video_OCP2026_CapitalGroup_投資人看NPO_20261015

Thank you, Eric. Uh Good afternoon, everyone. Uh thank you for this opportunity to share my views. As Eric said, usually you don't get investors presenting at these forums. At the end of this conversation, you'll say, "Ah, I know why." Uh but in the meantime, I will still try to add some some value, uh not talk about technology, because this audience, my fellow panelists, they know much more about this. They do this for a living. So, I'll try to give you a viewpoint from really an investor's perspective, the things we look about, uh the the things we focus on in terms of investing, uh and where we tend to get it right and often wrong. Um as Eric said, I work for a firm called Capital Group. We're a mutual fund manager. I focus on uh public market equities, and especially AI infrastructure. Um so, that's my background. Uh covered the space for many years, and watch it go through some cycles. So, uh that's just as background. So, let me just start with a big picture, which is how do investors think about this NPO question? And that makes some sense only once you see how far down the stack it sits in an investor framework. So, the first question an investor asks is whether the big capex numbers are sustainable. Frank talked about some of the numbers. I won't put up the obligatory capex slides, but just to put it in perspective, both Google and Meta are expected to see their capex grow 10x from 2023 to 2027. And I thought they were a lot spending a lot of money in 2023. So, it kind of puts in perspective how big these numbers are going to be. Capex estimates have gone up every quarter for about last couple of years, given the demand for AI compute. And the hyperscalers alone are expected to spend over a trillion in capex in 2027, combined with neo clouds, AI lab self-builds, sovereign builds, enterprise, that number's going to be north of two. I don't know if it's going to be uh you know, some of the much larger numbers I've seen, but it's going to be north of 2 trillion by 2030. So, most of the debate you read about, most of the debate we as investors focus on is about the sustainability and financeability of that spend. Beyond that, as investors, we're trying to find areas of relative strength and weakness to determine winners and losers. So, we think about how much of that CapEx is AI? Is AI growing as a percent of the mix? So, is that AI CapEx growing faster than that? Within that, how much is allocated to compute, memory, connectivity? Market was led by compute in the last year. We've all seen how much memory has been a shortage, an area of growth. And with uh you know, scale up, scale up, cross scale in, all of these things we're talking about, you know, connectivity is going to be a big area of growth going forward. And then within connectivity, there's the next level of debate. Copper, optical, that's where this audience focuses on, but also Ethernet versus NVLink and several others. And only then you get to NPO, CPO, pluggables. And the reason I went through this is to make two points. First, each layer is a separate debate. And as investors, we're trying to answer every layer. We sort of have to be right, especially at the top layers, to make a successful investment. Being right about NPO wouldn't help much if you're wrong about CapEx. So, that's an important thing. And then second, most investing decisions are also debated at the top layer. So, when you see the volatility in the stocks, people wonder, you know, what happened? It's usually people are concerned about something at a macro level rather than something about individual companies or sectors. And that is the risk and the opportunity. You could be You could get an architectural shift right and have something macro go against you, but also if you understand not just the top layers, but also what's going on in each of the layers of the network, you can identify good opportunities. Um Vlad will, I'm sure, walk through market forecasts understanding them is very important to modeling estimates for companies in the optics value chain. So, I'm not going to go that through that. But, the point I would make is at times of major change, like we're in now, it's often more important to get the architectural shift right. And for optics, I think the key point is I expect it outgrow AI CapEx meaningfully. Reason is simple, it's well known to this audience. Accelerator chips are their size limit, so compute capacity grows by connecting more of them together, larger and high bandwidth scale out clusters, scale across clusters, scale up domains, and and now I should add scale in. Connectivity therefore grows faster than AI CapEx, and within connectivity, more of every link move from copper to optical. So, optics grows faster than connectivity. Uh three numbers to dimensionalize this uh between ECOG and AI infrastructure summit. So, these are some of the numbers that are thrown around. Meta said uh in its 1 gigawatt Prometheus campus, they use approximately 4 million kilometers of fiber, and the 5 gigawatt Hyperion site will be several multiples of that. To put that in perspective, that would be larger than the entire cloud data center fiber market a few years ago. Transceiver units more than doubled each of the last 2 years, and could double again next year. And when scale up stretches across multiple racks, number of optical links between these racks increase by four to eight times. So, in any dimension you look at it, any segment of the network you look at it, the growth is very large. And this is all before the rack backplane becomes optical. So, the important point from an investor perspective is optics will outgrow AI CapEx, and the outperformance holds at whatever rate AI spend grows. It doesn't protect you if spend goes negative, but it means optics does compound faster than the CapEx line. Um this audience knows Richard Sutton's computing bitter lesson. General methods that scale with compute keep beating specialized approaches. They know this, but we seems like we have to keep relearning this. There's been questions all around about do scaling laws hold? And so far they have. Investing has a bitter lesson, too. It has more than one. And investors relearn it just as reliably in this case. I started covering optics right before the last bubble. So, I've learned my share of bitter lessons in this sector specifically. Or maybe the fact that I'm here suggests that I haven't learned any lesson at all. Well, we will be able to tell. But the lesson for investing is that being right about the technology is not the same as being right about the investment. As an investor, when I underwrite a transition like this one, I'm making three separate calls. So, the first one is around technology direction. That's usually the easy part. You talk to enough smart people in the industry, you know directionally where this is going. Uh the trend towards optics has been visible for a long time. Industry tends to get the direction right. The second is around timing. And that's often where I've gotten it wrong historically. And it's where investors are really paid or or punished uh in in their investing. Being early to a theme can look exactly like being wrong for a long time. This was the first bitter lesson. The move from copper to optical communications increased for need for increased optical connectivity. That was very clear architectural case for many years, but needed a use case like AI to really drive that growth. The stocks of companies that participate in this market went sideways for several years. While the thesis was correct, timing was off. And the third is identifying sustained winners. And this is often the hardest to call. When you have a rising tide, it initially lifts all boats. And so, the whole sector can look like a great investment. But the key to long good long-term investing is identifying companies that have a durable advantage through the cycle. Leading at the inflection is important, but not sufficient. The advantage has to survive the next architecture. This was the second bitter lesson and we saw this in the in the cloud era. Cloud drove a big increase in CapEx. The early winners were the vertically integrated players, but over time disaggregation won out and if you didn't recognize this, you might have been right on the early part of the investment but missed the transition. The point for this room, in my experience, the end customer decides which innovation becomes mainstream. So, that means my fellow panelists, the people who are building this, many of you in the audience make the decision. The most useful thing an investor can do is show up at events like this and listen to people deploying the technology to understand not only direction but also timing and sustain advantages. That's why I spend time at forums like this. That's why I'll be all emailing all my fellow panelists for follow-up conversations and asking these questions. The next point I wanted to make is somewhat NPO specific. For NPO, the relevant transition as everybody in this room is copper to optical and it's really three transitions happening at a different times. First one is between racks. Between racks is already optical in scale out. What changes is density with scale up. So, when scale up stretches across racks and to operate as one large accelerator, the number of optical ran links rise between 4 to 8x depending on on the use case. This is likely to be the first major volume application starting in late '27 wrapping in '28. And I'm giving you the investor perspective. This is what the market is assuming for for optical shocks. Second is inside the box. Today this signal leaves the switch or accelerator on copper traces and only becomes optical at the front panel. NPO moves that next to the chip. So, copper traces give away to fiber inside the server and switch. Similar time frame, late '27 early '28. And then finally within the rack, the copper backplane is the last move and I expect it to be partial when it does. Passive copper is the cheapest and lowest power option, so everyone I speak with wants to extend it as long as possible. The rule rack designers generally seem to follow is copper if you can, optics only if you must. This is obvious to several people in this room, but it is an important insight from an investing perspective. Timing. Nobody doubts the direction, but you have to get the timing right, or you're waiting for the optical content to show up, and it takes longer. It goes back to the point on being right on timing. I've heard people make the low power argument as well. And but I've heard that argument made for things like LPO, and there's not an ecosystem uh you know, uh driver and ecosystem around it. I feel like hyperscalers don't move away from solutions that work well and are cheaper, so I think timing is gated by the longevity of copper, and timing is key to investing in architectural transitions. And then my final slide, you know, where does NPO fit from an investor perspective? So, after thinking about all the big picture things, when you think about NPO specifically, what are the things I think about? I think of it as the most probable first step in the move to integrated optics for three reasons. First, you'll hear the same theme from speakers this afternoon, I'm sure, heard it at other conferences. NPO is designed as a standardized pluggable socket that keeps the advantages of pluggables while getting optics closer to the chip. From an investor perspective, that's important. Things when things change as little as possible and existing technologies do well, that usually is a winning formula. So, what why is that? First, it keeps the optics ecosystem open. CPU in its truest form probably needs a switch or accelerator vendor to deliver it. That lowers the number of vendor options. Today, the optical ecosystem has many vendors. You don't want to lose that. NPO keeps that structure. Second, it doesn't require redesigning the switch die. Optics remain replaceable, so good yields, serviceability, repair time, those are very important operator concerns. And finally, history. The last time networking faced a choice between a vertically integrated stack and a disaggregated one, the hyper tailors chose disaggregation. And that shaped who captured value for the next decade. I expect that pattern to repeat. And I'll wrap up with, you know, a couple of points. First, things I'm watching to evaluate progress. Fiber attach yield at volume given complexity, serviceability in early deployments, standard traction with standards like copper is able to hold its position at 200 gig and 400 gig. And finally, investing is about strong opinions held lightly. So, here's what would change my views from what they exist right now. First, CPO reaching volumes yields faster than expected, which would favor the vertically integrated model, or copper extending its reach and longevity, which would push the whole transition out. So, I'll stop here. Thank you again for the opportunity to share my views, and I apologize for having to present and leave and not be able to take questions, but I leave you in the hands of my capable fellow panelists. They have the answers, and I appreciate your time. Thanks.