of the people I really respect in the industry once said, he said, "Look, innovation has a shelf life that is defined by execution." So people can come and talk about all kinds of innovation. But at the end of it, if you don't execute, that shelf life of the innovation pretty much takes scale. So we keep executing, and we execute with also quality. And quality always is our focus, #1 focus. And the other point is it's really a robust multibillion-dollar ecosystem that's formed around switching routing silicon. And we've really democratized networking market and enabled a broad wave of innovation, right? Now think about it. This is -- we are doing to the networking industry or rather done to the networking industry, and we'll continue to do it, what industry-standard CPUs have done to the server market a long time ago. Some of you might even remember the story of SunSpark, right? They came out and said, "Hey, look, we have the SPARC ASIC chip that will kind of compete against the 680 industry standard platform. And by the way, we will also sell the SPARC chip to anybody who cares. Well, HP, do you want to buy? Dell, do you want to buy? Or somebody else, do you want to buy?" Come on. Those guys are not going to buy your chip. You are a vertically integrated player, okay? So as much as you'd like to become an industry standard, you are not because your business model itself is very confusing. Are you trying to be a system player? Are you trying to sell your Java software? Are you trying to sell something else? So rest is kind of history. You have to pick what you're focused on and do it extremely well, and that's what we do. We just focus on silicon. We're going to build silicon very well. The next thing is customers really value what we do. We've had to make many hard decisions in the silicon development process. Some of these yield important simplifications and cost savings for our customers, right? And what this results in, we show up on time. Every 18 to 24 months, we show up with a product that doubles the bandwidth or adds a bunch of other capabilities. And the ecosystems kind of comes -- has come to count on us. And we also get a lot of good feedback from them, and we keep evolving and keep building onto this virtuous cycle. So lastly, right, we continue to innovate. Look, it's been said before, the paranoid survive. I am paranoid. My team is paranoid. We don't take our success for granted. And over the last decade, our products have evolved. And they've evolved from simple switches to very advanced switches, right? And then from simple routing to very advanced routing. And most recently, even to the mobile fronthaul. Look, 5 years ago, if I told you, the core of one of the world's largest service providers and the edge of the service provider and even their sell-side routers would be running on Broadcom's merchant silicon, you'd probably laugh this off, right? You would have said, "Oh, great. Your switches are switches. They're not routers. You can probably use it in maybe some megascales because you don't need a lot of features and so on and so forth." But that's not the case. Over the last year, 1.5 years, you've probably heard from visionaries like Andre at AT&T and the others, they've come to us 5 years ago and said, "Hey, look, you've done something very interesting at the megascale. How can we transform this whole service provider networks we build the same way?" And we said, "Great. All we're looking for is an awesome partner," and we've done that today. And this is not just in the U.S. We've done this many places across the world. And as you know, in some places outside of the U.S., things like 5G and the 5G backhaul networks are being built out at a much more rapid pace than here. And we have, from our vantage point, the ability to engage more broadly -- very broadly with different partners, and we've been doing this. So we evolve our capabilities, and we'll keep evolving to what's next, right? We've gone from basic layer 2 switching to layer 3 switching. And anybody who understands the OSI stack comes, okay, what comes after layer 3 switching because layer 3 switching is routing? And I'll let your imaginations run wild and what's possible next. But anyway, so what I'd like to leave you with is there's a lot more we can continue to integrate because as we integrate, we make the chips simpler, we make the system simpler. We've democratized networking. We all grow together, right? And what we want to do here then is how do we get the traffic in and out of our switching and routing silicon much more simpler and faster and more efficient. And to talk about that, I think the most exciting part of the presentation today, I want to hand it over to Alexis. Thank you for your time.

Thank you.

Thank you, Ram. So Ram just provided an overview of Broadcom's advances in switching, highlighting how our technology enables the broader networking ecosystem. However, Ram's products end at the silicon edge. The real challenge lies in connecting it all together, from the chip's edge to server storage and out to the wide area network. This requires petabytes of cross-sectional bandwidth within a single data center. And since I'm pretty sure all of us have been binging Netflix recently, I'll put that in context. With 1 petabyte, you can transmit more than 10,000 2-hour movies in 4K resolution every second across the data center. Achieving this bandwidth is a challenge of truly massive physical scale. In my presentation today, I'm going to start by showing you the real-world challenges of connectivity, what's known as layer 1, the physical layer, and introduce Broadcom's silicon photonics technologies to address these challenges. Next slide. Today, most of you are already aware, hyperscale data centers are being built out at an unprecedented rate at unprecedented scale. In fact, the size of these data centers are at times limited only by the available power that can be drawn to the site. In this slide, you see a picture of a hyperscale data center in the image on the left. This is Digital Realty's Loudoun III data center campus in Ashburn, Virginia, also known as Data Center Alley. It's 2.3 million square feet in area and delivers over 225 megawatts of IT capacity. The picture on the right shows an internal view of Google's Council Bluff (sic) [ Council Bluffs ] data center in Iowa, highlighting the scale of these operations on the inside. Equipment in these sites span the entire building. You can see racks upon racks and rows upon rows of server and storage, requiring kilometers of fiber to connect it all together. Cloud service providers are spending over $100 billion on CapEx annually with a staggering 15% compound annual growth rate. We all know these data centers are moving tremendous amounts of traffic and tremendous amounts of data, both within themselves and across interconnecting data centers on campuses or even now at regional zones. Cisco's Global Cloud Index predicted 19.5 zettabytes of cloud data movement in 2021. Ram just highlighted all of his advances and gen-on-gen advances and high-bandwidth switching, which -- tens of thousands of which can connect up to hundreds of thousands of servers and storage elements inside these data centers. But the bigger challenge and what we'll talk about today is that it's millions of optical interconnects are required to connect this all together and then back out to the WAN. On to Slide 25. So we have here, what we're showing is a visual functional diagram of how compute storage and even specialized compute clusters for HPC or AI and ML workloads are all connected together in order to provide any-to-any connectivity from any compute element to any storage element or back out to the WAN. There, I'm showing here the multiple layers of switches that are all interconnected by the blue lines, which are the optical interconnects. These together constitute the connectivity fabric. Data center traffic is expected to grow at a compound annual growth rate of more than 25% for the foreseeable future for reasons you already know: workload movement to the cloud, AI, big data, just to name a few. So in order to keep up with this growth, cloud service providers have broken from the traditional server architectures and have moved into virtual and disaggregated pools of resources, both within the data centers, which coined the phrase decade ago, warehouse-scale compute or the data centers as the computer; and now, of course, across regional campuses and zones. As compute scales out, the fabric or the interconnect itself can become the bottleneck to moving, processing and storing data. One additional data point. Today, only 4% of the IT spending is on the public cloud. As this percentage grows, the hyperscale cloud operators need to invest even more time and talent, maximizing each unit of compute efficiency, inevitably furthering this disaggregation and increasing the requirements that are being placed on high-performance fabrics. So indeed, the connectivity fabric is growing. It's increasingly a increasing percentage of spend and power consumption in the data center. So crucially, this all requires power-effective and cost-effective optical interconnects. Moving on to Slide 26. I'm going to bring you back to Ram's image that he showed earlier today of what a traditional switch system looks like: a pizza box with the networking switch silicon and the discrete pluggable optics. In these systems today, optical interconnects already consume 50% of the power in the network. And on a dollar-per-gigabit basis, cloud service providers already spend 10x more on optical bandwidth than they spend on the switch ASIC bandwidth. Without a breakthrough in the technologies that are being used and a profound change in the architecture of the system, we expect the upward trend in power consumption and in cost differential of the optical interconnect to only continue. In the past, optics were thought of merely as the plumbing connecting everything all together. But now with the constraints on power and the increasing bandwidth and performance requirements, the optical interconnect has become central to the data center design. Our customers are telling us that it's not only driving their networking CapEx, but it's actually a central and key factor in the design choices of the next-generation data center architectures. So the main point here is the traditional system architectures where we have switch systems with pluggable optics can't keep up with the exponential growth in bandwidth required within the constraints of power and cost. A true step function is needed. The time has come to reinvent the architecture of the network platform or the system itself. What our customers need and what Broadcom is uniquely positioned to deliver is the comprehensive engineering solution from the switch silicon all the way through to the fiber connectivity to prevent optical connectivity from limiting efficient and performance bandwidth and compute scale out. Broadcom's solution is to integrate this optical functionality onto the switch itself, delivering a photonic-integrated switch, all-in silicon. This is a super large opportunity for Broadcom, increasing the silicon content of the switch system by tenfold, so into the tens of thousands of dollars, and at the same time, offering tremendous value proposition for our customers in terms of performance and cost breakthroughs. So as you'll see on the next slide, 27, I'll walk you through what Broadcom is doing in this space. So Broadcom is unmatched in our capabilities and technologies to achieve this. We've been investing in leadership networking, signal processing and optical capabilities for over 3 decades. On this slide, you'll see 5 key pillars that are required to re-architect the network platform. And Broadcom has demonstrated leadership capabilities over the first 4 pillars already. So you're already aware from Ram's talk of Broadcom's exceptional networking switch ASIC competencies, including leadership SerDes and DSP. But we are integrating these capabilities with our 30-year heritage in mixed-signal ICs, lasers and fabs and advanced semiconductor manufacturing and test. With all of these, we have the foundation to deliver a disruptive solution. But until now, a critical element to create this solution was missing, the photonic-integrated circuit or silicon photonics. Well, you all are aware, the electronics industry has evolved from discrete transistors to highly integrated ICs over the past 50 years. But silicon-based photonic ICs are still in their relative infancy. They only have a couple of years of production under their belt. Today, we're introducing Broadcom's first integrated silicon photonics solution. We start with our market-leading switch. And we extend the switch capability to incorporate optical I/O directly on the switch substrate, breaking the traditional segmentation and paradigms of system design. As Ram said, it's all a part of the simplification of the system to enable the broader ecosystem. This will enable industry-leading economics, volume and power efficiency at scale. So moving on to the next slide. I'm excited to share with you what our silicon photonics looks like. We have designed, just as Ram has said with the switch silicon, a purpose-built silicon photonics platform specifically driven for integration. It brings all the discrete elements of the ecosystem together onto a single silicon package with a unified architecture. It has 3 critical capabilities. One, it has a massively parallel optical processing on a single chip. Two, it has high-density packaging capabilities with other silicon and electrical components. And three, it enables this profound shift in how the network platform and systems are designed and segmented. So unlike many of the early silicon photonics entrants who focus on optical transceivers and discrete elements, our purpose-built silicon photonics solution has been designed from grounds up with breakthrough technology to enable seamless integration of optical I/O with ASICs on a single standard industry package. It's focused on very high-density photonic ICs that are integrated directly on substrate with Broadcom's other market-leading networking elements, the switch, the DSPs and, of course, our lasers. Now I'm going to jump into a little bit of technical content on the key components that are in the silicon photonics platform itself before moving back out to the system-level view. So in this design, we've addressed power. Most importantly, we've engineered a system that directly drives the photonics from the switch SerDes or from the edge of the switch itself. What this does is it eliminates all of the system retimers found in today's switch systems and in the optical transceivers themselves as well as the high-cost PCBs that are required to route the high-speed signals to the front of the panel in the switch box. This allows us to achieve an electrical interconnect power of less than 1 picojoules per bit, a significant improvement over today's designs. And we've developed a thermal management system that can handle the optics and the ASIC in a single package, operating under single heatsink. We've, of course, also addressed density. As Ram was talking about 512 lanes of SerDes, we have to match that with the optical capability. So we do this by shrinking the optics to match the switch I/O, offering the industry's highest bandwidth density at 500 gigabits per second per millimeter of die edge and with 64 single-mode fibers attached to a single die. For comparison, today's transceivers enable roughly this bandwidth per 20 millimeters of optical module. Finally, we've addressed reliability and operational efficiency as a design priority, utilizing remote laser sources with no electrical connectors to ease the deployment and field serviceability of this new solution. Moving on to Slide 29. I'll outline the value we're providing with our integrated photonics switch platform. If you step back and recall the image I used showing today's switch systems made up of discrete components and high-cost pluggable optical transceivers, what we're really doing in its simplest form is we're delivering our integrated silicon photonics platform by effectively pulling those 32 optical transceivers directly onto the switch substrate. To reiterate, through integration, we're miniaturizing the system and eliminating the high-cost PCBA, retimers and pluggable optics by adding this -- by adding all of this capability directly onto the same substrate as a switch. You can see this in the image to the left, where we're showing a schematic of Tomahawk 4, Ram's 25.6-terabit switch, with 4 connectorized optical silicon photonic tiles mounted on the same substrate as a switch chip. On the right-hand side, we're showing a system-level schematic of a Tomahawk 5 system with full optical integration, 8 co-packaged silicon photonic chiplets mounted onto an industry standard substrate. Both systems make use of the existing DSP or the SerDes in the switch chip to directly modulate the optical signals into and out of the chip, simplifying the data signal path and making the most effective use of the silicon in the system. This is the next step in the evolution of semiconductor I/O. Once again, Broadcom is uniquely positioned to forge this new frontier by tapping into our market-leading semiconductor capabilities and our current capabilities in high-throughput switching, advanced SerDes and DSP technologies and, of course, leveraging 30 years leadership in semiconductor and optical manufacturing and test. I can't think of anyone else who has all of these capabilities and the ability to provide a disruptive new platform. The results are absolutely compelling. By reducing our -- the power consumption of the system by 30% and improving rack density by 50% and, of course, offering a cost advantage of up to 40%, our customers will begin to view the interconnect not as a pain point but rather as an enabler to facilitate new network architectures and continued growth in their services. So moving on to Slide 30. You'll see how we are introducing our silicon photonics in a phased approach, integrating with a multitude of our products and expanding Broadcom's semiconductor solutions franchise. Our first product is called Thebit. And it integrates silicon photonics with our 800-gig PHY and is sampling later this year. In 2022, we'll be introducing our first optically integrated switch, Humboldt, co-packaging the Tomahawk 4 with optical I/O and followed soon thereafter by Bailly, our 51T solution. Over the longer term, though, we really view silicon photonics as an expansion of our semiconductor platform and envision an optical ASIC capability. By integrating both with our own devices and in-custom designs, we can enable the scale-up and scale-out compute architectures alike, co-packaging to expand our capabilities into different markets. So in conclusion, I'll summarize the key points that we shared today. As the growth of data and pervasive cloud services continue, so does the increasing demand for data movement and fabric connectivity. Breakthroughs are required in optical connectivity technology today to keep up with the explosive demands on bandwidth and minimize the massive power footprint that's drawn from the fabric. Today, we introduced Broadcom's purpose-built silicon photonics platform, the next generation of semiconductor I/O. This, together with Broadcom's unmatched cadence in delivering leadership switch and routing silicon, utilizing our SerDes and DSP leadership, our mixed-signal technologies and our decades of leadership in lasers and mass volume semiconductor and optical manufacturing capabilities, we're uniquely positioned to deliver this disruptive new capability at scale. Finally, what does this all mean for Broadcom? We believe this means -- equates to approximately a $3 billion SAM expansion through the introduction of integrated silicon photonics on our networking platform. And with that, I'll turn it back over to Hock.

Thank you, Alexis. Well, I hope that you leave today's presentation with a much better set understanding of why we are all so excited about the opportunities in networking. Beyond that, I further hope you now have a better sense of how Broadcom executes in its core franchises, how we compete and invest to sustain and even grow businesses we already excel. With that, let me turn it back to Harlan for and open up the floor for questions.

Yes. Thank you, Hock, and thank you, Ram and Alexis. That was just a great set of presentations and insights and gives us a clear understanding of why Broadcom is the category killer in the networking business. Let me start off with just a couple of high-level questions first for Hock, and then we'll dive into questions for Ram and for Alexis. But a question for Hock. So on the earnings call in early December, you talked about strong bookings and record backlog of $14 billion. Based on what we're hearing from some of your peers, from some of the companies that have presented today at our conference, it feels like the demand environment looking into the first half of this year continues to look strong and broad-based. Is the Broadcom team continuing to see strong bookings? And in what areas are you seeing the sustained strength?

That's a very direct question. So let me answer it simply by saying, yes, we continue to see strong demand, strong bookings for our products as a whole. It doesn't apply across every single product category that we have, however. We live in a very interesting space because in terms when -- 6 months ago or even 3 months ago -- 6 months ago, for sure, we were seeing a ramp in addition to strength in networking and in broadband. And networking is obviously coming from hyper cloud and service providers. And broadband is obviously coming from telcos in the work from home environment. We have -- we further augment that strength through the ramp of our wireless business. Today, [ we're in ] -- to expand on that, the wireless business is very seasonal. So today, as we expand, there's decelerating bookings on our wireless business as it passes through its product launch and goes after the launch and decelerate bookings. However, we have seen continued, if not more, strength in these same areas of networking as well as broadband. And that has sustained the level of bookings we have. And you're right. You're correct. We have visibility all the way through to the middle of 2021 and possibly even beyond at this point.

Great. And maybe as a follow-up for Ram because Hock mentioned the sustaining or maybe increasing book of business on the networking side. The networking division was up 17% year-over-year in Q4. Tomahawk 3, Jericho2 has seen a strong upgrade cycle, driven by 200-gig and 400-gig optical with your cloud, hyperscale and service provider customers. We know that at least 2 of the Tier 1 hyperscalers have started the transition to 200 and 400 gig in 2019 and 2020, but we believe that basically all of the other Tier 1 and Tier 2 guys have yet to begin their upgrade cycle. So are they starting their upgrade cycle? Is this what is driving the strength in the networking business, Ram?

Yes. I would say the strength in the networking business is actually pretty broad-based, right? It's not just driven by particular upgrade cycle or specific to the megascale data centers. Our business is spread across enterprise, service provider as well as the megascale. Now within the enterprise, obviously, there is kind of the campus and there's a branch, and there's kind of the data center itself. So when we look at our overall business, obviously, there's the megascale in transitions to 400 gig and with it Tomahawk 4 shipping, we're also seeing kind of 800-gig platforms getting out there. But also just think about the 5G, right? When you think about 5G, a lot of people talk about the base stations and so on and so forth, but all of the traffic has to kind of come back to some kind of a mobile backhaul, right? There's a fronthaul and there's a mobile backhaul. And a lot of the mobile backhaul bearers are happening across the world, not just in the U.S. but across the world. And those mobile backhaul technologies in multiple things, things like PDN, there's IP RAN, there is FlexE and so on and so forth. So all of that is also kind of growing. Because even if you're connecting on a DSL to the house, eventually, you have to drop that traffic even on to some kind of a wired network in the back. So we play there, too. And then last but definitely not the least, when you think about the enterprise networking, a lot of the vendors, the partners who use us are continuing to grow, and they're gaining share, so which is also kind of tailwinds for us. So for us -- so it's kind of a combination of multiple things. the megascales, the transitions to 400, 800 gig, partly 5G, backhaul build-outs, generally wired infrastructure build-out and then our enterprise partners gaining share.

Maybe -- and I appreciate that. Maybe if we can step back for a second. Maybe if you could just give us a, Ram, just give us a snapshot of your networking business. Can you just break out rough -- in rough numbers, what percentage of your networking franchise is cloud and hyperscale versus service provider versus enterprise?

So Harlan, we have not done that in the past, but there's always the first time to everything, and I guess, you're very special. So to give you some perspective, I would say the cloud, hyper cloud's probably 25% or less of our overall business. I would say the service provider, kind of the telco, is going to be in that 35% to 40% range. And the rest of it is in the enterprise. Now it might not feel very intuitive to you, right? You might say, "Oh, wow, enterprises are really kind of that big and -- or what was I thinking?" But you kind of have to think about enterprise very broadly because in the enterprise, you have the campus, you have the branch, you have the data centers. And then you think about this enterprise very globally, right? There's a lot of global players in this enterprise market, and they are growing and they're growing fast. But yes, so hopefully, that allows you to connect some dots here: 25% or less; 35% to 40%, telco service provider; remaining, enterprise.

Great. And I have an investor question here for you, Ram. So why can't competitors implement the same "purpose-built strategy" on the switching and routing franchise and try to squeeze the profit pool? What is Broadcom doing that is helping them to extend or maintain your competitive advantage and 1- or 2-step lead over competitors?

Yes. So valid question, right? But this competition should have started 10 years ago if they want to be where we are today on this purpose-built strategy. And then two, being able to execute from one generation to another successively, right, and then be able to do it successfully is a huge investment. Every one of these chips is a $50 million to $100 million investment and to be able to maintain that cadence of investment on a technology platform, obviously, for me, some of that is subsidized in the sense that I get to leverage this Broadcom's technology underlying platform that is supporting a very broad set of silicon that goes across the different segments of that industry. So I would come back and say, so far, you actually really haven't heard anybody announce that they have this purpose-built strategy, and this is something that we've made a decision almost 9-plus years ago. And now then you have to develop those product lines or building those. That means it's going to be at least 24 to 36 months before they can actually execute on it. And then once they do that, they have to keep revving those chips every 18 to 24 months. That means they need that scale of investment and be able to get it right. And I think so when you think about it, the moat is one way to think about it. This moat is very large, very wide and very high.

Got it. Appreciate that. And then question for Alexis, investor question. Can you just talk about your views on co-packaged switch silicon with optics as it relates to some of the competitive initiatives out there, i.e., Intel, Barefoot are working on it? Compare and contrast your approach relative to the Intel, Barefoot initiative on co-packaged optics.

Sure. Well, whereas I won't comment directly on a competitor's approach, I can say that the industry truly understands the need for breakthrough capabilities in this space. And there's a lot of investment here. That said, our approach is based on industry-standard silicon, Ram's merchant silicon, and we're putting together purpose-built silicon photonics solution that's probably the most tightly integrated solution on the market. I think I mentioned about a 1 picojoule per bit electrical interface. That's probably almost 10x better than the current systems that are out there, where one has to rely on additional retimers and PHYs throughout the system itself. So I think that Broadcom, because we're viewing this as an extension of the switch platform, has really had the opportunity to define it at a system level, taking into account every piece of the enterprise system. And it's a gnarly engineering problem. But I think what we've been able to come out with here is the cleanest and tightest integration with the network silicon that's possible. So I do think that there will be a healthy and vibrant ecosystem that forms up around this, and we welcome that.

And then a follow-up -- and another question for you is how should we think about the drivers of the $3 billion photonic SAM that you laid out in terms of units, dollar, content, time frame? Anything that you can maybe a little bit quantify in a little bit more detail for us on this SAM opportunity?

Yes. I mean I think that one of the fastest-growing areas in the data center infrastructure is that which is spent on the network. So we think that the data center CapEx, in general, is growing at greater than a 15% compound annual growth rate. The spend on networking within that is growing from 12% to 15%. So we're seeing 2 pretty strong growth CAGRs. Now when we look at this $3 billion SAM expansion, it's really targeting specifically -- I would say, when we bring this out in production, let's say, 2022, we're able to capture a portion of this market that's currently today served by optical modules. So you could think of this $3 billion SAM expansion as Broadcom entering into the optical modules, not into the optical module space, but we're not in that space. We've exited that space purposefully, and we're driving this from an integrated silicon photonics solution. That's where the $3 billion SAM expansion comes from.

Got it. And then a question for Ram. The team recently started production shipments of Tomahawk 4, as you mentioned, enables both 400-gig and next-gen 800-gig optical connectivity, doubles the switching throughput relative to Tomahawk 3. You've established the 2-year cadence of doubling the switching throughput. So should we anticipate Broadcom introducing its Tomahawk 5 platform supporting 51 terabit per second throughput and 800-gig optical connectivity in the kind of 2022 time frame?

One thing I would say is I've made a compromise before breaking out the segments for you. Now it's my products. I will let history and the trajectory let -- be the guide for somebody, but I will not break that one here, Harlan.

Okay. No problem.

But your guess could be as good as anybody's just based on our history.

Well, the history has been pretty accurate. So it's been every 2 years. So I would assume that, that would be the case. But I appreciate your answer there.

Because -- by the way, one of the thing that happens is if I say yes to this, right, when I go to launch the product, my marketing and the PR team will say there's nothing to launch here because you've already launched it. So I'll leave it at that. Please go ahead.

Okay. Great. And then, Hock, a question back to you in answer to the strong booking trends. Let's see. Are you still requiring customers to place orders with 6 months or longer lead times? And in addition to leading-edge capacity, are you starting to see constraints on lagging-edge 28- and 40-nanometer wafers as well? And when do you anticipate these supply constraints to ease?

To put it in perspective, we're one of the largest buyers of silicon wafers in the world today. We're one among the largest. So we use across a fairly large wide spectrum leading edge to some legacy edge. And that constraint has not changed dramatically for the last 4, 5 months, even 6 months, as we had last -- as we indicated in the last 2 or 3 earnings call. And it continues to be very constrained.

Got it. Okay. And then question for Ram. The whole idea of disaggregation or separating the hardware from the software or the white-box model, as we typically tend to call it, obviously adopted by the cloud guys, Broadcom saw this coming many years ago. Recently, as you mentioned, service provider customers are starting to adopt the disaggregation model with the strong adoption of Jericho2 by AT&T with that initiative. But in enterprise, the trend is still focused on OEM-specific silicon hardware, software-integrated networking platforms. With the strong enterprise customer relationships that you have, let's say, via the infrastructure software franchise, is the Broadcom team also trying to enable, motivate these large enterprises to move to a more disaggregated networking silicon model, and therefore, maybe drive more adoption of your merchant networking silicon into these customers?

Yes. So what I'd say is the simplification of the system that we are driving is of compelling value proposition in customers across all the segments, right, the enterprise service provider and the data center because it really leads to lower power, simpler software, high -- much higher availability. So when we think about merchant silicon adoption in the enterprise, we don't necessarily have to drive it through the notion of disaggregation. Just all the other OEM partners are using merchant silicon if they are able to win and grow their share by virtue of the -- basically, a much more efficient platform, that is sufficient enough for us. So we don't necessarily go and say, everybody has to disaggregate, but we actually talk to all the customers about the benefits of being on merchant silicon, independent of which form factor it takes, whether it takes a white-box model or whether it takes an OEM-integrated model. But yes, I mean, we continue to grow well into the enterprise space. And as I've mentioned to you before, almost 35% to 40% of our business is in the enterprise segment, and it continues to grow.

Question for Alexis. How do you see what -- so the question for Alexis is what happens to the optical transceiver market? With these co-packaged fully integrated solutions, it looks like you have some products that are actually going to start to demo or come out 2022 through the 2025 time frame. What happens to the optical transceiver market? Does this totally just cannibalize this entire optical module market over the next 5 years?

I think there's a strong trend towards integration in terms of overarching solutions. But there are a vast number of different optical segments that will still be serviced by pluggable modules. What we're focusing on is the high-density switch integration and switch interconnects today which can deploy across Tomahawk, Jericho. It could deploy across any of these areas. So I think it's a tremendous opportunity. But I think that for the next few years, the solutions will coexist as the ecosystem develops. That said, it's a tremendous opportunity and a tremendous value proposition that's going to be hard to turn down.

Harlan, if I could interject, too. You realize, right, from U.S. manufacturing and U.S. industry point of view, there's hardly any optical transceiver market -- business to talk about. This all built, assemble and built in very low-cost countries where they are much more efficient in building it, particularly in places like China or Vietnam, Taiwan, to some extent, but largely China, where, in fact, some of the raw, like hardwares, sheet metal, are even cheaper. There's no business to compete in optical transceiver, except the engines to drive those transceivers, the lasers, some of the key components like drivers, TIAs, that may exist. But as a full assemble, there isn't.

Got it. And I appreciate the insights. Question for the team is -- maybe for Ram is and for Hock is, how do you guys manage cost per chip given that 800-millimeter squared? I mean that's probably pushing the limits of the reticle sizing. So we all know that cost per transistor is going up. And so does that mean that in order for you to continue to drive margin expansion, that prices have to go up on a per gigabit of throughput basis, that price has to go up every single generation of Tomahawk or Jericho on a go-forward basis?

Hock, you want to take that?

Ram, you're welcome to take that question.

Okay. Yes. The slide I showed that actually talks about when you double the bandwidth of a device, right? You go from 6 devices to 1 in all that interconnected runs between all of these devices, whether sometimes at interconnect, there's a copper interconnect, either on a PCB or in a cable or sometimes that interconnect is actually an optical interconnected at a port level, all of that kind of gets collapsed from 6 devices to 1. So on a cost per gigabit basis -- when I say cost, I'm talking about our selling price on a gigabit basis.

Yes. Yes.

Overall, will be lower, right? Cost per gigabit will be lower, but the ASP of the device absolutely will be higher. And when they actually think about all the second order bits that are around it, the interconnect, the OpEx and the system costs, when they are not actually integrated, all of that is something that actually benefits us and benefits the customers in terms of moving to the next generation.

Question for Hock. Would you still consider M&A in semiconductors if it made strategic and financial sense? Or have you just completely shut the door on looking at potential semiconductor M&A in favor of only software targets?

Oh, of course, we are just as open. I mean we are probably one of the broadest-based largest semiconductor and for SoC, probably one of the largest SoC company out there, if not one of the largest. So to us, we continue to see semiconductor as a very interesting deep profit opportunity and profit pool. The thing is, though, as I mentioned, we have 3 very distinct, clear criteria any acquisition have to meet. And as long as they meet it, we're very open. We haven't shut down anything.

Got it. A question for Ram. In large, fast-growing markets, it's inevitable that you're going to see more competitors. And this goes back maybe to the prior question, trying to break into the switching and routing silicon market. Over the years, we've seen many try and fail to do battle with Broadcom, Cavium, Barefoot, Intel, Fulcrum to name few. You have a few new up and comers like Innovium, MediaTek, Cisco's new merchant one silicon products. Do you see any of these competitors gaining traction? And I think you already talked about this, but the question is what are the barriers to entry the team has built to maintain its strong share?

Yes. So the barriers to entry in terms of being able to maintain share, first, is execution. There's no alternative to execution. Look, I think before Broadcom, I've been at other places and so on and so forth, I've been in the networking industry for a very long period of time. When an incumbent loses position, it's primarily because they haven't executed. There's always an interest in the marketplace for an alternative to an incumbent. But as long as the incumbent actually executes, they generally maintain their momentum. And from our standpoint, I mean, we kind of actually think about this whole market, right? The systems business is over a $50 billion market, $50-plus billion in terms of what we serve. Today, the entirety of all of this discussion we've had at a silicon level is a fraction of that, okay? Then you have to intuitively ask yourself, if the silicon is the system, who's capturing the value and why? And the guy who's capturing the value, will he be the fox guarding the hen house if you try to get into the silicon business, okay? So the reason that is important is you got to think about this, say, this is a business of scale, it requires a lot of investment. And the beauty of it is the market that we play in is in a market we are enabling, there's a huge difference in terms of the total addressable market. So I keep telling people. People might say, "Hey, you are an ugly fish, and you are a big fish in a small pond." That's our problem, right? And so really, what you'd like to do is kind of change the pond that you're playing in and keep executing. And we think, in terms of everything that I've talked about before, purpose-built strategy that we've built many years ago, we've already executed on to it multi-generationally. And more importantly, when we executed on it, it's not always about a few years later, I'm going to try to do something else, something else. If you just look at some -- all the announcements that these guys have made over a period of time, every 3 years there's a new announcement, a new architecture, something that they've shipped before, shipped before does not work. None of that. It's continuous execution, continuous innovation on separate product lines. And we built this business around the economics of the silicon market. We've not built this business around trying to essentially subsidize it through some other business. And somewhere in between, I kind of get confused on what my business model. None of that, right? So no, look, we always take competition seriously and if we eventually lose market share, shame on us because we haven't executed or we haven't thought about the art of what's possible. But I can tell you, in the last 8, 9 years, we've seen multiple of them. They've all come and said that they have done -- they will do a better job. And I think we have shown that we are able to stay at the leading edge of doing a good job. And as I've said before, the Broadcom technology platform is a very, very important enabler that you cannot discount.

Yes. Absolutely. We've seen just great, solid execution from the team, especially on the networking franchise execution. We're just about out of time. So I'm so pleased that we could be the first to host this analyst day with Broadcom and looking forward to the next session, maybe as a participant. So Hock, Ram, Alexis, thank you so much for the great insights today. It's clear why Broadcom is the leader in the networking market and will continue to benefit from the strong demand in cloud activity bandwidth going forward. So thank you. Thank you very much for the opportunity to get some insights into the networking franchise.

Thank you, Harlan. Thanks for hosting. We're glad to be here.

Thank you.