Semtech Corporation (SMTC) Earnings Call Transcript & Summary

Welcome to Semtech's Data Center Analyst Day 2026. Please welcome to the stage Semtech Senior Vice President, Investor Relations, Mitch Haws.

Good morning, everyone, and welcome to Semtech's Data Center Analyst Day. We have a great program in store for you today. A lot of discussion today around where data center fits within our portfolio and how we're very well positioned for the future. We'll also talk about how that translates into our long-term financial goals. Looking at the agenda, we'll start with -- sorry, Hong will kick us off with a discussion around the transformation of Semtech, where we've been on that journey and where we go from here. Following Hong's remarks, we have the pleasure of the various team panel to talk about the future of data center networking. So we expect a lively discussion. Following that, Imran Sherazi, our General Manager of Single Integrity will go into detail on the data center road map. Importantly, he'll talk about the legacy of our data center business and how that positions us for a very bright future. During lunch, if you see at the back, in addition to lunch, you'll have a chance to look at all the demos. Most of what we talk about today in terms of products will be featured in those demos. So a good opportunity to see the technology in action. Mark will begin at 1:00 with a discussion around the financial results for Q3. He'll talk about our long-term financial model and how that looks through the end of fiscal year 30. Following that, we'll open the floor to questions. The senior leadership team for Semtech will all be here. So a good chance to dig deep. A couple of housekeeping items. We will make forward-looking statements today. Those statements carry certain risks and uncertainties. The risks associated with that are available today on the website and as well as our SEC filings. In addition, during the course of the discussion, we will talk about non-GAAP financial measures. There are reconciliations for those are on our website at investor.semtech.com. With that, thank you for joining us. Let's get started. [Presentation]

Please welcome to the stage Semtech President and CEO, Dr. Hong Hou.

Good morning, everyone. Welcome to the Data Center Analyst Day. For those of you who come over from the OCP welcome. It's very fitting for us to have the data center Analyst Day alongside with a great event of the OCP Global Summit. That's where the innovation was so case. That's the direction we are heading to. Having the CEO for Semtech slightly over 2 years. And during this time, you often ask us to hold an Analyst Day, our T&D to talk about where we differentiate and what are the acronyms of different innovations and technology. Here we go. The reason we waited for 2 years is because we were busy in transformation. But right now, I will say -- and hopefully, you see the proof that Semtech is a different company. When I visit different sites, our employees often tell me I have been working in this company for 15, 20 years, but I feel it's different very -- a lot of energy over the last 2 years, our customers are telling us how did we get to this point? Today, I wanted to walk you through our journey of Semtech Rising, Semtech transforming and Semtech [ exceling ]. So let's get started. Its inflection point is happening now, and we are fundamentally changing the company's profile. With a sharper portfolio and after the divestiture announcement, of a wireless module business that allow us to double down to have sharpened the focus on the core portfolio, and we have stronger growth momentum. The AI data center spending is increasing and the networking connectivity is bottleneck, and that is the area we are focusing on that is the area we're doing well. And with the increased demand, we got tremendous booking momentum. We got record level backlog. That's given us visibility for many years in the future. So it's the right time for us to provide a multiyear financial guidance. And the early proof shows that we are on track to build a predictable high-growth and high-margin business. I know I met many of you, if not all of you, over the last couple of years, but I still like to spend a few words on myself, I'm still as excited as leading this company at my day 1. I feel very privileged to lead the next phase of Semtech transformation and growth. I'm technical, had a technical degree in electrical engineering and probably one of the few CEOs has worked in 3 national labs in 3 countries: China, Japan and U.S. and also conducted research on photonics and communication at AT&T Bell laboratory in my early days. My industrial experience spans over communication semiconductor and data center. I've been moving around quite a bit, but that background really prepared me I feel like for this job at Semtech. The 3 products may experience the most relevant to what I do at Semtech. Emcore, at the time I was the CEO, started as Director of Technology, and we let the high-power laser development and commercialization with supporting high-power laser support probably 80% of the broadband infrastructure for cable TV. You probably don't know that. But that's why when we had the acquisition, we know exactly what we're buying. The second part, when I was working at Intel to lead the cloud networking group, that's where I get other connections and get the familiarity on the data center requirement. Before I step in as the CEO, I was the independent director for a year, that prepared me to get to know the challenges and opportunities. All of this experience translate into the time. I was able to hit the ground and running without any learning curve without a trial and error. So I feel fortunate to come to a right place at the right time, bring the right skill set. Now a little bit more about Semtech. We got a long history. We were founded in 1960. We became a public company in 1967. Just to calibrate yourself, that's 1 year prior to [ Intel ] with funding. So when I engage with investors very often, I heard, "Oh, I know Semtech, my ask is which part of the history do you know about Semtech? " We started defense rooted pivoted to analog and mixed signal and the acquisition in 2012 of [indiscernible] really bringing the market-leading asset. So that's our crown jewel. We did a little bit detour in 2023 of the acquisition of Sierra Wireless, at the wrong time, which stretched our balance sheet and diluted our focus. I'm not going to dwell on that, but I don't want you to pretend that never happened. So I joined after that, and right at a time, the industry it's investing in AI infrastructure tremendously. We analyze our portfolio. We feel we have the right to win, and we are embarking a journey of transformation for Semtech. Very established company, but it can't have the renewed vision and focus. So today, we aspire to be the leading connectivity company of the AI area, focused on moving more data with less energy from the heart of the data center to the edge of the physical world. We calibrate our strategy, our action against this vision statement. So we embarked our journey called Semtech Rising and 3 arcs and one story to transform the company. Semtech rising to fix the fundamentals, we started that second week as I was the CEO, and we can be clear now mission accomplished for that phase. We are squarely in the Semtech transforming that is about to build more growth engines to accelerate our business growth. And our destiny, our goal is to be the leader of the industry in what we do, that Semtech exciting. Again, there are 3 phases to support that vision. Let me talk about the Semtech rising. As I mentioned about, we embarked this journey second week after I became CEO we didn't have time on luxury to do a long history study, but we know what stress does. We know what need to be done. Our positioning was not clear, and we were a mixed signal analog diverse company, that's not enough to get us focused on achieving our mission, achieving the alignment with the market opportunities. Balance sheet was stretched. To say the least, we had a leverage ratio of 10.5x at the time. And that wouldn't allow us to do anything more aggressive and then our employees, morale was low, and the engagement with the customers was lacking. So what we did, the strategic positioning is the delineated core and noncore and put omitted resources in the core area in R&D. And even for the core areas, we did a thorough review that we only prioritize the project that matters to the customers the most. And in the balance sheet strengthening, we did a number of transactions, and we brought the leverage ratio to slightly over one at this point. And we also prioritize our capital priorities in R&D and be really nimble in G&A, and we were able to go on [indiscernible] as we improve our balance sheet and financial capabilities. On the customer engagement, I'm going to have a slide to talk about that in more detail. But on the employee side, we communicate often and communicated at our levels through our hands meetings through monthly newsletter through site with it we reestablished trust and transparency. And I am very proud today, we're going to be giving out many numbers. This number I'm most proud of, our engagement score from 2 years ago, I will hand it over with a scorecard of 57% engagement to today, the 78%. That is really things matters the most. If we did not have the employees aligned and to share the clarity and the vision, we wouldn't be here today. So with that, we were able to turn the highly leveraged company into a focused and aggressive data center pursuit. So that to double-click on one of the fundamental culture change, it's a customer engagement process. We provide I see the key components now including photonics to the module manufacturers and then they, in turn, to supply the integrated product to hyperscalers and XPU companies. In the past, we basically engage with the module manufacturers on this chart. And we do not really engage much with the define, the definitional customer who designed the data center who defined their architecture. Now it's different. We engaged all levels. By leveraging our -- the analog and mixed signal expertise and that is a track record and experience, we are mile increasingly more by the hyperscalers and a technology provider as a technology partner. We can help to understand their pain points and by applying our differentiated technical analysis from simulation to purpose-built silicon to provide differentiated solutions. One, we provide that what is doable and they can define their architecture and our solution by doing this engagement is almost automatically designed in. Of course, we engage with our direct customers very closely. This help us to accelerate the design in process help us to translate into accelerated growth. As I said, we got early proof today, we are invited to the room in the early architecture design phase by the hyperscalers, by the XPU companies. That's really a great place to be, and we are increasingly getting trusted to be the solution provider. Semtech transforming. That's where we're squarely in this phase. The aim is to build more growth engines. Now we have the financial capability. We want to double down and intensify the R&D investment. Just to give you a ballpark number. This year, we are projecting to increase our R&D spending by 30% sequentially. In the core area, we probably even more heavily indexed on the R&D. Why is that? To build more growth engines. Many -- most of the projects, I would say, stemmed from the alignment with the customers and trying to solve their most pricing issues not by the scientific curiosity. So we believe that the investment will inevitably develop a new product and generating new business opportunities the feedback allow us to be able to invest more. So we are really in this good versus cycle to accelerate business growth. We're in a really good place. And on the portfolio optimization, we did the first divestiture of noncore assets that allow us to deliver predictable margin profile we're constantly looking at our portfolio, but we wanted to build more growth engines, data centers today, the largest one and we have the LoRa and we believe we can have more growth engine based on the current portfolio. So portfolio optimization is an ongoing journey as we increase our -- expand our portfolio and expand our ambition, we will identify more gaps in capabilities and [indiscernible]. M&A focus is not to increase the scale of buying revenue but more enhancing capability and applied the leverage. So we have some ideas, and we constantly review that, that's a journey, strengthening the winning culture. Our employees get an early test of winning that is great and winning is indicative winning [indiscernible] and winning is gratifying. Winning is our largest driver behind our employees. So you're going to have a chance to meet our application engineers when we're in the break and roam around to see our demos. But we outlined in our strategic imperative growth oriented customer intimacy, innovation with differentiating solutions, do the best in whatever we do in whatever area we are in, and employee engagement. So we'll continue to strengthen that culture, and we are well underway for transformation of the company. To show a proof point that we are positioned at inflexion and Mark is going to be talking a lot more detail on our financial projections. But here, I would just wanted to show the historic 3 bars at the trend chart. It took us a year to grow the top line about high -- our mid-single -- mid-teens from the Q3 '25 to Q3 '26. But we're going to be having over 50% of growth from Q3 '26 to Q3 '27, that from the updated guidance we're going to be giving. So our growth is accelerating. Margin profile after divestiture of the wireless module asset as expected in to close in Q4. Our gross margin baseline is going to be 64%, with tremendous tailwind of the mix shift into the high-margin business, and that is a floor, not a ceiling. Our earning is accelerating even more so because of the operational leverage when you have a larger scale without [indiscernible] G&A, and that would translate into more earnings per share. So we see the early proof that the inflection is happening. It is not just a projection. This is the data. This is the fact. Here, I said we are -- today, it's not a full long Analyst Day in reviewing our product and portfolios but more data center centric is fitting to co-host with the OCP event, but also showing that in FY '24, our data center revenue only account for about 7% of total revenue. This quarter, we expect the data center revenue account for about 40% of total revenue. It's not because the other part of the portfolio is shrinking. It's just because the data center is -- the growth is accelerating. We'll show you that the FY '30 forecast later. Mark is going to present that. And in that, we're going to be having the data center revenue to be about 2/3 of total revenue with a much bigger number of the total consolidated revenue. So we are still in the early innings of this growth, even though this market inflection points and the portfolio optimization continues, that's helped to bring the margin profile and also get a better bottom line. And our strategy is to become a leader in the data center. We may not be brought, but we want to be the best in everything we do in every product we offer. They can be copper, it can be fiber, but the theme is to provide high bandwidth, low-power solutions. We all see this chart this from BofA, recently and also this trunkchart, is different maybe after this week. It's going to be revising a little bit maybe up or maybe down but you can't argue there is no -- any of this level of investment is unprecedented, and we are sitting in the front and center of this in data center the gating items used to be compute capabilities. Today, when we engage with the customers, we were often told you, the networking connectivity, it's a bottleneck. We have the opportunity to unleash almost as much as 2x of the compute power. And what the customer wants, it's a high bandwidth low power, low latency and high bandwidth density solutions. That's exactly what we do. So that's what we play in data center. We don't do cooling or power delivery. We don't do compute. We don't do memory either, but we're just laser focused on providing the best network connectivity solutions. So talking about connectivity, you heard the terminology many times, the different connectivity, but they all share the same requirement again, high bandwidth, lower power consumption and high bandwidth density and low latency, starting from scale across. This is the so-called in the past, telecom, metro, but now it's proliferating into more -- on the same campus between different buildings. The reach is 2 kilometers and above. Through [indiscernible] acquisition later on, I will discuss in more detail, we offer the industry-leading game chips to be integrated in tunable lasers and to be the high-end laser for metro and our scale across applications. Our laser is ready for coherent light application as it's now, and we are engaging with multiple customers on that. So we have some play in scale across that's only going to expand what is most of the ramp we're experiencing is a scale out. So that is for interconnecting the Ethernet switch fabric in data center. And our solutions is incorporated in the fully read time, optical pluggable modules, LRO, LPO, AOC, you name it. A lot of acronyms and Imran is going to be doing justice in his presentation to explain in more detail and lift up the hood and show what we have in there, what's our content in there. We're supporting 800 gig. We're ramping 1.60 with designing 3.2T. We provide multi-generational solutions co-designing that with our customers. In front of the center, that is our most of the activity at this point. Good thing is volume compared to scale across estimated. You tell me a 15, 20, 25x of the volume is changing every day. Scale-in, scale up first, that is even higher volume, 10x more. Scale up is to interconnect the servers, the switches and at the right level and cross regs within a cluster. The scale-up we provide solutions both copper and fiber, copper through a redriver or linear equalizer and when it's integrated in a cable [indiscernible] is a great backplane interconnect solution, and that business is ramping with key hyperscalers. There's a lot of design activities of linear equalizer integrated on the Board because later on, when Imran presented, you will see that signal integrity is going to be compromised when you transmit high data rate across the PCBs, the most elegant way to solve that problem is integrate re drivers, some use retimers a little bit penalty on power side, but you'll be able to clean up the signal and allow the signal to reach the other end without a distortion. Use copper when you can, use fiber when you need to. That has been our guiding principle and validated by [indiscernible] in the GTC meeting. Otherwise, many of the analysts was asking me for a terminal value of copper, Don't put the terminal value yet. Copper is going to be playing in 224 gig recs. Copper is going to be even more playing in the 448 gig regs. I mentioned about a 3.2T design activities that's replicable. The [indiscernible] on board is tremendous activities in there. They see this as increasingly a very viable solution. Then MPO, the CPO, the optical interposers, I assume those are really the key demos got to get a chance to walk on the floor yesterday for a couple of hours at OCP. I probably already counted 15, 20 different demos of the NPOs. Our solution is there. If not there today, we're tweaking the form factor and it's going to be incorporated tomorrow. The reason for that is current NPL design is 200 gig per lane. That is an IP that performance is already validated through the pluggable tranceivers, NPO in a way, is LP steroid is a highly packed form factor of LPO performance is already validated, and we are in the front end center in providing solutions in there. So that is a market moving toward us. Market is moving toward analog, and we do nothing but analog solutions at this point. Scale in a lot of discussions and people really more forward-looking. And when you run into the connectivity wall and power wall and you can use interconnect between different chips with the chiplets or optical interposer. So we are closely in tune with our customers on the different solutions. So that is a connectivity topology. And as the data rate increases, just by providing the best PMD of physical media devices, TIA and driver, it's not sufficient because people will give you the spec with the broadband. We don't have a lot of performance margin to give out. So that's why we need to expand our portfolio to be able to have optical components alongside with our electronic components. This co optimization will allow us to provide a best chipset solution to the customers with reference design. Fitting in a theme of building additional engine, we did a couple of acquisitions of photonic assets. In March, we announced our acquisition of HIEFO is a company going through different recognitions, but rooted with Caltech, Ortel, Lucent in 2000, spent $3 billion by the company and Emcore when I was there, and we bought this company from Lucent and subsequently, we build high-power lasers. And we leveraged 46 years of the heritage, and we know that we are leaders in the optical component in providing the best our conversion efficiency the design allow us to provide the best beam profile allowed to low loss coupling into the fiber, our waveguide and also the noise figures. So this is really a great asset and grid extension. We have vertically integrated operating the fab in L.A. area from MOCVD epi to fab processing to test and chip-on-carrier packaging, we can utilize the entire ecosystem to increase our capacity. Speaking of capacity, today, we get more demand than we have the capacity. So we are in the process of increasing capacity. We project to increase capacity by 3, 4x by the end of this year and to serve the customers later on in different applications. One other way to increase our capacity is that we did another acquisition of photonic asset in June of this year, D-Tech. So we brought on board the world-class designers for high-speed photodiode. And the acquisition also allow us to have the immediate access to 7,000 square feet clean room facility. So we got preown the equipment. We're moving into this space. That's how we increase the capacity by 3, 4 [indiscernible] end of this year. There is a little different facility, but within driving distance of 10, 12 miles. And we are also focusing on design photodiodes and project to provide samples and demo in the upcoming OFC. So our photonics portfolio right now is still limited, but we are not planning to build everything for everyone. We want it to be answer ourselves on the strength and really proliferate strategically. For example, on laser side, I mentioned about our game chip is already class-leading provide the highest power in the tunable lasers. We have a solid customer base with now HIEFO inside of Semtech for 6 months, we are able to increase our customer base pretty significantly. As I mentioned already, our lasers in the same design already provide low noise, high-power and narrow [indiscernible] which is our ideal for [indiscernible] and light optical transceivers. We're engaging with multiple module manufacturers and then having our samples in their evaluation and qualification queue. Some of the early results indicating that we're better than any incumbent lasers they are using right now. The large market opportunity, the industry is still experiencing a lot of shortages is for IMDD, intensity modulated direct drive applications. You're hearing the laser shortages. Of course, we got a lot of interest in that area. But we are -- we don't have a lot of capacity to serve them. Instead, we use our differentiation. We use very high power lasers delivering power, the 200 to 250-milliwatt power CW lasers, which is different from, say, 30-milliwatt CW lasers, you have to use one laser per channel and lasers in optical transceivers. Now with our VHP CW laser, you can have one laser to drive 8 channel and silicon photonics modulators, but we have to engage selectively because we wanted to engage with the customers, they really value the U.S. made and for security -- for the supply chain surety. And we use this as a platform to design for a future and develop OCI, CW scale-up players working with the partners. We provide UHP 300, 400-milliwatt lasers with the semiconductor optical amplifier together with their Wavelength definition capability, we're able to provide DWDM lasers at high power for CPO scale up. Again, we're not trying to be everything for everyone. In Photonics, we are building strength outward and with the customer pulling us forward. The portfolio expansion can translate directly into the content increase in our business. So here, it's a chart. Probably you can take a picture, you can remember that I mentioned about the 800 gig, our content is single -- high single digit. When the industry evolving into 3.2, for example, we'll be able to monetizing more with the same component functionality with more complexity like a TIA and drivers plus adding more content like CW lasers and photodiode and even modulators. This is really very important because the most of the architecture shift in the future trend, increase our content. That's why we're saying the market is moving toward us rather than away from us. So right now, the content on 800 gig and through LPO increases through NPL increase even more, and we are getting ready for the components to be inserted in the 3.2 generation. Okay. About the market size. So we use 650 Group, and we'll have the founder and CEO of the 650 Group, Alan Weckel, is going to be on the panel later on. So we use their data, and they are most in tune for the networking side. If you look at the slides in the TO color, that's the size of PMD, the TIA and drivers, by the way, we adjust it only for our product, not the entire TAM. And of course, this does not include any additional opportunities that we're evaluating from M&A point of view. But that incorporate what's in -- on our road map for organic expansion opportunities. Lasers and game chips and SLAs, the addition of photonic components increase of SAM pretty significantly. I still feel that active copper have more opportunity than the market giving credit for because when you are going to a higher data rate, you just don't have other ways to get to the other end without redriver or retimer. Either way, so in 4 years, you see the SAM can increase by 5x. That's tremendous. That is a great opportunity. For us, we grew revenue through avenues. We can simply write with the wave and the size and trend will benefit us. That's number one. The number two, the content increase. I mentioned about 10x potential from the previous slide. Number three, we aspire to continue our journey like we have demonstrated on 800 gig TIA gain market share on future generations. So that gives us a more growth driver as well. So what give us the right to win? Why Semtech? I mentioned about a technical leadership and Imran is going to be talking more eloquently about that later on. I'm not going to dwell on that. We have the track record. We have the data rather than proposal, customer intimacy this engagement model works, and we'll continue to gain customer trust, and we want it to be also easy to work with the customers. This is a sentiment from the customer unprompt. It's not what I wanted to do, and we really engage with the customers without arrogance and we try to be problem solving. Operations excellence. So we have started to planning process with our customers on a monthly basis. Now we proliferate that number to over 20. That gives us an insight not only for what immediately they need and also what they're planning in the future. So that allows us to proactively design for capacity. And on Tuesday, we announced a multiyear strategic agreement with our great strategic partner, Tower Semiconductor. That agreement provides secure capacity to support our business growth in the horizon we are going to be providing a forecast. So we are also the hallmark of the operations for Semtech team is to provide operations excellence all the time. So those are the 3 pillars to support our vision. I will be doing a disservice if I don't mention about spend a minute to talk about the LoRa because that's our second growth driver. And we also did some realignment with the market on the LoRa strategy. We focus on silicon develop the best capability, lowest power, longest reach and higher bandwidth and provide market enablement. So now we have the dual band to provide enhanced bandwidth. We have the LoRa provide enhanced the capability on another protocol we have Amazon to address the mass market of consumer side through Sidewalk, we have been traditionally been very strong in providing robust connectivity solutions for medical devices the enablement through innovation forums through software development stack through the webinar and customer engagement to list the entire ecosystem. So they work very hard alongside with us, that allow us to proliferate the application verticals from traditional 3 to 4 to now 8, 9 and 10 or even more drilling. That is a great growth driver. Early on, I mentioned that we have planned to review the rest of our portfolio to continue to develop more differentiation, continue to develop better drivers. So what does that translate to? So Mark is going to be talking about this in a lot more detail. We're projecting by FY '30, our revenue is $3.2 billion, mostly driven by data center and LoRa, that translates into the consolidated growth rate of 38% compound. Just get it translated into the multiples CAGR translating to 9x growth in 4 years in revenue. 30% CAGR translate into nearly 3x of growth of lower revenue from -- in 4 years. And then consolidated is going to be about 3.5x. I wanted to provide a little bit more color on this projection and we have seen a lot of project in the hockey stick. But I wanted to stress that we heavily indexed on the near term because we have the visibility and the confidence and conviction and so the FY '27 in FY '28, later on, Mark is going to be providing, for example, data center side, we project over 100% year-over-year growth. The rest of the horizon we use the market CAGR. But as I said, we aspire to provide 50% higher growth rate than the market. So as for the bottom line, Mark is going to be discussing that in more detail. And now let me bring back our journey to our destination. Semtech rising, done. Semtech transforming, we're in progress is accelerating. Where do we want to bring the company to? Semtech excelling, wanted to be providing higher growth than the market, better profitability than the market to customers. We aspire to be their partner of their choice to employees. We want them to be energized, motivated into investors. We thank you for the trust you place in us and in our team and we want you to be handsomely rewarded. The last slide, the takeaway Semtech is fundamentally different companies today. I'm not going to go through every bullet point. And hopefully, you see the data, again, is not the aspiration and we are pivoting in this really -- we are inflecting and this is really in the early innings. Let me close on a personal note. When I started as the CEO a couple of years ago, I ask our employees and some of our investors in the audience, give me the benefit of doubt. I will not let you down. Hopefully, you see that the early proof we delivered some results. But to me, to my entire team to the Board, this is just the beginning. The best of Semtech is yet to come. Thank you very much. So next on the agenda is the industry channel -- panel, is the AI connectivity impairment -- imperatives. So I am -- it's my pleasure to introduce our moderator, Mark Edelstone in the company's trusted adviser and also the Chairman of Morgan Stanley Global Semiconductor banking. Mark?

Good morning, everybody. Never underestimate the importance of leadership. That is one of my fun things. I've been doing this for a very long time. And to see Semtech today is quite impressive, quite amazing. Hong in his earlier comments said that knowing Semtech is a question of what period you're talking about. I've been doing this since 1989. And so I can vividly remember my discussions with Jack Poe, who is the CEO here from 1985 to 2003. And the difference in this company today under Hong's leadership is impressive, amazing, and it's great to see. This company is in a great spot. It's an honor on to be able to do this for you today. We've got a great panel. I'll have them come in a minute, but we're going to be able in the next hour to provide, I think, a pretty good perspective on what's going on in the market. Many of these things you'll know. Other things I hope it will be a deeper dive for you as well. But Hong and his team have actually put together some real subject matter experts around this topic, starting with Alan Weckel. Alan is the Founder of 650 Group. He's basically been doing research for a couple of decades. He's been running this company since 2017. David Ramku is the Head of Infrastructure at Meta. So we'll get a pretty good perspective of what's going on with those folks that are spending lots of money on AI infra to deliver what's happening going forward in all the AI workloads. And then Alexis Bjorlin, who is deep technologist, such matter expert on this as well, really great background. Currently now the Chief Strategy Officer at General Catalyst which is a well-known VC investing in this sector. So I'd like then, of course, Hong as part of the panel as well. So I'd like the 3 of them to please come up and join me on stage. Okay. Thank you so much. So what I'd love to do, I gave just a brief background of the 3 of you, but it'd be helpful, I think, you're such domain experts in things, and we're going to cover things from the market technology all those kinds of things. So it might be helpful to just quickly give a little bit more detail on your background. What are the things that you've done that make you impactful to basically help share inputs to this audience. And then as you do that, also, the first question you could just do at the same time basically would be to just talk about the single biggest change you expect to see in data center networking here in the next 5 years. So maybe starting with Alan, and then we'll just go down the list, and we'll go from there.

Yes. And thanks so much for having me. So I've been in research for over 20 years. Prior to that, I did network design at Raytheon. So my entire career has been in networking and recently in data center. And we track over 40 different markets from compute storage networking to semiconductors. So we're kind of at the end of the day, unit counters when all is said and done out there. And the amount of change is just incredible, right? I think the way I think about this from a networking perspective is we're going through a great transformation in terms of speeds and feeds, but fundamentally, we're increasing the content as we go. So networking is kind of transforming twice, both from a higher speed, higher bandwidth perspective and then also significantly more content. And that ultimately drives the networking market from tens of billions to hundreds of billions and then eventually, right, to that kind of magical trillion market just for networking itself.

And the biggest change?

The biggest change is the pace of innovation, right? If we think about what's occurring in networking, we're going from 1 network to 4 networks in order to support AI that's something that historically we've never seen at an industry level before.

Okay. David?

David Ramku. I'm a part of a Meta infrastructure organization. My team is responsible for technical program management, product management, ecosystem and partnerships engagement. We also kind of lead the internal prioritization of our programs. we're part of the infrastructure team that's responsible for the physical hardware that we place in our data centers, and that ranges from service and racks that we design that are based on merchant silicon and every name you can think of is obviously represented there. We also spend a fair bit of effort on building our own silicon for multiple applications, including AI accelerators, network products, et cetera. Prior to joining Meta about 4.5 years ago, I spent 18 years at Intel. My last role there was leading the infrastructure processing unit business within the Intel Connectivity group.

And biggest change coming?

Biggest change coming. How does the saying go predictions are hard, especially about the future. Probably the network evolves from being the plumbing of the data center to being the enabler of compute architectures that unlock new possibilities for our AI applications.

Alexis?

Excellent. So thank you for having me here. So I currently serve as the Chief Strategy Officer for General Catalyst, but have a deep career spanning NVIDIA, meta Broadcom and Intel. And before that, some really excellent source photonics in the optics and networking space, my entire career. So sorry about that. So I think one of the most important elements that we'll see coming down the pipe for disruptive change, of course, is co-packaged optics. Optics closer and closer to the silicon, whether that manifests itself in co-packers optics, near package optics, what we're seeing is an unlock of bandwidth that's required to enable next-generation architectures, as David's saying. So I think that's the big disruptive change that's coming up. But when we think about the evolving landscape of compute and memory architectures, it's all still bottlenecked and constrained by the network, and the network becomes the enabler to high-performance heterogeneous compute systems in the future.

Fantastic. All right. So -- and everybody knows Hong already, and he's covered his materials. So obviously, there's lots of amazing things happening in the sector. I think it's pretty clear that what's happening in networking in the data center is going to be very impactful in terms of how the rest of this whole AI deployment is going to play out. And I think we all know we're very, very early here. I mean, ChatGPT was only 4 years ago as an example. So tip of the iceberg, in my opinion, despite what we're seeing in the marketplace today, and we'll talk maybe a little bit more about that, too. But just to set the stage a little bit. So Alan, you just said $1 trillion opportunity in networking over time. And we know that this year, the hyperscale is just spending trillion dollars plus or minus. Just give us your sense of how you think it breaks out on the networking piece between scale out, scale across scale up? And then what do you see as the biggest -- the fastest growers within that?

Yes. So scale-out is the largest domain size today. That's the first market that's going to cross that $100 billion threshold. In the longer term, scale-up will become the larger market therefore, is the fastest growing. But that does not take anything away from scale across, right? We're still in the early days of locking scale across, whether that means building to building in a very large campus or connecting these campuses together. This is something that historically the hyperscalers didn't have to worry about in the volumes that we have. But kind of agentic is unlocking that agents are by definition chatty and trying to engage with resources that sit in other locations. So as we move from training to inference, we really do get the unlock and scale across and that will have a kind of growth rate. Each of these is a very healthy market, and you kind of add them up, and that's where you get the significant TAM expansion. Fundamentally, networking is becoming a larger part of CapEx as time goes forward.

Okay. Fantastic. David, you're in a unique position here representing, obviously, what's happening in the players that are putting forth all the infrastructure and things not specific to Meta because I recognize that puts you in a sensitive spot. But just kind of give your perspective as someone who is trying to put in all this infrastructure, what are the sort of key things that you're focused on relative to seeing payback from the types of things that are required to basically make this all happen for the rest of the workloads that people are using.

So if you're asking in the context of how we think about kind of setting our budgets and kind of projecting out our infrastructure, obviously, we look at internal demand, and we look at how the demand plays out over the years. The -- obviously, demand is plentiful. So the next key factor becomes power availability. And that generally becomes the gating factor of how much infrastructure you can add. Obviously, we look at TCO for the various solutions that we're implementing. We also try to map out the demand shape on training versus inference, as our models get bigger, we need larger inference clusters. On cluster size, you're trading -- you're generally training time versus the cluster size. And the last, but obviously, it's become really important. Supply availability is really important. So as we look out in time, do we think that there is supply available for us to build to our ambitions. So to kind of recap, we think about power relative to our demand, we think of our demand shape and then we think of supply assurance and availability. There are some investments we make that happen well ahead of the actual NTI program that we deploy. So we know that we have to invest in capabilities before that actually intercept the program, optics being an example of that. We've invested in optics, and we've done Optics research for a pretty long period in anticipation that when a product comes that intercepts, it gets integrated with that program. Otherwise, the duration of those deployments would get much longer.

Great. So maybe following on that thread a little bit of excess. You've got amazing experience, you've worked for the biggest and baddest companies effectively in AI. Your most recent was GDX4. NVIDIA, but before that, Meta Broadcom, so you've seen a lot put yourselves in the shoes of someone who has to make these decisions around ROI. How do you see that today, recognizing that there's obviously a whole lot of front-loading that needs to take place, as David just talked about, how should investors think about the ROI on these investments today as we're thinking about the monetization of tokens and all the things that we're going to get out of AI?

Yes. Obviously, I think the world is trying to figure out how to think about this. But the current way that we look at this is the return on investment of deploying AI infrastructure really is looking at the monetization models from. Currently, we moved away from simple metrics of dollars per GP per hour, $2 dollars per token. And I think where the world will find itself is in dollars per outcome. So one of the most important things to focus on is AI diffusion into the enterprises into sovereign applications and understanding the outcomes that we can unlock with AI and those outcomes becoming value that we're able to attribute down to the investments. We're not there fully yet because we haven't been able to map out that full value chain. But what we are seeing is a tremendous rise in the use of AI in the use of tokens in the balance between the proprietary closed-source models and the advent of open source models that are achieving very -- almost -- they're in rapid pace to achieve similar outcomes. So as we see the enterprise is deploying a mixture of proprietary and open models what it affords enterprises more choice in the cost structure of their deployments. So when we think about the deployment of more and more AI infrastructure, I'm fairly bullish on that. So I think people are looking at this in terms -- first of all, just in the infrastructure itself as a capital return. So you look at it in the land power shell, you look at how do we achieve an unlock of the bottlenecks that we have today. The model neck currently, everybody is looking at is the access to powered land to power itself so that we can scale with the ambition of the industry. And then when we look at an ROI at every different level of this, it's land powered shell, typically, people are looking for a 15% IRR. How do you bring that together? How do you execute that? When you're looking at deploying the AI infrastructure itself, it really moves up the stack. So I think we're going to see a tremendous value shift across the chain, where value is accruing in certain parts of this today. It used to be accruing in the model and application and in the semiconductors. We're going to see that value shift across the chain as this becomes as we continue to deploy at massive numbers of gigawatt scale.

And maybe just as a follow-up. I know you're relatively new in your role at General Catalyst. But as the Chief Strategy Officer, presumably, you're thinking very hard about dollars that are going to be deployed by your funds. And so what's your observation as you look at it and look at the math behind it, do the IRRs look like they are going to be achieved based on the opportunities that you see across the AI infra opportunity set?

Yes. Yes, I think that there is a compelling case for the value to be achieved for, as I said, at each layer of the stack, there's a different math. There's a different algorithm behind it. But one thing that is for sure is that the importance of compute and compute infrastructure as we see this evolution, whether the evolution as anything, this is a rising tide, and we'll see minor corrections back and forth, but the trend that's there to the ability for AI to deliver significant value and outcomes, we believe it's going to be across the entire stack.

Great. For you, Hong, and for those in the audience, I sat in those seats for a very long period of time as a research analyst. I did that for 18 years. I've now been a banker in this sector for 19. So kind of 37 years of seeing how things evolve. And at Morgan Stanley, we've looked at all these cycles. You've probably seen research that we've done over the years at every generation of compute, starting with mainframes, going to minis, going to PCs, client server, wired Internet, mobile cloud, all those things. It's typically rest in about a decade and it's been an order of magnitude increase in terms of units, demand, all those kinds of things, returns. And now here in this AI era, it feels to me that it's well more than an order of magnitude coming. You laid out Hong, a very robust view of what the world looks like in 2030. I agree with it. I bet you're probably low on those numbers is the [indiscernible] point of view. But we all recognize that things can happen beyond our control. And I know that as a leader here for this company, you, Mark and us, your team, are very focused on under-promising and over-delivering and making sure that you're looking for inflection points where things could change. So with that being said, what are the sort of the near-term demand signals that you're looking at to make sure that things are on track and making sure you can deliver when you look out just over a shorter period of time, like over the next few quarters as an example?

Yes. Thank you, Mark. Yes. So we provide a multiyear guidance. And as I said, the near term, we have much better visibility the booking momentum has continued to accelerate. Over the last 4 quarters, we have been reaching the record level of backlog every quarter. And while also triangulate and compare notes with our industry peers through, say, Morgan Stanley event and other industrial events. What we see that share the same sentiment. In the near term, for the next 2 years, the demand horizon is very solid and there is no single customer coming in at hedging and say, "Well, can we dial down a little bit.?" So I get more cause these days to now help us to accelerate they pull in some of the demand, even though we are scheduled for Q2 next year, and we wanted to have it earlier. And there are no sense we sensed that the industry, the channel is stacking with the inventories. So that gives us a conviction and confidence for the near term. Then longer term, certainly, right now, the market is a little choppy. They look at all the different things. But either way, is unprecedented investment cycle and a great opportunity. At the AI user, we know that today, I would say the company-wise, we don't really know how to work without AI. We're using AI in every area in our design, market analysis in our node conversion in our disposition of material yield analysis failure analysis, everything, in every area. So that gives us the confidence. We are increasing the investment spending on the side. And I think we asked ourselves in the beginning, what is the return on investment. But now the baseline is with AI, we don't ask that question anymore. Simply put, we don't really know -- it's not going to be coming -- going back to the -- the days without AI.

Yes. And it feels to me as well that just this opportunity here is so enormous as we just have to completely change the way that compute architectures have been done in the past, Jensen talks about AI factories and so on. And it's just gating factors here that prevent in this cycle things getting overheated versus what we've seen in prior cycles. So land shell power, wafers, there are just some natural gating things that prevent access is getting built up, which, to me, if demand holds and I think we're still early on in the applications, one would think that it should go for a while. One of the things though, Alan, that we hear from people is just -- the overall cash flows, in some respects, people could argue the world this turned upside down. I love the way the world works today because all the companies that I think have just done more to transform our world in technology, being semiconductor companies now have pricing power. They've got variable cost business models by and large. And a lot of the CapEx burden is now being undertaken by their customers. And they're obviously powerful, like David's company and others. But what's your view of just how long that can last? And it sort of comes back a little bit to Alex's point of view on what the ROI is on those investments. But what's your just high-level view of how long can we see this? Or are there issues that we should be focused on relative to hyperscalers, neo clouds, enterprise all the spend that's necessary to deliver the goodness that's coming down the road?

Yes. I think it's very sustainable. So today, we could say this is focused more on hyperscaler space or hyperscaler plus neo clouds. And what we're seeing on the early days of inference and Agentic is an incredible unlock. I can even mention that like internally to themselves, right? But the big opportunity here is going to be the enterprise unlock shortly after that, like the physical AI unlock. And those are kind of tremendous cycles. And if you look at the amount of investment relative to the unlock or relative to GDP or cash flow at a company level, we're still very much in the early stages of that. And I think that early stage plus sustainability plus what you mentioned, right, some natural gatekeeping items in terms of growth. allowed us to be a multiyear or, say, decade plus cycle to it.

Fantastic. Maybe on that point then, Alexis, if we see any type of slowdown in just the cadence of models and creation and how that evolves. Does that you think caused things to slow down for networking? Or is inference so powerful that it's just going to carry the day?

I think we're seeing a massive conversion from data center and AI infrastructure being used from training to inference. And what we see working in the space today or what I personally have visibility to is an insatiable demand for compute for inference. So I think even if the models were to slow down, we are just at the very beginning of this curve of people adopting AI and leveraging the current model fine-tuning the models fine-tuning them on their own custom data. We're starting to see real unlocks. And the beauty of it is, as the deployments become more efficient, we're seeing even more demand. So it is that [indiscernible] runs paradox of once people get access to the right model, they distill it on their own data, all of a sudden, they're unlocking higher value and being able to deploy further. I mean, I really believe we're just at the beginning of this. So I don't think that, that model slowdown would have an impact on the insatiable need for infrastructure.

Yes. Agreed. So David, maybe on that point, because I would fully agree with what Alexis just said, how does that impact the way that you're basically building the network going forward from an inference demand perspective?n

When we think of the requirements of training versus [indiscernible] training requires a larger cluster that is more synchronous in its bandwidth requirements. Inference is a bit more bursty, capacity can be a bit less predictable. You end up focusing more when it comes to inference more on scale up versus scale out. The other thing we think through is the larger context windows, which people get accustomed to appreciating require greater memory bandwidth and the agentic workloads, obviously require a tighter coupling of GPU and CPU. So as we think about the capacity we put in place, we think about those inference requirements relative to trading.

Great. Maybe for you, Hong, then -- what's your sort of view? What do you see as being unique here about the networking block effectively compared with compute and memory and how does that -- how does it matter? And where's the value creation that's going to come from here over time?

Yes, I'm sorry, I couldn't -- can you repeat your question? .

Yes, sure. It's really the question, what's unique about the networking block compared with compute and memory that's going to create the value around the networking opportunity that exists?

That's -- so as I mentioned, when we understand and we often get asked by the customers that the compute power is plentiful in there and is innovating in an accelerated pace. Every 12 to 18 months, there's a new generation of processors coming out. But we do not in the connectivity space, innovate as fast. So we've become the bottleneck. And power is another key consideration. And every time the current connectivity infrastructure rely on DSPs, it almost feels like there's a crutches you have to use. If you can design a clean signal path, you have DSP to do the cleanup job. But what we try to do is using an analog solution to provide a clear path for it. I do believe we can provide more value in the entire data center by innovating and providing a better solution in connectivity. I think we're getting a lot of encouragement and pressure and become accelerated innovation pace because we are the choke point of the entire stack to monetize, to provide more value. So we can be enabling factor for the future.

Yes. Maybe when you think about that, Hong, as just a follow-up, it feels to me like there are so many different companies, the large ones as well as, obviously, lots of start-ups that are focused more on the compute side of things. And yet on this whole networking connectivity that's going to create a lot more -- a lot of value to make sure that we can break through the copper wall and through the memory wall, all those things. that the number of players that have the real core competencies to do this and what Semtech can do, feels like a pretty small group of people. So maybe just talk a little bit about that. What you see is the real core competencies and the competitive dynamics around that vis-a-vis Semtech?

Yes. So that's a very good point. You look at the entire ecosystem, there seem to be more people work on the compute side and the networking side I would say people start -- the ecosystems just start realizing the criticalness of the connectivity. For example, Alexis once she was at Intel which is back in 2016, use launch to the effort of the CPO. At that time, we did a demo in the common comments is that, "Oh, we don't need that much power. We don't need that much bandwidth." So even though we had a technology, but it didn't get commercialized until the compute really crank up so much data for AI computing and bandwidth density need to be increased dramatically. Now all of a sudden, that becomes a trend. I think the connectivity in a way is always -- at the entire time, we call adjacent, it's not a frontline center. The front end center is still the compute. So that's probably part of the reason the entire [indiscernible] ecosystem seems to be focusing on the compute more than connectivity and that analog is kind of like an old-fashioned way. And so now it's in the front stage because they recognize the power advantage. For example, the [indiscernible] provide power saving by 90% compared to retimer. And people did not believe it can and said, "Wow, you really, why?" mean now they start recognizing that. But you see more innovative companies and starting up. And I think the market is going to be driving the ecosystem to innovate. They're talking about the links by DLB lasers, by micro VCSEL arrays, micro LEDs and even the RF. So people start paying more attention to the connectivity at this point.

I think it's a really great point, Hong, because we talk about data centers. I mean, in the world that I'm living in where everything is measured in power. It's how many megawatts and then you back calculate from that, how many accelerators, how many CPUs you could fit into that power envelope. And so we've seen the switch from simply model hardware codesign. It used to be hardware software co-design to model hardware, codesign model, silicon. And now it's really -- we're optimizing the entire system around power. And so some of these -- the analog approaches and anything that could save power or where we're innovative in that network. And then how do you build the orchestration layer of the software to be able to take into account the underlying layout of the compute system, the network, the network bandwidth, it's all being co-optimized. So it's a tremendous opportunity. When we see a constraint like power, it's really truly grid-connected power where now everybody is building all the behind-the-meter power solutions just to unlock the capabilities. I mean innovation is happening everywhere. And ideas that didn't meet the ROI for investment before are now being looked at again. So it's a renaissance across the network, I would say.

Yes. David, would you add anything to that relative to -- we've kind of laid out where some of the constraints are? Obviously, a lot of innovation is going to happen to deal with those. But thinking about power rack density optics, anything else you'd add to what you just heard?

Yes. I mean if I think of the network power relative to everything else we use is probably digit percentage number. Now obviously, we're deploying gigawatts of total capacity. So single-digit ends up mattering but the real value of the power -- of the benefit of optics, in my view is from the architectural innovations that it enables. So for example, I'm somewhat -- if you guys had a chance to walk around the OCP floor, you've got to see that some of the racks now are getting a little bit bigger, but that only allows you to extend the scale-up domain so much. If you want to get to a scale-up domain of, let's say, 576 GPUs, you need to find a way to connect [indiscernible] GPUs together. It's -- and for them to act as one optics is what makes it possible. So I think of optics as a lot of the value delivers is the added performance that those larger scale up domains, as an example, enabled. That wouldn't otherwise be possible or at the campus level at the cluster level, we can now -- optics allows us to have these large clusters that are pretty significant distances or a lot more distant from each other than they used to be. Our -- so I think of the value of optics significantly higher in that front. Our [indiscernible] cluster, our New Albany, Ohio, where we trained our latest model, news, wonder. You've got buildings that are a couple of kilometers away. You've got buildings that are as much as 60 miles away that we're adding to that campus. It's a gigawatt cluster, and optics allowed us to make that possible, make that act as one. Obviously, we had to put a lot of innovation in place to allow for that. We build a back-end aggregator site, which is basically a super spine. It allows us to aggregate regions of 20,000 or so GPUs in one. We have to build deep buffer so that traffic that's coming in from a couple of miles away doesn't get dropped. It gets kind of merged and inserted with everything else. But again, like in my view, the power is important. But the architectural capabilities and the performance you get from those is where a ton of the value comes from.

I'm glad to see there's just no complexity whatsoever in your business and the things you need to make decisions on. But maybe as you think about the things that are quite important to you or to other operators as well, how should people think about what to build internally and own internally as your real value prop versus what should come from the merchant market?

Yes, that's -- we basically think through about kind of 3 high-level questions if we decide to build versus buy. Do we have the ability to optimize the full stack, hardware, software compiler if I build something internally and can I do that better? Is the volume sufficient to justify the NRE because these aren't small investments that we have to make? And then is there a TCO advantage as I look to scale out this internal solution? Generally, we've taken the approach, though, that even when all those -- all 3 of those questions are positive, we still want the fleet level to maintain a portfolio that includes both merchant silicon solutions as well as our own internal silicon solutions. The reason being is we think that fleet allows for a more resilient data center, and it allows us for better options as we scale our solutions. So -- but generally, those are kind of the 3 high-level questions on how we think on [indiscernible] buy.

Great. Maybe Alan, coming back to you because we just talked a little bit about power and obviously, that's a big issue, and people now are talking about compute as a function of power. And the numbers I keep hearing just keep changing, it seems like every quarter, every year in terms of like what is the amount of spend on a capacity effectively. So from the research that you're doing, what do you think that is today? And does that change in the future? Because I keep every time I hear someone tell me sitting in David set or someone else's see, it's a pretty high number. So just sort of cost per gigawatt to deploy.

Yes. So I think that's really operator-specific, right? It's a very different answer if you're a GPU as a service versus more of like a search or social platform and what you're trying to do with those workloads. But networking right now is kind of in that, we'll call it, 12% to 15% range. And that's going to go towards the 20% range in terms of spend per gigawatt and ultimately, networking is that tool that unlocks everything that we said like networking is a tool that unlocks power. It allows you to go that additional power source. Networking is a tool that allows you to unlock context windows, right? Pull that memory together. So networking is a required tool in all of these unlocks. It's like the only universal tool that lets us solve all these problems.

But just so we get the denominator right, what's the -- what's your view of total dollars per gigawatt all-in, compute, everything that's going in to basically make that happen.

Yes. So total dollars per gigawatt is somewhere around $35 billion to $50 billion depending on what you're doing, build versus buy, et cetera. And I would expect that to increase with time, not decrease with time.

Yes. I just heard someone yesterday tell me $80 billion. Just -- it's like the numbers just keep going up. And obviously, that's in part because complexity is going up as well, and this stuff is not inexpensive for sure either. Okay. So Hong, kind of your perspective of proprietary versus open networks. How should we think about the opportunity set of what things are going to be just open standards versus things that might be more proprietary as it relates to the networking piece?

Yes. So it is great that right now, I think we are doing both the standards and the purpose but. So OCP, by the way, David on his spare time, he's leading the OCP open compute project and the Chairman. So I really appreciate you break away from the show and come to join the panel. But the standards define some common ground so that the innovation ecosystem and innovating in the most efficient way. But with the accelerated pace of innovation that we talked about early on, you have to engage to allow some purpose boot and so the engagement with hyperscalers, they have the scale. They move very fast. They know exactly what they need for their architecture, for their networking. So that's what we found this new way of engagement is a combination of both standards and early engagement. So that's a part, as I talked about the engagement model. We changed through the Semtech rising phase is that having the end customer engagement to define what the future generation, for example, right now, the SPO and NPO, the NPO probably still doesn't have a standard yet. The standard pinout and all of this connectivity requirement. But people started innovating already. There may be some form factor will become the de facto standard through this course, but you cannot sit back and waiting for industry to argue out and define a common standards and then you develop and that will be just too late. So that's a challenge for us. That's also excitement for us to see that connectivity is finally is not just a conduit to connect something, is it enabled [indiscernible] the efficiency of the AI data center.

Yes. So maybe, Alexis, turning to you again, and you highlighted a little bit on this before, but maybe just kind of give us a [indiscernible] of where we are today in the journey between pluggables going to MPO to CPO, XPO and so on. So where are we in that path? And as we get to CPO in particular, what are going to be the key things that you think will determine success for the suppliers because it's going to take a lot to get embedded altogether?

Yes. I mean I think this really comes down. From my perspective, it's -- on having operated large-scale infrastructure, it really comes down to an operational metric. What people are trying to optimize for is the uptime and the reliability of the infrastructure itself. So as we look at the decision points and we look at where the failures are in the data center itself, there's obviously, the entire industry works on pluggable optics today. But we're seeing the switch over to co-packers optics. I know that there's a big debate on co-packaged optics versus NPO. But when you step back from that, both of them require removing the entire blade to resolve. So my personal view is the Leap drug to co-package optics because it really comes down to a mean time to repair blast radius mean time to repair and how do you replace that blade. So I think the ultimate reliability of a co-package optic solution will be higher, and the goal of the industry is to bring that to a standard silicon level reliability and then you still are using these -- the optical light source is still a pluggable solution. So I think that, that light source remains field repairable. And then it just comes down to the architectural decision of how many light sources per pluggable unit. So not to make it too technical. But I really, when you step back, it's -- I do see the evolution from pluggable to co-package because it's a significant improvement in and it's at least a 30% improvement in power consumption and then the signal integrity and the performance is extremely high. Now there is -- I think there will be people who choose to go to the NPO route because it's more comfortable. But we'll see every different hyperscale adopt a different solution on that. But ultimately, that would be my own personal view.

Yes. Did you want to add something, Hong?

I agree. I mean, ultimately, the CPO provide the best power efficiency, people do per bid level, the best integratability and also the stability. People have then kind of in a way, misconception, this is serviceability of reliability, what is the laser having the reliability issues. Well, traditionally, laser has been the route of our evos. If the optical transceiver failed 9 out of 10 times is because of lasers, this state is different and laser reliability is not easy. What Alexis mentioned about the uptime, the reliability, it's more the network, the linked flaps and had to say this when you have the DSP, when you do not test out the corner conditions in the network, it's prone to link fabs. So you don't bring the optics so close to the XPU, you don't have -- you don't need a DSP, you don't have the chance to kind of getting that complication integrated. So I think that's great. NPO in the meantime, is a interim open ecosystem play. So on CPO, they are establishing the entire manufacturing ecosystem because based on the silicon photonics, it is not as mature as a digital logic from design, from packaging ecosystem. But the industry is building that and they're catching up so quickly. I have no doubt that CPO is going to happen in the future. NPO In the meantime, provide a viable interim solutions allow different hyperscalers to get their customized XPU interconnected. And while they don't have an industry standard again defining pin-out and mechanical key power cells.

Yes, agreed. And it feels to me like next 5 years, CPO is likely to be one of the most disruptive things we see here in infrastructure. Maybe Alan, from your point of view, I recognize it's sort of changing, but where do you think the knee of the curve is for CPO deployment? Like when does initial ramp really happen? And when is it just catch fire?

Yes. The initial ramp is now, right? There is CPO in production. And then I'd say, decent or significant volumes beyond proof of concept. The inflection is going to be second half of '28 or into '29 where that inflect. There's kind of 2 bottlenecks there. The first is manufacturing CPO is new, and it takes a lot of new tools in order to get that through manufacturing test qualification. That's new. The industry needs to grow into that. And then secondarily, you've got a time into these road maps, whether that's on the, we'll call it, network ASIC or the XP GPU, right? And so that kind of creates a question like, do you go NPO to CPO, do you jump right to CPO? That's really going to be a personal decision of each hyperscaler or each accelerator. But the inflection second half of 28%. And as you mentioned, like it's going to be highly disruptive, but in a massive opportunity way.

Yes. So David, [indiscernible] to you on OCP given your role. It seems like it's been just a home run and you've outgrown this area. So I understand it's moving in the future. But it seems like one of the key things coming out of the show has been [indiscernible] versus other approaches. So do you have a view on that, [ MicroVixls ] MicroLED? How do you think that ultimately plays out and recognize that it's still early, but decisions are needing to be made now for these ramps?

Decisions are being made now. And if I kind of put my OCP had on, generally, our approach is to allow the ecosystem to debate and arrive at a solution that there is a kind of greater alignment on there -- it's probably not the way to go. It's probably not the fastest way to get to what ends up being a standard. But we find that it's the most enduring way because you've allowed the technical debate take place, and you've allowed the voices of support and opposition to kind of bear out. So -- and OCP is obviously really well suited to allow for that to happen as various work streams of companies come together, collaborate and ultimately turn that into actual products that enter the market. .

So very nice consensus building thought process. Alexis is a deep technologist, or Hong, maybe the 2 of you, which you when you think of it from just a pure technical perspective of solving all the challenges that would exist, which do you think is going to be -- which I think wins out in the end, recognizing we'll coalesce people's perspectives, but the technologists new, where would you think this shakes out?

[indiscernible].

[indiscernible] as an example.

[indiscernible]

[indiscernible] and wide for scale up. So now they seem to be the recent trend is the argument is to MicroLEDs, you leverage the consumer electronics industry, the MicroLED for display you do the slow and wide. Do you know I know your opinion on that.

I mean -- so yes, so I mean, as a pure technologist as an individual, I believe, in the MicroVCSEL or VCSEL approach or a MicroLED approach, but that's just because of the signal integrity and the high-bandwidth communications. So also, when you think about the challenges of the connector and how large the connectors need to be, it's really about system in, but that's -- personally, I also lever towards the higher bandwidth solutions, higher bandwidth, less connections because so many fibers, so much connection I've just personally seen inside the data centers and the racks how difficult it is. But what is your opinion Hong?

Yes. No, Alex, is I totally agree in a slow and wide it needs to be reasonably fast. If it's too slow, you're adding a reverse gear box you're getting the native signal converted back to slow signal in the other side to do the other conversion is inedibly adding more power consumption. Kind of, in a way, [indiscernible] purpose. And -- but of course, the better to be fast and wide. You may need that. And that is what's really the OCI is promised to deliver this A channel, DWDM later on 16 channel, even had the promise to do 32 channels. Each one go with a 64 gigabit or 128 gigabit per second that provide high bandwidth and high bandwidth density too. But I do see that the industry thinking about all the different ways to innovate not only for performance, for availability, further cost for reliability. Those are really the things that OCP and a consortium to try to promote, right, not just focusing on technical aspects but the totality of the supply and the viable solutions.

Yes. Okay. Maybe a question for both Alan and David here. If we look at the transitions that are happening here, from like 400 gig to 800 gig to 1.6, 3.2. The cadence is shrinking. And so that's great. We're ramping stuff faster than ever before. Is that sustainable? Or does something ultimately have to give maybe from a market perspective and then from the user's perspective, David?

Yes. I'd say from a market perspective, it is sustainable, right? 800 gig is over 2 exercise of 400 gig in terms of volume, and we can expect that at 1.6 will be a 2x plus the volume size there. So the market is very much able to absorb those types of volumes. But we are running into a little bit of a dilemma wall there. We've more or less maxed out the ratings from a switch perspective, and we've maxed out the beachhead from an ASIC perspective inside the server. So that requires the more intuitive ways. That's why there's so much excitement over MicroVCSELs this year, right? It solves some of those beachhead dilemmas. So the market can absorb it and the volumes are absolutely there with the cadence. Cautionary tail is if you're at [indiscernible] how much 400 gig are you going to have? So you ultimately want to make sure that you have a home for those technologies as we go through these faster adoption curves.

Yes. And from a user perspective?

Yes, I mean, so I'm obviously not going to answer the sustainable question. But the way I think about it is that evolution of the network, is kind of matching the evolution of compute. And the evolution of compute is largely matching what our kind of needs are. So if I think of [Audio Gap]

[Audio Gap] And -- it's been a great discussion. There's obviously lots of complexity, lots of opportunity here ahead in this whole opportunity set for networking to drive AI infrastructure and all the benefits that it's going to deliver in the future. But as said, it's a complex world. There's lots of markets concerned. What's the one just assumption you think that the market just has wrong today? About -- because to me, it's obvious where things are headed, but the market doesn't always see it the same way. So what's the one thing you think is an assumption people are making that just isn't -- doesn't hold water, start with Alan.

Yes. I think it's that these are binary answers, right? We're going to get to CPO. There's no doubt there, and it's going to happen, I'd say, roughly in that 5-year horizon. But the misconception is that these things are binary, copper goes away, copper away. If we do MicroVCSELs, we're not getting rid of transceivers. So these aren't going to be binary events. It's the right network for the right part of the server or the right part of the memory, and that ultimately leads to a lot of markets running in parallel.

David?

I'd be -- I'm humble enough not to make a statement about what the market is wrong. I generally observe the market to learn from. And over time, find that it's actually more pressing than most forward-looking statements out there. If I were to add one area where I think it's probably underappreciated by the outsiders is the amount of innovation that's happening in the hardware space right now at all levels. There is more emerging companies, there is more start-ups. Many of them have fantastic ideas they're working on. And the rate of innovation is kind of unmatched to anything I've seen in the past. 25 years I've been in this space. And I think that's generally underappreciated by the outsiders, those outside looking in.

So I agree with both of you have said, I mean I would take it just at a slightly different angle is I'm an incredible optimist for the power of AI for human outcomes. I think this will be an era of incredible abundance that we're about to unlock. So it may not be what the market is getting wrong, but I think that there's a lot of fear in the market that could really be converted to optimism. It's -- we're at the era and at the precipice of an incredible time. And it's going to require an incredible amount of compute and compute access. So just like -- I mean it's the opportunity to -- I fundamentally believe it's an opportunity for all of human kind globally to seize this moment and really unlock the power of AI. So that is -- that's more of a wish on my part. But I get to operate at the very frontier of AI, I see where it can go. And so I really am very bullish because of that. And I think that we're going to unlock this moment. This moment will be really better appreciated maybe in the coming year.

All right, Hong, you get the last word.

Yes. So I think the disconnect what I see is today's assumption or sustainable or not is based on the technologies today. But we continue to innovate. For example, you're going to be hearing this afternoon XPO, increased density for 4x power efficiency can be improved. So the cost of debuting data centers, the power consumption, the labor and footprint you need is all going to be revolutionized. People have now factored that into their model and look at the macroeconomic situation. Another thing is a disconnect between the macro concerns. And really, we are in a tranche of building the product and we have the great visibility in a daily drill, we have to go through in expediting. So I hope sometimes it will be converged somewhere. But I'm very excited about this. I do believe where they said, the best is yet to come.

Yes, I couldn't agree more, and you can't change the world in a straight line. So -- but I think there's this tremendous opportunity ahead. I want to thank the panel for just great insights. Thank you all for being here as well.