Invest Like the Best with Patrick O'Shaughnessy - Andrew Homan & Chris Miller - Redefining Semiconductor Progress - [Invest Like the Best, EP.396]

Starting point is 00:00:02 Hello and welcome, everyone. I'm Patrick O'Shaughnessy, and this is Invest Like the Best. This show is an open-ended exploration of markets, ideas, stories, and strategies that will help you better invest both your time and your money. Invest like the best is part of the Colossus family of podcasts, and you can access all our podcasts, including edited transcripts, show notes, and other resources to keep learning at join colossus.com. Patrick O'Shaughnessy is the CEO of Positive Sum. All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of positive sum. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of positive sum may maintain positions in the securities discussed in this podcast. To learn more, visit psum.vc. My guest today are Andrew Homan and Chris Miller. Andrew has spent two decades at Maverick Capital and is a managing partner at Maverick Silicon, where he leads the firm's technology investments.

Starting point is 00:01:07 Chris is a professor at Tufts and the author of the New York Times bestselling book, Chip War, which details the geopolitical battle to control the semiconductor industry. Today we get into a comprehensive discussion on the semiconductor ecosystem and the silicon backbone of our digital age. Andrew and Chris share insights on how venture capital is navigating this complex industry and what it means for the future of computing. We discussed the AI-driven revolution in chip demand, the geopolitics of semi-manufacturing, and the next wave of innovation behind NVIDIA. Please enjoy my conversation with Andrew Homan and Chris Miller. So, guys, I thought a fun place to begin would be to talk about one specific thing in this ecosystem, just as a jump-off point. And Intel seems like an interesting business to do that with because Intel's been around for a long time, there's been periods when Intel was an unbelievably dominant and pioneering business.

Starting point is 00:02:00 and we're in the midst of a paradigm shift. So maybe you can describe why Intel the business is an interesting one to study to learn about what's happening in the field of semiconductors today. So I guess maybe to kick it off, I was at an event last week with the global semiconductor alliance. And this topic came up and someone asked, what film genre would you describe in terms of what is happening at Intel right now? Is it a horror show? Will it end as a romantic comedy? And I'm not exactly sure what term I would use to describe that right now. But someone from the company got on stage and his comment was I would describe it as a Hong Kong action film, which is probably a pretty good example of explaining the state of play right now there. And I think there's a lot of

Starting point is 00:02:48 moving pieces. You've got geopolitical tensions laid on top of all of them. You've got competitive issues on the design business. They're trying to break into the foundry business, kind of a next new leg of the stool for them. But there are a lot of moving pieces, I guess is how I would kick it off. But Chris, you've been following the company for a long time. I think they illustrate how hard it is for a very successful company to pivot to a new business. Because Intel's problems, in my view, are actually a function of their success. They were too successful in producing chips for PCs, too successful in data center for a long time. And it made them risk averse and hesitant to invest in new products. They missed the mobile revolution. Famously, Steve Jobs asked them to produce

Starting point is 00:03:28 shipped for the iPhone. They didn't do it. And now they're struggling with AI because it was always a huge risk to try something new. Why does this happen? I know that's the reason I've read the innovator's dilemma like everybody else. I understand juicy existing profit streams. But also, this is so well documented across business history, especially in technology. These are all very smart people running Intel with lots of experience. Why don't they just say we're not going to fall prey to this specific kind of disruption? And we're willing to do what it takes, including betting lots of our free cash flow or whatever to not lose this way again because we know this is the way companies like us lose. I think it's painful for institutions and for individuals to actually

Starting point is 00:04:09 embrace the innovator's dilemma because it means you have to sacrifice the products that you made your career on. You have to promote people from different business streams who are working on new products that threaten the old ones. And so there's just a lot of tension that's involved. It's much easier to keep the status quo, especially if the status quo looks like it's profitable. And it was profitable for a long time. And by the time you realize it's the long and profitable. It's often too late. Interestingly, Andy Grove, one of the initial leaders in entire. Exactly. He wrote a book about this. And when you read the book, you're like, why is this not mandatory reading the entire industry, but certainly the Intel management team,

Starting point is 00:04:43 because as you talked about, you missed this transition to mobile. It seems like they've largely missed the transition to AI. And so you've got one of the founders of the company that had highlighted these risks and these paradigm shifts and how you need to be ahead of them. And for whatever reason, And it's been challenged for them. Can you describe that paradigm shift that we're going through right now from your perspective? So I think when you look at technology, every decade there seems to be what I would describe as an architectural change.

Starting point is 00:05:10 And so you went from mainframe in the 80s to the PC. Then you went from PC to mobile. Cloud also was happening simultaneously. And now this next paradigm shift that we're just in the very early innings of would be AI. And the implications from that I think are probably broad. or perhaps than any of those other paradigm shifts that we've touched on so far. Did you see it any differently?

Starting point is 00:05:32 You've studied the history of this as much or more than anybody. Is it that simple or is there more nuance to appreciate? I think that's right in terms of the end markets. I think simultaneous to that, there's been shifts in the business models along the way. So the history of the industry has started with companies that were entirely integrated. They designed and produce their own ships. And you shift to the fabulous foundry model with TSP producing ships for everyone. And now you're in a position where you've got a small number.

Starting point is 00:05:56 of AI companies or cloud companies that are not just big buyers of chips. They're also reverting to the prior practice of designing their own chips and getting really deep into the weeds of the technicals of how their chips are produced because they want to have that entire integrated process where they can control from the software all the way down to the silicon. Could you pick one company example that's making that transition away from being just a hyperscaler and going vertical and describe what's happening in a lot of detail? What this actually means in terms of terms of outlay of investment, what they're actually building, why they're doing it. One of these transitions would be interesting to highlight what you mean. There's two directions that's happening.

Starting point is 00:06:34 It's happening at the chip companies are going up like Nvidia. We can talk about that. But at the hyperscale level, they're all going down. They've all bought AI accelerator startups. They've all tried to integrate them. They're all trying to understand how to bring the networking together. And they're having to hire hundreds and companies as thousands of new employees who have areas of expertise they've never had to hire in before. So it really is the same story, I think, meta at Microsoft, at Amazon. They're all essentially following the same playbook. Maybe describe the logic as you see it. I think you look at the enormous capital outlay that all these companies are making right now primarily for Nvidia, GPUs, and they look at where that spend is going to go in the future, which I think

Starting point is 00:07:13 that number is only going higher. And all there are alternatives that they can do, at least for certain workloads, where they can custom build a chip that is very purpose built for one specific workload, whether that be search, for example, or some sort of social media app, that will, allow them to potentially lower their cost to serve those customers significantly. It's fascinating to watch investors try to figure this out. And you can find a brilliant investor for just about any argument for where the value is going to accrue in all this. I'm assuming here that AI is going to be a massive huge thing on par with or exceeding mobile or cloud or whatever. Obviously, there's some version of the world probably where that's

Starting point is 00:07:50 not true. I don't know. Let's focus on the world in which it is true. I was with a very well-known investor in the private markets last week that said, look, we just think it's 100% going to be in the application layer. And so that's where we're making all of our bets. You're focused much more on the infrastructure silicon layer and everything around the chip. Why are you making that bet? What is it about your reasoning through where the profits will accrue in the future because of this boom that has you focusing so specifically on silicon and the surrounding technologies? So when I think about the broader AI ecosystem, we think about it as a three-layer cake, right? So the bottom layer, chips in cloud, that's where we're focused.

Starting point is 00:08:29 The middle layer would be the foundational models, and then the top layer would be the applications built on top of those. So the foundational models would be OpenAI, Anthropic, Gemini, etc. That chip cloud layer would be Nvidia, AMD, TSMC, etc. And then that top layer would be chat GPT, office co-pilot, Tesla full self-driving. And I think the challenge right now is that on those top two layers, it's very foggy in terms of who the ultimate winners are going to be and how much value is going to get accrued to those layers broadly. Eventually, yeah, there will be big dollars that go there. But in terms of who actually captures, though, I think it's very opaque right now. When I think about that chip layer, though, to me, almost in any scenario, I see that being the area that's going to capture a significant amount of economic rent through this paradigm shift that we're going through right now with AI.

Starting point is 00:09:20 What would be the reason that it doesn't? Let's say that we get to simulate the future in a whole bunch of different ways. And in some percentage to those futures, the infrastructure layer does not capture a lot of the profits. It's big, but there's pricing pressure and it gets competed down to the cost of capital. And there's just not a great competitive advantage period to earn money in. Draw that picture. Or can you draw that picture? How could that happen? I think there would be two dynamics. So one would be, did these scaling laws change? Right now, everything we see would say, no, they're not going to change. The more compute you throw out these problems, the better the model, the quicker you'll be able to train the model, et cetera. I think the question is, do you get into

Starting point is 00:09:55 a situation where there's much more competition, particularly in that AI accelerator specific area of the market where prices come down significantly? I think ultimately we do want prices to come down. That's the history of compute. But I think you've seen in prior transition, whether it be mobile or whether it was with PC, that silicon layer still captures a ton of value with PC and server. Intel clearly created an incredible business during those periods of time. And then specifically with mobile, you look at companies like Arm and Qualcomm and others that monetized significantly on that trend, despite there was more competition. Prices did come down. But the TAM was so large that these still great companies were built and created over those time periods. And I think the PC mobile era

Starting point is 00:10:36 both show that describing it just as silicon actually misstates the source of those modes, because it was silicon plus software and the interface between the two. Why are there Intel chips and half the world's PCs. It's because of the software's built on top of Intel's architecture. The same thing true for whether it's Apple and mobile or Walcom or Arm at the base of it all. And so you've actually had a long history of the Silicon Software interface playing a big role. And the software is often a big portion of the moat. And one other area that's worth mentioning on the scaling laws is with a release of open AI Strabery model, an area of AI that was thought to be compute light, that being inference,

Starting point is 00:11:11 is now compute intense as well. It's for that model to be better at reasoning. it requires much more and a much longer period of inference. It's really interesting to think about the history just from an investing perspective up until now, and I'd love you to reflect on it, where if you ask about semi-investors, in private markets, they'll often tell you, just don't do it. It's capital-intensive and brutally competitive, and you're just not going to make money as a semiconductor investor if you're like a venture investor.

Starting point is 00:11:37 And in public markets, there are certainly people that have done tons of semis investing historically, but there aren't a lot of semis fund, semis dedicated funds. There's plenty of software dedicated funds, but not a lot of semis dedicated funds. So why is the technology that has undergirded all of technology progress of the last 40, 50 years? Why has it been like somewhat of a backwater for public and private investors over, say, of the last 20 years? I think there's a dynamic where there were a lot of semi-focused funds 20, 30 years ago, and many of them did very well.

Starting point is 00:12:09 And then you had this rapid rise of software that kind of captured everyone's imagination. and many of the kind of, particularly the early stage community, pivoted to software and did fabulously well, right, investing behind that trend. And then you had a situation where a lot of that expertise just retired and left the industry. And it's probably, it is, semis are challenging. They're technically complex. They are capital intensive. Though I think like comparing them to software today, looking prospectively over the next 10 years versus, say, the last 10 years, my guess is that we're going to see more and more capital start to flow into this hardware layer because so much value, I think we'll get created there, and there's just so much need for innovation.

Starting point is 00:12:48 Where do you think the highest perspective returns are within the infrastructure layer? Everyone obviously is focused on Nvidia because that's where all the market cap is. They've made GPUs, everyone's talking about GPUs. There's so much more going on around them. Data centers are getting built out famously. A record pace with crazy amounts of power. There's all this cool stuff happening around the GPU itself. So where else are you most interested to study, look, meet with companies, and so on?

Starting point is 00:13:14 This ties into a point that Chris made earlier, less focused on that core GPU itself. And there's many companies trying to go directly after that AI accelerator market. I think it is challenging to go after the predator that it's at the very top of the food chain right now, NVIDIA being that player. And there's all these interesting areas around that GPU, whether it's interconnect or memory, or power-related, where you're just going to need innovation because the way of doing business that existed five years ago is just breaking. And the data center that existed five years ago where an Intel processor is king of that data center and there's some simple networking and then

Starting point is 00:13:55 there's some storage attached to it. The example that we get when we're talking to folks is that essentially you had a tsunami, an AI tsunami that you basically wiped that data center out and Jensen's riding his surfboard on top of that tsunami. And as a result, you're seeing everything needing to get rebuilt because the power needs of these GPUs are off the charts. You're trying to connect as many of them together as possible to create super GPUs. And so right now we're connecting 100,000 to get together. The next step is going to be connecting a million of them together.

Starting point is 00:14:24 And the complexity of doing that is creating, I think, a lot of opportunity. Chris, I'm curious where you've studied the history of this in detail. We won't go through it all. People should read your book if they're interested. which is excellent. But I'm curious if you take all that history up through to today, and then you spend a lot of time with companies, with governments, with investors, with sort of everyone in this ecosystem, what people are not talking about enough that is not yet enough in the mainstream narrative, but you're hearing a lot about in these conversations. I think everyone is beginning

Starting point is 00:14:54 to register how much money will be spent on compute over the coming years, decade, driven by data center buildout, driven by AI. The numbers are staggering. They're excited. if you're a data center builder or a power provider or if you're Nvidia, but that's actually a bad thing. We'd like compute to be as cheap as cheap as possible. And so I think the key focus right now should be finding ways to minimize the amount of money that needs to be spent or to build out the computing we need for AI. And we're right now at the stage of admiring the scale of the problem, admiring the billions of dollars that will be spent. And actually, the real goal should be finding ways to economize, to bring new technology to bear, to find ways to bring down that cost,

Starting point is 00:15:33 because that's going to be one of the key gating factors of AI development is can we bring the cost down? Can we talk about the political considerations around both the infrastructure and the first two stacks? We'll leave the apps for later maybe. But if we're talking about just the infrastructure, the fabs, everything that needs to go in to run these things, and then the model layers. US companies seem to really dominate this, maybe except for TSM and a couple in Europe. Talk about the most important political considerations given that if this technology is, the most important one ever. It seems like every country's going to want some answer to their foundation model or their data center or their infrastructure. What are the key variables in the political

Starting point is 00:16:12 landscape right now as you see them? I think if you start at the level of where chips produced, you've got a whole lot of concentration in Taiwan, as you mentioned, and the economies of scale in the industry are such that you're just not going to have multiple locations at the size of Taiwan where chips are produced, maybe a couple other countries, Korea, but it's not going to be every country producing is on chips. Totally impossible. At the data center level, if you're talking to about data centers for foundation models. Those are huge, brutally expensive projects. And so there too, although there's a lot of governments that would like to have their own, the scale of investment is such that unless you're sitting on a lot of oil or unless you're a very large

Starting point is 00:16:46 economy, you really can't justify the spend. And so that's why, if you look at where big data center projects are happening in the U.S., happening in China, there's a desire to have in the Middle East, and there's political issues that come up right there because the U.S. wants most of the world's GPUs to be used in the United States and is wary of letting too many get deployed in other countries, both because there's concerns about China and competition with China, but also because the U.S. thinks it'll be really benefit to training most of the world's foundation models. If we had like a country leaderboard of who is the most advanced in a couple of different areas,

Starting point is 00:17:20 let's say foundation models and infrastructure, who do you think will be the largest movers up and down over the next, say, like five years? I think the surprise player right now is the U.S. The Emirates had put a ton of money into AI. Their leadership is very clued in on this issue. They've spent a lot of time studying it. And if you look at the Falcon models that they've developed, they're pretty sophisticated. And I think most people wouldn't have expected five years ago that the UAE would be front and center in AI.

Starting point is 00:17:48 I think actually this surprising downside player is China. If you look at the huge capital expenditures that U.S. big tech firms are outlaying right now, that's not really matched in China with maybe the exception of bite dance. And that's partly for U.S.-China reasons and controls on technology, but also for China's own domestic economic situation, is not making firms excited about putting out billions or tens of billions of dollars in data center spend and investment in AI. Andrew, can you talk about the moats in this business that people might have a hard time appreciating TSM's an interesting one to talk about because everyone knows what a political danger or hotbed this could be. Most of the world's cutting edge chips are built very far from us and very close to arrival. and it seems sort of insane that we just don't replicate this in the U.S. I know we're trying. But the fact that we haven't begs the question, why is it so hard? What is it about the infrastructure

Starting point is 00:18:39 businesses that makes them so hard to replicate or copy? It just seems crazy. It's changed over time, right? Because there was a period of time where Intel actually was at the leading edge and was the most technologically advanced manufacturer of chips in the world. And then TSMC, LepFrog, then that probably happened around 10 nanometer time when Intel, was struggling with that node. So call that. Gosh, that was probably almost seven or eight years ago, something in the zip code. And so, like, why has TSMC been so successful at what they do?

Starting point is 00:19:07 I think in part of it, it's because they're just maniacally focused on just one business model, right? All they are doing is the Foundry model, whereas the other players, Intel has a very large design business. It's designing the chips that will get made in the Intel Foundry. Samsung plays in multiple areas, memory Foundry. They also design some of their own chips. But if you're just singularly focused, right, on the task at hand and making the highest performance chips, right, and then learning from your customers because they have all the most sophisticated design customers at this point making their chips on the leading edge

Starting point is 00:19:40 at TSMC, it just gives them the ability to continue to improve and sharpen that manufacturing process to a level that I think at this point is be very challenging for anyone to catch up. I think the thing that strikes me is that ecosystem in Taiwan is just extraordinarily well developed. And we talk about the software ecosystem of Silicon Valley, but there's a manufacturing ecosystem in Taiwan. That's the chemical suppliers, the materials producers, the people who know how to fix the tool and the tool breaks. And they're all right there in a very small country within an hour and a half of each other on the high speed rail on the western shore of Taiwan. And so if something goes wrong in your factory, we need to prototype a new machine or new process, it's very

Starting point is 00:20:18 easy to get it right away in a way that if you're in Arizona or if you're in Japan, you just don't have that efficiency. And so TSM will say, I think they're right, that the bullet train that goes up and down Taiwan bringing people back and forth in a single day between their fabs is one of the key sources of their comparative advantage. Why is the yield so low in these things? If you're producing cars in a Ford factory or something, there's very few lemons, whereas you tell me what the numbers are, but I think for some chips, this is as low as like 70% of the ones produced actually then go get used in production, which means you're just like literally throwing away a lot of the product. Why is that still the case despite all that sophistication?

Starting point is 00:20:55 The manufacturing is the most complex that humans have ever done. So in comparison to a car, cars can have fault tolerances measured in millimeters for certain parts. In chips, you're talking about nanometers, so a thousand times more precise. And if you look at just the chip in your phone, for example, and pop it open, there are billions of tiny transistors, each one of which is the size of a coronavirus, and almost all of which have to work for your chip to function, to let you like things on Instagram or send a text message. And it changes every year. Every year, TSM rolls out a new manufacturing process

Starting point is 00:21:27 that is dramatically better than the old one, which means that the transistors are smaller or packed more densely together. And so the tolerances get even tighter year after year. And so the question of yield, how do you produce chips that work, means you've got to be yielding a new process on a regular basis.

Starting point is 00:21:42 So you sort out last year's process and you start all over again with the next process with even smaller tolerances for error. Andrew, I'm curious what you think like the five-ish most important companies are right now outside of NVIDA and TSM. I would put Broadcom on that list. I would put SK. Hynix on that list on the memory front.

Starting point is 00:22:02 I would put AMD on that list right now. I think when you look further down, you would put some of these other smaller component players that are in the AI infrastructure orbit as well. There's actually probably a longer list of those more than two or three. but you definitely have a whole group of companies that are benefiting from what I think is going to be one of the most significant infrastructure builds that we've seen. If you think about a company like ASML or something that seems pretty singular, which companies, if they just broke completely, would most mess us up?

Starting point is 00:22:35 Well, I think ASML is definitely on that list. Maybe that's a time to chat about EUV and all these different acronyms that people throw around and what that means because I think it ties into the yield question and why you only get a 70% yield or maybe even lower than that in some cases. But we were talking about at lunch how incredibly complex the EUV processes and what exactly you're doing. And we're going to use this laser to heat this tin to a hundred times the temperature of the sun. And then we're going to bounce it off these mirrors that are the most technologically advanced and smooth mirrors in the entire world. And then we're going to essentially the layer of complexity, I think the analogy is I'm going to hit a

Starting point is 00:23:13 golf ball from Earth and get a hole in one on the moon. That's what you're doing. And so when you're dealing with that scope, I think that kind of gives you a sense for why sometimes these chips don't come off the line and perfect. What do you think are the top concerns of the U.S. government? We can talk about Chips Act. We can talk about where we will accept external capital to finance companies or projects. Just by revealed preference, what does the U.S. government care the most about? And in your interactions, how on the ball are they? Like, we're talking the day after Gavin Newsom vetoed this bill in California, which I think everyone in technology is very happy about. How well calibrated are our politicians around this stuff? Do you think they've been effective?

Starting point is 00:23:53 What are they doing? I think there are three broad goals government has. One is to mitigate chipmaking concentration around Taiwan. Number two is to keep U.S. firms out in front technologically in every segment that's possible. And then three is to prevent adversaries, China above all, from accessing cutting-edge AI chips. And I think we've seen a lot of moves the last couple of years on each of these fronts, the Chips Act, which subsidizes domestic manufacturing controls on technology transfer to China. Government is relatively clued in. I think the people who are working these issues understand them well. But I think government works at government speed. The chip industry works at Moore's law speed. And so what really matters in the long run is actually that Silicon Valley, the U.S.

Starting point is 00:24:36 ship industry, can keep that exponential growth going. If it does, then the political issue, she'll sort themselves out. And if it doesn't, we have more problems in just political issues because our entire economy is levered to Moore's law. Why can't China catch up? If so much money, so many incredibly smart people, so much talent there, obviously the will seems like more and more of a centrally planned, more centrally focused effort, they can just say we're going to devote X amount of effort. We don't direct the private sector the way they might. Help me understand why they can't get to parity. And earlier you said they might be on the downgrade list. No one can do it on their own right now. The U.S. can't produce cutting ships on its own. Japan can't. Taiwan

Starting point is 00:25:15 certainly can't. They rely on imports of chemicals and software and technology. And so the fact that China's struggling to do it on its own is not a surprise. They're normal for struggling to do it on their own. In fact, they're trying something that's harder than anyone else has already tried, which is to actually be self-sufficient. The trend over the last 30 years has been to become less self-sufficient. We've collectively decided that we'd rather have a specialization. So Taiwan specializes in manufacturing, Japan specialized in the chemicals, U.S. specializes in design because there's huge efficiencies that come from that. And did we accept a bit too much specialization in the case of Taiwan? One can argue that.

Starting point is 00:25:50 But I think China's effort to do it all by themselves is much, much harder because they don't get the benefit of access to world markets and access to the whole world's expertise. Because this is such a burgeoning space, I'm curious, Andrew, where, if I got a room full of the semi-investors that you respect, whoever those people are, where you think your, personal views would most diverge from that group? A couplefold. I think there's been a view, you know, maybe take the biggest of them all, right, invidia, and what is the view on that company and how has that view changed in the broader public markets over time? Because it's interesting. A lot of active managers that I've

Starting point is 00:26:27 spoken with actually, Nvidia has not been a huge position for them, which is interesting because I think people always felt this is too expensive or this story has already been discovered. And so that's probably been one of the most, I don't know, puzzling or just interesting observations that I've had over time is surprisingly like few people actually own Invidi, despite it feeling like it's a very well-known story, right? And I think people talk about it all the time, right? It's in the newspaper all the time, etc. But I think people have this fear that, oh my gosh, the stock is too expensive or it's run, it's already up so much. But if you actually look at what's happened there, at least on a fundamental basis, over the last two years since ChatGPT really was released, the fundamentals have tracked.

Starting point is 00:27:07 debt in line with the stock price, right? And so, if anything, I think the earnings multiple is probably flat to slightly down versus where it was when Chatsybt was released. And so they've actually delivered on a fundamental basis, like over that time period. I think many smart semi-investors probably do own Nvidia, so I don't want to say that's necessarily the case. But I think broader in terms of the broader market view of that company has always struck me as a little bit interesting, given that dynamic. Can you talk about the changing cost structures and business models of the digital slash software world in this new era where it used to be very people engineering intensive. Now it's becoming much more compute intensive.

Starting point is 00:27:47 I know we talked about we want that cost to come down. But just the game on the field right now, the difference in the business models that you're seeing. This struck me probably like early 23. I was tagging along a meeting with the Maverick Ventures team, meeting with one of the frontier model players. And this was early in their life cycle. And we were getting their financial model for what the business could look. like over the next three years, five years, what the cost structure looked like. And this was one of these eye-opony moments where you looked at the cost base and literally 90% of it was compute,

Starting point is 00:28:19 right? So these are dollars that we're spending with Azure or AWS or GCP or Oracle Cloud and literally just for training these models. And that was another one of these aha moments. People don't appreciate the magnitude of spend that's going to be happening here over the next decade. The stats that people throw around, train a competitive model in 2022 was $10 million. In 2023, it's $100 million. In 2024, it's a billion dollars. In 2025, it's $10 billion. And then I think the latest that Larry Arleson was highlighting was $100 billion in 26,

Starting point is 00:28:55 just like staggering amounts. What types of companies can actually keep up with that level of spend to be able to be on that frontier model edge than lead? That's one area. just you're seeing just the cost of doing business and software in particular just drastically change where you had these old models that were very people intensive and you had all kinds of salespeople running around selling the product, very R&D heavy and programmers, etc. And is that now changing where just the cost of business is being much more compute and chip

Starting point is 00:29:23 driven versus people driven? I think you're seeing that shift in the hyperscalor cap X numbers, et cetera. Can you outline what you view as the key sectors within semis, whether that's design, whether that's manufacturing, whether that's whatever. What do you think over the key three, four, five areas that matter most? So the design piece is critical, right? And that's where the companies like Nvidia dominate, right? They're designing the best chips.

Starting point is 00:29:50 And then TSM is making those chips. And then I think what is interesting is that you're seeing as Moore's Law slows. And this is something that people are well aware of. There's been press talking about this for now for probably the last 10 years. where are the opportunities to create value as it gets harder and harder to get those Moorslaw benefits on the front end, so the initial manufacturing of that chip itself. And that's why you're seeing so much interesting innovation on the backend side, so the advanced packaging side of the equation, for example.

Starting point is 00:30:18 So, all right, we've made this chip now at TSMC. It's on three nanometer. How can we package that chip with other chips in memory, for example, and get that packaging to be as efficient as possible because historically people just hadn't cared about it because you're getting these massive cost downs, with Moore's Law. And every two years, doubling the other transistors in that chip itself. And so you're finding these areas that historically have been ignored that are now being,

Starting point is 00:30:42 in some cases, bottlenecks like Co-OAS is probably a term that many people have heard thrown around that was actually limiting the number of Nvidia GPUs were able to get manufactured. And so finding those areas that historically perhaps they have been ignored are now actually some of the most interesting. If you think about the emerging players, everyone's heard some of these big names mentioned a lot, A-SML, NVIDIA, cadence, and so on. What are the most interesting emerging players in your mind that will matter a lot more, like my rankings question, names that will hear a lot more in five years than we're hearing about today? I think the first thing you'd have to say is that you'd be surprised by

Starting point is 00:31:18 how short the list is relative to any other sector of technology. Look at AI and list the five most important companies in AI. Half of them were founded in the last decade. Do the same in semiconductors, and it's a very different ratio. I think that speaks to the fact that there are some big incumbents who play a very big role. And it's led to less company formation than I think you might expect, given that the history of the industry is actually in startups merging in Silicon Valley. Andrew, do you think about returns?

Starting point is 00:31:48 I know you and I've talked about this in the past. It's really just this is like a massive cycle that could be bigger than the ones that have come before it. And obviously, technology investing is about riding those waves and identifying them. How do you think about what that means for prospective returns relative to everyone's opportunity cost, the S&P 500 or something like that? At the chip player. It's interesting.

Starting point is 00:32:09 When we look at opportunities on the private side, this scenario where at least to date, you've seen valuations actually be quite palatable, much different than what you see in software, particularly over the last few years. And so I think the opportunities for returns are actually going to be quite attractive. Interesting in looking at this performance of the semiconductor index broadly, it's actually done quite well, I think something like 25% Kager over the last decade. And so I think when we think about the opportunity set in front of us right now, you look at the private world of semis and then you look at the public world. In theory, that private world should outperform the public companies,

Starting point is 00:32:45 given that there should be any liquidity premium, et cetera. And what we see in the market now is actually a pretty attractive set of companies that should be either large, independent players in this space looking out three to five years or will become important parts of many of the large existing incumbents, right? Because as you see it, very challenging for these public companies to merge with each other, they're going to have the need to do these tucking M&A strategies to really beef up their internal flagship efforts. Talk about antitrust. It seems like a really important I don't know, feature of the system when it comes to returns. If Nvidia wanted to buy all the companies that you backed and they're not able to because of antitrust, that's going to impact. That's going to

Starting point is 00:33:27 impact your returns, maybe. What's the state of play with antitrust and what's going to be allowed? We've heard a lot of antitrust last week with discussion of a potential Intel Qualcomm deal, whether that would face antitrust regulatory issues. Hard for me to know. There's actually two types of antitrust that matter in the chip industry. One is real antitrust from Western regulators, and two is fake antitrust review from Chinese regulators. And a lot of the deals that we've seen be sunk in recent years were from the Chinese refusing to give approval, not because there was actually real concern, but this was a part of the political game that they're playing with the United States. And so I think for every major transaction and even some minor transactions in the chip

Starting point is 00:34:06 industry, the question of Chinese approval hangs over. And it's something that companies themselves can't do much about. It depends on the state of U.S.-China relations and when's the next presidential summit. That's what matters more than the actual substance of the antitrust. Can you give an example of that? Just the case study of that playing out? The most recent, I think, was with Intel's effort to buy Tower semiconductor. So Tower semiconductor is a very small foundry company producing different types of chips than Intel produced. They're both in the foundry business, but towers a small company. And Western regulators all waved it through, and then China refused to review it. And why can't we just say, okay, why isn't there somewhere around that? I think this is a really interesting question.

Starting point is 00:34:47 And at some point, we're going to see a merger that's approved by everyone but China and the company is just going to say, screw it. But the problem is China is a pretty big economy. And everyone's got assets in China and everyone needs the Chinese market. And so unless you're in a niche where you don't care about Chinese customers and you don't have any assets in China, you don't know exposure to the Chinese market. You've got to listen when they say no. There was this meme going around earlier this week that in the late 90s, we really didn't have an internet bubble. We really just had a telecom bubble. Why might this not be of a similar character or nature to that? All right, I've got a lot of thoughts on this. The Kappex debate rages on. And I think that Gavin Baker did a nice job,

Starting point is 00:35:25 setting the state of play on this about a month ago. So let's go back to that bubble, right? We spent a lot of time thinking about it. I started at Maverick in 2004, right, as it was in the aftermath, right, of that explosion. And we pulled the data and looked at the biggest offenders and biggest spenders of KAPX during that time period. So the global crossings, the level threes, the quests, et cetera. And you look at the amount of capital that they were spending was just totally unsustainable versus where we are today. And I'll get into that. But the Kappex, so the dollars that they're spending, whether that be on fiber or Cisco routers or other kind of optical gear, was on average about 200% of their operating cash flow. So for every dollar that the

Starting point is 00:36:05 business generated on operating basis, $2 was spent on KPEC. So they weren't self-funding. They had to borrow in the high-yield markets or issue equity or what have you, but it was not a self-funding sustainable situation. In some years, we pulled the data and it was even more op-popping. It would be 500% of operating cash in a given year. And then you look at what was the utilization rate of that fiber in the ground. And I think as a view of like as late as 1999 and early 2000, only like 2% of it was actually lit up and being used. It was all on this build it and they will come mentality because the internet was in the early days back then. And so the thought was this is all going to get consumed in very short order. And it just turned out that it wasn't. And I think it's very different

Starting point is 00:36:47 from today where you look at the spending that is primarily done by the hypers, so meta, alphabet, Amazon, and Microsoft, what are they spending relative to history on CAPEX? And if you look at their operating cash flow in this case, so cash the business is generating, right now their CAPX is only about 50% of that, which is dead in line with where it's been over the last five years. So we are very much in a sustainable level where you could double. cap-ex and these businesses would still be self-funding. And you wouldn't have to run into an issue where they're having to borrow or issue equity or something like that to fund these builds.

Starting point is 00:37:25 Furthermore, when you look at the utilization rates of these GPUs in their clouds, it is nearly 100%. There's nothing sitting around idle underutilized. The dynamic, I think, is very different, looking at that telecom period of time where you had this massive amount of infrastructure that was built. It wasn't really built in a sustainable manner from like a financing perspective. it wasn't even being utilized, whereas today we can debate what ultimately will get generated by these large language models and the value that will get created. But I can tell you, the GPUs are definitely being highly utilized right now. So it feels like a very different period of time. So if I were to take your analysis and apply to today, the place that looks like TELCO is the

Starting point is 00:38:06 foundation model companies. They are spending 5x. They're way more. They're operating cash flows. They have negative operating cash flows. That does exist. So does that mean they're all toast, if we get to a $100 billion model, there's just not many places in the world. You can get $100 billion. And the very few places you can are the companies that you've mentioned that have these amazing operating cash flows. So does that just mean that this technology is like an incumbent dream and that these upstarts, even the opening eyes of the world, are in trouble? That's an interesting way of cutting it. And I think you look at the types of companies that could sustain that level of spend. That list is pretty short. And we talked about them just a few minutes ago.

Starting point is 00:38:46 So that's a very interesting way of frame it in, I guess, the implications in terms of who ultimately can afford these. The list is not very long. Well, it ties back into the Middle East, which we discussed before because if you're looking for very large pools of capital, that's one of them. Will that be feasible? Will the U.S., we have in various cases pretty close ties to the countries with the most money? Talk about the range of feasibility. I think the U.S. is pretty comfortable with Middle Eastern governments investing in U.S.-based companies to build U.S.-based data centers. And we've seen, for example, Microsoft Mobata announced a joint investment vehicle.

Starting point is 00:39:21 I think the U.S. is a lot more worried about investment in the Middle East because then the U.S. has less leverage, the Middle East has more. Certainly there will be data centers built all over the world, including the Middle East, but the U.S. government really doesn't want the most advanced frontier data centers being built in the Gulf rather than in the U.S. And so there's a lot of regulation coming down the pipe right now about that particular issue. What are you most worried about? I worry about the time horizon in which we see the return that Andrews talked about.

Starting point is 00:39:48 I think making sure that lines up is going to be critical to sustaining. If the scaling laws are true, the CAPEX scales up alongside the size, and we need real returns right away to fund that CAPEX. I think the next couple of years, we have a runway in which the hyper-scalers will spend regardless of return, and then at some point we're going to need to see real business models emerging. And so the time horizon for that, I think, really matters. I think the other thing, going back to the cost of compute, is the more we can do to find ways to economize and compute, the more sustainable all this becomes. There's a huge incentive right now to economize on compute because Nvidia is shown how large the market is to take a bite out of.

Starting point is 00:40:26 And the best way to take a bite out of it is to produce something that is just as performant or even more performant, but at lower price. Andrew, how about from an investing perspective? What worries you the most? You're obviously going to lay out a lot of capital to support companies trying to solve all these problems. What keeps you up? It goes back to some of the points that Chris just made around what is the sustainability of all this. And is this a situation where is there a digestion period? I've been doing semis for long enough to know that this is still a cyclical business, despite what people will say.

Starting point is 00:40:54 I think it's absolutely a growth cyclical business. And so I actually probably welcome any sort of down cycle because it's a great time to put capital to work. But I think just making sure that there will be a tier of new companies that are formed that are able to capture this broader paradigm shift that we're going through. And the challenge that the world, I think, faces today somewhat is you have one company, an incredible company that's captured the lion's share of the economics to date. And when I think about Nvidia and what they've done, this was a company that was initially a chip company, right, that then became chip and memory company that then expanded onto networking and now they're selling this full system. And what happens

Starting point is 00:41:37 if there's one company that's just capturing all the economic value of a transition like this. And you're going to want to have, I think, a very vibrant startup community that's helping solve some of those problems, particularly like in and around the GPU. I think going directly after the GPU itself, we can talk about that more too, I think is a challenging task, but there's no shortage of need for innovation, everything around that GPU. And so are we going to be able to create an ecosystem here, opportunity that exists? around all those things that we mentioned that are breaking earlier, around power, around interconnect, around memory, et cetera. And I think when you look at the big customers that we just talked about,

Starting point is 00:42:17 they are highly motivated to have a diverse and healthy silicon supplier base. And we were chatting about, let's look at the last big paradigm shift and what happened with mobile. And you had a situation where when Apple announced the iPhone, AT&T is the launch partner, AT&T had a market cap of $250 billion, When the iPhone came out, Apple had a $70 billion market cap. And you fast forward 14 years. And Apple's market cap now is $3.5 trillion. And AT&T is $150 billion. Apple just extracted all the economics out of that transition,

Starting point is 00:42:53 capturing obviously the hardware with the iPhone, also the software with the App Store. And I think when you think about the biggest buyers of a lot of this infrastructure right now, they want to make sure that they don't become, whatever you want to describe, a dump pipe, right? and they want to be able to capture value from this AI transition that we're going through right now and not that all accrued to one of their suppliers. I think it creates attention in the ecosystem when you have a supplier that has a bigger market cap than you,

Starting point is 00:43:20 that has more power than you, et cetera. I think there's a real incentive to have a healthy ecosystem. Generally speaking, what are the corporate strategy options available to those players that want to accomplish that? Just playing forward the Apple AT&T thing, how do you avoid that happening to you this time? Well, they're trying to develop their own chips. And so that custom-a-work is probably very high in the list of what they're trying to do. And I think they're very forward-leaning. We were having lunch with one of the hyperscalers last week where they have a team.

Starting point is 00:43:48 Their whole role of this team is just to be talking to startups in Silicon Valley, understanding what technologies are coming down the pike so that they can be put into their data center and to see if they can help solve some of those pain points that I illustrated. So I think they're trying to have a very wide aperture in terms of what they're looking at in terms of what other opportunities are out there, even for the incumbents, right? There's, I think, a strong desire to see AMD succeed in this market as well. I think we will see more and more success come from them as they catch up and are releasing new chips that are higher performant.

Starting point is 00:44:19 They figure out some of the software challenges that they've had historically. So I think that's another piece of the equation. So it's going to be a combination of custom chips that these hyperscalers are developing themselves. It'll be looking at other incumbents like the established semi-companies. And I think the third bucket is going to be the startup. ecosystem, providing value to help them diversify. I know you said maybe less interesting to go directly after the GPU and better to attack

Starting point is 00:44:42 the area surrounding. But let's say the GPU is the Death Star at NVIDIA and there's that one little arrow for Luke Skywalker to shoot the laser beam through to destroy it. What's your best guess as to where that vulnerability might be for NVIDA? I think the challenge will be that Nvidia's chip is very flexible and nimble. And that's why people like it. If the models change, the GPU can still solve the models of tomorrow. If there is a scenario, this transformer technology, for example, ends up being the structure that everyone is going to be building models on going forward.

Starting point is 00:45:13 That will potentially be that opportunity for companies for startups that are able to just isolate that one very specific use case and do it much more cheaper. That will be an area, I think, that could end up proving to be a challenge for them. Like, how does invidia respond to that? They can obviously make custom chips that kind of go after these very specific markets to date. They've chosen not to do so. Okay, I want to talk about the ways to stay informed in this very fast-changing world. And I'd love for each you to call to mind three specific people. If you only got to ever talk to three people with the goal of always knowing what's going on,

Starting point is 00:45:48 just imagine you could access wherever you want. What three people would you pick? I think it would definitely be someone at building out the infrastructure at one of the large hyperscalage. They're on the leading edge of all this. or you could say it even one of the frontier model folks. I think that's definitely at the top of the list. I think it's going to be someone at TSM that's on the leading edge of the manufacturing

Starting point is 00:46:08 front. I think it's definitely someone at Nvidia that's on the front. I mean, there's someone at Vida, right? You can say who that would be. Those are obviously three good answers. I think the other aspect of this that I look at it from is understanding the business model shifts. And that, to me, is something that you understand, not by talking at someone who's building the data center at the hyperscaler, but actually who's running one of the

Starting point is 00:46:27 hyperscalers because they're looking at this. saying, Nvidia is now bigger than we are, what's our approach to the fact that our suppliers now beating us in scale and what are the shifts we need to take to deal with that? I think that's one. I think the second is the capital allocation side of it. Where is money flowing from into? And I think you learn a lot talking to some of the smart investors in this space. And I think the third would be actually politics matters more here than it has in half a century.

Starting point is 00:46:54 And so you really can't understand what's coming in terms of who has which access to which market in terms of which technology can you sell abroad without considering the political side as well. What questions would you ask? Let's say we had those six people in this room with us. What would be the couple most pressing questions that you don't already know the answers to that you wish you could have an answer to? I think on hyperscalor build out, what are the key problems that they want to be solving next year, the year after? That will provide you a roadmap into A, what are they going to be buying? But also B, how are they positioning themselves in the industry. And to me, that's a hugely important question. That's where all the money is allocated

Starting point is 00:47:32 in each hypers' decisions. That's half the market right there when you're talking about AI. Yeah, I think the government piece that Chris brought up is really important here too, right? And just understanding what is the folks, particularly of the U.S. government and where they are going to be putting potentially restrictions in place or probably more importantly, what are they trying to enable, right? Whether that's domestic manufacturing or getting the startup ecosystem in the U.S. operating in a much more higher level, I think that's another key question and topic to explore. What about beyond the hyperscaler's half the market? What about the other half?

Starting point is 00:48:02 What's going on there? I think the other market that hasn't really taken off yet, but Will, is on the edge of networks. We're going to see a whole lot of innovation coming as we, in order to deploy some models in the data centers, some on the edge, and we're not sure what ratio yet. I think we're the very, very early innings of innovation in semis on the edge to be super power efficient to provide just the right amount of processing power for edge devices. That'll be different for cars, I think, than it will be for phones or for, wearables or others, unlike Data Center, where we're inning three or four right now for

Starting point is 00:48:32 Data Center, we're still in inning one on the edge. What does that mean? So does that just mean Tesla and Apple are going to develop better chips? Does it mean the end user devices are pretty concentrated, and it's the biggest companies in the world? They already designed chips because it's just them doing a better job of what they're already doing, or is it some new paradigm shift there, too? So in certain cases like Apple, obviously, they will most likely be making their own chips for their iPhones. I think when you look out a little bit more broadly, there's definitely going to be the opportunity for new companies to take advantage of this edge AI theme, I guess is how I describe it. The analogy, and this is actually interesting because it ties into Nvidia as well, where Nvidia has clearly shown that it dominates the data center. I think hands down, I think most people would not debate that.

Starting point is 00:49:18 Whereas I think on the edge, the story is very much unwritten at this point. and I'm not sure if the game has really even begun. And if I look at Nvidia and where their focus is, I think Jensen's pulling all or the vast majority of the resources to focus on data center piece of the business because that's where all the value is getting created right now. Whereas I feel like they perhaps even decrease focus on some of their edge products.

Starting point is 00:49:41 And you can see this even in their financial statements where their data set of business has grown at a rate that I don't think I've ever seen any company grow at that scale over the last eight quarters. Whereas on their auto business, for example, that business has largely been stable over the past eight quarters, whereas you've seen this explosion in the data center. Part of that is just it's not in their DNA as much to be offering low power chips, which is really critical for the edge when many of these things are going to be

Starting point is 00:50:07 battery operated and performance for why is the key metric. The joke in the industry is that Vidia makes a killer gaming laptop, but if I'm getting on my flight from JFK back to San Francisco and I plug that laptop in, the plane, the whole plane drops 200 feet because the power usage is so high. Is there an opportunity on the edge to create, I think, some transformative companies that can take advantage of, I think, what will be intelligence just moving from that data center into your phone, into your car, into your PC, into the broader industrial footprint that largely hasn't seen a lot of innovation in probably four decades. And so that's, I think, an area that's really exciting and worth explore. And I think the other point, just in terms of who's buying these things, is this whole rise of

Starting point is 00:50:52 the startup GPU cloud, right? Companies like Corweave, for example, that came really out of nowhere to build these just dominant franchises right now. There's one in the AMD ecosystem called Tensorway that's trying to do a similar strategy. And I think there's a real need for these startup clouds because they're just, they can move more quickly. And it's interesting that we view Amazon and Microsoft, these incredible businesses, which they are. but they're also at a scale that they just don't move that quickly. So there is the benefit of partnering with a core even, Invidi did like an ingenious job, really bringing them up to scale where they can get chips out six months before.

Starting point is 00:51:31 I think if you look at Corweave, they had their H-100s out in March of 2023, and you couldn't spin up an instance of H-100 in Azure or AWS until probably almost like six months later. And that's a lifetime, right, given how quickly things are moving. right now in the broader AI ecosystem. And so if you're able to be a nimble startup that can move quickly and align with a chip company in this case, Nvidia, or TensorFlow's case, AMD, there's an opportunity there to create a lot of value, both for the chip company and for that startup cloud itself, because you just get this product out into the market.

Starting point is 00:52:08 And these GPU clusters are very finicky, right? And so if you just specialize in making those work, there's a big market. On the edge, one more point there is that I think since we're in the early stages, we underestimate the amount of change in products that will emerge. It's easy to assume it's going to be AirPods and iPhones that are defining the edge in 10 years. But if you looked in 2000, what would be the primary use mobile computing? You might not have guessed the smartphone. And so I think we should be pretty open-minded about where will glasses be or whatever else.

Starting point is 00:52:36 It's interesting to watch Facebook for two years ago. The Metaverse seemed like a idea that wasn't going to materialize very soon. And now you won't call the Metaverse, but actually there's some interesting glasses applications coming out on the consumer side. And I think the other thing is the industrial side, which is less exciting because it's not something we're going to buy for our own use. But as Andrew said, there's been so much less innovation on the industrial side. And there are so many potential use cases for AI on the industrial edge. We were talking over lunch about tractors, having GPUs in them. I think that's probably just the first of many use cases for industry, which could be a very large market.

Starting point is 00:53:10 Meaning a John Deere that has a couple of Nvidia GPUs on it, anything that could benefit from the deployment of a cutting-edge AI model will actually have these cutting-edge chips, not cheap chips like they're in most things, on devices. And any sort of robotics in the future is going to need a whole lot of computing power. And maybe some of that will be calling back to the cloud. If latency matters, like in a self-driving car, but also in a drone, you probably don't want to call back to the cloud to ask it a question on a regular basis. You want enough compute on the device. What about power? Any commentary on power? There's all these crazy headlines about building more nukes and all the things that we're

Starting point is 00:53:44 going to do to power the data centers and any layer deeper insights on what's going on in the world of power. So maybe this would be a little bit of controversial view. My thought on that, so power's critical and there's probably not enough of it broadly, but I think there is a dynamic here where capitalism usually finds a way to work. And I remember if you go back to solar, for example, in the 2006, 2007 timeframe, there was a fear that we were going to run out of polysilken to make these solar wafer isn't sure enough, you know, capitalism finds a way capacity comes online and then there

Starting point is 00:54:14 was actually oversupply in that case. And so with power in particular, when I talk to folks that I would view to be experts in this space, I wouldn't say I'm an expert in power by any means. But I think there is a dynamic when you see these forecasts for data centers that are getting built. Is there a dynamic where a data center says that it's going to need pick a number 100 megawatts, 50 megabytes. And will the real use is actually be a less than what they're saying kind of nameplate capacity will be just because there's almost like a double ordering dynamic? When they're telling the planner in that city or that state how much power they want, they're going to give the absolute kind of highest number that they can so that in the case that they might need that

Starting point is 00:54:51 much, they can have access to that much electricity. I think the other dynamic, and this is an area we're talking about intelligence at the edge, the broader energy infrastructure in the U.S., I think people don't actually have like a crystal clear sense of how much capacity actually is out there and what the utilization rate actually is. And there's this concept of you're building to handle that peak load, but are there creative ways where in Atlanta, for example, in the heat of summer, when everyone has their AC on in August, can you somehow incentivize folks to turn their air conditioning to a warmer temperature and bring that peak level down?

Starting point is 00:55:28 And then all of a sudden, maybe there is actually more excess capacity. then people sort of a provocative statement because I think the mindset broadly in the investment community is that power, and I think power is very critical, but I think there is a dynamic where maybe some of these kind of shortfalls find solutions. I think the other dynamic is that in data center chips, power has not really been a key focus the past decade. It's mattered, but it hasn't been the case that the primary driver has been bringing down power consumption. And as a result, we haven't brought down power consumption. Whereas in mobile, most people 20 years ago would be shocked by the compute per watt that is in your iPhone because Apple and others have fixated on bringing that down or bringing up the compute per watt up because battery life is limited. And I think now, Nvidia is getting a lot of calls from its customers saying, hey, we'd like more performance, but also you've got to take power really seriously.

Starting point is 00:56:18 I'm curious how things get done in this world. This is like a side door into asking about the way that you're structuring Maverick Silicon specifically, the players you have around the table, the investments of your time and energy that you're making that will ultimately be called back to benefit the companies that you invest in. It strikes me as a world that you could count on two hands and pretty quickly get to players. that really have the power and whose influence and support you might need to be successful as a breakout company. Talk through that dynamic of how things are getting done, who matters, how you're cultivating relationships with them. It seems like a much more hands-on style of investing than maybe even Maverick has done in the past where it's like mostly a public markets investment firm. You buy some Nvidia. You don't have to do anything. You can benefit from it going up.

Starting point is 00:57:06 It seems like you're going away the other direction with a really hands-on approach. So how and why are you doing it that way. Yeah. It's funny, is we spent more time in this and pulled on this thread. I think we've gotten more and more excited about the opportunity here. And when you talk to many of the founders that we've worked with, there's three big pain points for a semiconductor startup. Paypoint one is getting the design tools. So that would be from a synopsis or a cadence. Pain point two is getting the right IP building blocks, right? Because a lot of these chips, you don't need to reinvent the wheel and everything. There's off-the-shelf IP that you can buy for various pieces of that chip. And then the third challenging bucket is obviously the Foundry piece in that initial tape out and then broader manufacturing.

Starting point is 00:57:47 Is there a way to work with the key players in each of those buckets? So EDA, IP, and Foundry to just ease the burden for these startups. And the concept that we use is that many of these startups are on the airplane. They're in the middle seat. They're in the last row of the plane. How can we get them upgraded to first class? All the large semi-player benefit when they're interacting with those Foundry players, the IP players, the EDA players, et cetera. And so it's been exciting to have those conversations

Starting point is 00:58:16 because many of those incumbents on those three buckets will benefit from having these startups in their ecosystem rather than one of their competitors' ecosystems. Not all these companies will be successful. Some of them will become big customers in which company will be the next Annapurna or which company will be the next Estera Labs. There will be others like that in the future. And I think those different incumbents are very excited about working closely with them. But it's been fun. You've worked with Jim Keller is a well-accepted genius of the field who's also entrepreneurial. Is that generally the description of the founders in this space versus the 22-year-old dropout YC types

Starting point is 00:58:54 or the sort of classic Silicon Valley Disruptor? Talk us through like the nature of the people building these things. I would say the vast majority tend to be much more experience. So these are individuals that have run divisions at the large public semi-company. right, that have had the experience of building out chips at scale, which is, I think, much different in the software where you have a much, I think, younger average age of your startup founder, maybe even 20 year difference or something along those lines. There's some exceptions to that. I know with Etch, with Gavin, does a fascinating work there. And I remember laughing with him the person I met. I was like, did they teach you any of this at Harvard? And he's,

Starting point is 00:59:31 nah, not really. But it is, I think for the most probably, that's probably more of the exception than the rule. This is going to be a much more experienced-based management thing that you're going to be investing behind. Who do you compete with for deals? One of the things I've learned talking to you is the interesting distribution of where capital is available for investing in, especially in private companies. And obviously, this can change really quickly and supply rises to meet demand and so on. But it does strike me that there aren't a lot of people doing in a dedicated way what you're trying to do in private markets for the SEMIS ecosystem right now. So who in a given deal that you're trying to invest in that you think is the best deal of the year, who are you up against most often?

Starting point is 01:00:12 When I think about the lay of the land of the semiconductor, probably semiconductor investing landscape, you have some very good early stage folks like Lipu at Walden, for example, Stefan at Sutter Hill. These are just experts in their craft right at that early stage company formation period. And then I think when you look at the very latest stages and even the pre-IPO rounds, You have some, what I would deem to be traditional public market players. Fidelity has got a very strong team, Adam Benjamin, leading that, that will come into those pre-IPO rounds. But in that middle section, let's call it Series B and beyond, where the capital needs really start to pick up, I think there's a real dearth and shortage of capital in that market.

Starting point is 01:00:53 And I think that's the opportunity that we're so excited about because as these capital needs to pick up for these companies, and that's right, is that tapeouts happening and the IP costs are starting to ramp because they're, actually starting to manufacture chips, the EDA and emulation costs start to pick up because you're testing that chip to make sure that it's working. And that can total $50 million. And there's just not that many players that are willing to commit that level of capital to these companies. I think once you're able to satisfy that need, the reward can be enormous. We've seen that with some of these recent semi-IPOs. I think just comparing it to software, for example, where I think there's a very long list of capital providers that are going after that market. I think it's, frankly, just much shorter when it comes to semiconductors. If you could fly anywhere tomorrow and have a meeting of your

Starting point is 01:01:40 choice based on what's going on right now, where would you go? I feel like the part of the industry that is harder to understand right now is in China, actually. Just went back earlier this year, the first time since COVID, but because of the politics, because the regulation, industries, that used to be very closely knit are being split apart. And so I think understanding, for example, Huawei, which is the AI leader in China, producing China's leading GPUs at a moderate scale. That is going to be the company that defines where China is on the AI. What's the scale comparison to like in the area or something? It's small right now relative to and they're trying to ramp up and the US government's trying to slow their ramp. And one of the

Starting point is 01:02:17 key questions, I think, over the next couple of years is will they succeed or will they face real supply constraints for the next couple of years? Can you tell the Huawei story, how it got to here and what's happens? It really is a fascinating company. And I think had it not been for the politics, we probably would see it like a Sony or a Samsung, a real electronics powerhouse. And in some ways, it still is. They started out in telecoms and networking equipment founded in Chen Zhen right across the border from Hong Kong and grew by selling first to the Chinese market, but very early on pivoting to international sales. And so they were able to learn from their customers abroad, improve their technology, some IP theft along the way. But was it more or less

Starting point is 01:02:55 and the average, hard to say. And it was only really in the last 10 or 15 years that the relationship between the Huawei and the Chinese state got closer. And now given the politics, Huawei's business model is to sell as a almost monopolistic producer into a Chinese market that is constrained both by the U.S. and the Chinese government. And so the company itself has had to shift a lot. And now it's the Chinese AI monopoly player. And maybe describe what the U.S. government did vis-a-vis Huawei just to complete the story. So the U.S. government doesn't want China to produce cutting-edge chips. And so Huawei works with SMIC, the leading Chinese chip maker to actually manufacture its chips. Smic is like the TSM of China. But Smic isn't allowed to buy the most advanced tools,

Starting point is 01:03:36 including the newer lithography systems from ASML. And so right now, Smic is trying to produce pretty close to cutting edge GPUs using second-tier machinery. And it's learning just how hard it is to produce nanometer-sized transistors using tools that weren't really designed for that purpose. And so I think they're a scrappy company. They are innovating on the fly, but they're also hitting real scale limitations, at least for now. When you think about the components of returns, you mentioned the entry prices can be reasonable sometimes in the space relative to some other stuff you've seen. What about just like the size they can get to, the capital intensity along the way that will dilute

Starting point is 01:04:17 you as an investor, if you're an early investor, and the paths to exit, those seem like the remaining three components. I'd love you to just like pine on each of those three. I think maybe we'll start with the last one in terms of the path to exit. If there's 10 companies, my guess is one or two will end up going public, the traditional kind of IPO route. I think the majority will end up getting acquired by the large established semi-companies that will be using these startups to help basically bolt on to their flagship effort. where they can buy an engineering team or they can buy a product that can really solve a product gap that they have. And these are companies that are generating collectively something like

Starting point is 01:04:55 $200 billion or free cash a year. So there's ample dry powder to pay for assets like this, even to the tune of one to $2 to $3 billion. And we've seen several examples of that happen over the recent term. That is how I think about the exit path. And then a couple of them obviously won't work. In terms of the capital needs and the dilution that's required, I think it's interesting because there's this mindset with semis that they are. very capital intensive, which I think is broadly true. I mean, each company's not making its own fab, right? They're making their chip at TSMC, but there are capital needs in terms of some of those startup costs that we talked about having a deep engineering team, et cetera, that are not

Starting point is 01:05:30 inexpensive. But I think comparing to software, yeah, software is not very capital intensive when you start. You can only have three or four people and you can do it all in the cloud and it doesn't cost that much to spin up the first product. But I think to scale, software is actually pretty darn capital intensive. It's just that capital intensity hits you later. In seven, it hits you early, but then I think actually the margin profile of a young semi-company can actually be far more attractive than software. And so you can look at a semi-company that is doing a couple hundred million dollars of sales and is operating at 35 or 40 percent operating margins relative to a software company that could be still burning significant cash,

Starting point is 01:06:07 which may be the right decision for that software company if the LTV of these customers that they're winning will justify that level of spend. But I don't think that you should kid yourself that software's not capital intensive. I think it's just when in that life cycle of the company that it hits you. So that's how I would frame that. What about the word bubble? When does the word bubble enter your minds, if at all? I think when people talk about bubble, it means you probably aren't in a bubble.

Starting point is 01:06:32 They usually one rule of thumb. It's usually only like afterwards. First rule of fight club. Yeah. So I think that's probably one just broad observation. Whenever people talk about a bubble, I think people were talking about this being a bubble like 12 months ago, right? And I think it's proven not be the case.

Starting point is 01:06:47 to date. And so I think that's just one overarching view. As I mentioned earlier, I've been investing in semiconductors for long enough to know that this is a sickle industry and at some point there will be a digestion period, but I will say that this is a growth cycle. And this is actually a very attractive end market to be putting dollars into. And I think if you look at where the industry structure is today and where the operating margins are relative to some of these other sectors in tech, it's very attractive. And so if you have the stomach to deal with, if there is a bubble or there's a digestion period or what have you, I think that it's actually can create opportunity because I think we know the endpoint is going to be up into the right. Is that going to be a straight line up there?

Starting point is 01:07:28 Probably not. But I think taking advantage of any down cycle, hopefully you can do that from a position of strength. Chris, if you had to study or encourage people to study any of the historical episodes in semis to better understand the future, what would you have them study? The PC revolution and Intel's role in it gives you an example of what it's like to be early in the stage of a growth cycle. The chip industry was always cyclical around DRAM, the memory chips that were in all sorts of products, computers, consumer devices. But Intel was early to the PC revolution and was able to ride that wave two and a half decades, three decades, of which there were cycles in the midst of it, but because more people were buying PCs, you didn't really notice those cycles anywhere near

Starting point is 01:08:09 the way you noticed the DRAM cycle because they were just less protected. And so I think that's the optimistic view of where we are right now is that it's a cycle, as Andrew said, heading up to the right, you can suffer some ups and downs. And I think that's probably the right analogy, because we're seeing an entirely new type of, or new category of chips emerged to prominence that we hadn't used before. And that's what happened in the 80s with microprocessors that were used for the first PCs. Really a niche market before then, and PCs made the mass market. How do we get to this state of the world where so much of the manufacturing is done overseas? I think there were two drivers of the shift overseas.

Starting point is 01:08:46 I think one was that U.S. companies naturally, in some cases, rightly, focused on what they were good at, which was design. And two is that manufacturing was cheaper overseas. Today, it's cheaper in Taiwan, not primarily because of labor costs. There's actually very few people that work in a fab, highly automated. It's cheaper because of the ecosystem dynamics. And those are very hard to reverse. And I think for that reason is also why companies are less excited about, investing a dollar manufacturing versus design, because if you're going to invest a dollar in

Starting point is 01:09:17 manufacturing, you've got higher costs and less developed ecosystem, you'd probably rather develop it, put it in design. Government's trying to change that with the Chips Act, putting $40 billion at work in the U.S., and that's a lot of money. I think it's also important to keep that amount of money in context. TSMC spent almost that much in CAPEX last year. And so even really substantial government money is enough to move the needle, but it moves the needle on the margin. And it's going to take, I think, a long time to really shift where manufacturing happens. What could most disrupt the trajectory of this whole story? What could be an example, like a new architecture, not a transformer, but something different

Starting point is 01:09:52 that re-kicks off a different kind of rate of improvement. Or you already talked about maybe the scaling laws level out and that could change things. What other things could happen that would drastically change the trajectory of the semis industry? Memory. One of the key limiting factors in performance improvements is the need to move. data back and forth from processor to memory. And so there's a variety of new memory architectures that are being experimented with. I'd emphasize experimented rather than prototyped, but a lot of

Starting point is 01:10:21 researchers refer to this as the memory wall. The key pain point is that your memory is too far from your processor and this slows everything down. And so I'm not going to bet on it next year or the year after, but down the road, we might need new memory architectures. Maybe just explain in a little bit more detail what is literally going on there. What is happening? And then if it were to get revolutionized what it would unlock? You literally have a memory chip and a processor chip. So in the case of an AI accelerator of your GPU and your high-bound-of-memory, and you need to move data between them, which means move electrons between them.

Starting point is 01:10:53 And the distance coupled with the speed at which electrons move, which is slower than the speed of light, for example, you could move it photonically and you increasingly do, creates friction in your training. And so if you were to design chips differently to have processor memory more intertwined or closer together, or if you switch to optical interconnects, those are all ways. you can increase the speed at which that happens. And the knock-on effect of that would be just a more efficient system. More efficient system, but also potentially changing.

Starting point is 01:11:19 Lower cost of compute. Lower cost of compute, yeah, and potentially changing who's good at producing them. I think that's fair. You have this dynamic in the industry where the GPU performance has been increasing at a much faster rate than the networking and memory performance. And so you're not able to get as much utilization out of that GPU as you would like because of some of these other bottlenecks, what happens if the networking piece and that memory piece can't keep up? Does that end up just slowing down just the level of innovation that we're

Starting point is 01:11:49 seeing in the market and the ability to train a bigger and bigger model? I think my expectation would be there's a lot of interesting activity happening in that networking and memory level to try to solve some of those pain points and catch up with the performance improvement that you're seeing with the GPU. But to me, I think that's a key issue to understand because if you're in a situation where you've got this incredibly powerful processor, but you can't get data in and out of it quickly enough, or you can't get enough data in and out of it. Does that end up being a bottleneck that slows down innovation? Just imagine that your whole portfolio is already built the general types of companies that you're going to invest in. On average, where do you think the source of sustainable power and differentiation will come from business power, not a literal power, will come from for those companies?

Starting point is 01:12:34 Is it mostly scale? Is it intellectual property? Is it cornering the best researchers? As you think about not just who's going to grow, but who's going to grow in a defensible way, where do you suspect most of that defensibility will come from? I think it's got to come from the intellectual property piece, writing the value that these companies are bringing to the table and the problems that they're solving right? Because the scale piece comes from the foundry players.

Starting point is 01:12:57 These startups are designing the chips from the most part, and then they're going to outsource that piece. So the scale gets solved by that third party foundry, whereas I think the value really is going to be created from the, pain point that company is solving and are they able to do it in a way that no one else can. Is that world like AI model research where frustrating how few people seem to really matter and it's like dozens of people that seem to have any real impact on frontier models? I think it's broader than that. To be totally candid, I think you've got, in Silicon Valley

Starting point is 01:13:26 in particular, right, you've got thousands of engineers, tens of thousands, probably frankly even hundreds of thousands, right, that are very technical? Are they all going to be startup founders? Definitely not. I think it is actually probably at a wider, then maybe what you're seeing with some of these AI kind of frontier model places where it seems like there's these 100x engineers, right? If you're able to have one of them, that solves all your problems. I think there are folks like that in the semi-industry, for sure. You highlight Jim Keller, his reputation, and then clearly precedes him. And there's others. But I think in broadly speaking, I don't think you have quite that same pain point. What do you two most commonly debate or disagree on?

Starting point is 01:14:00 I don't know if we describe. I mean, we talk a lot about government stuff, right, in terms of when headlines come out and there's a lot of news flow coming out of D.C. and just trying to understand what it means and what the intentions are of the various restrictions that are coming out or whether it's chips act related. What are the end goals that they're trying to achieve with both the fab side and the R&D side? Chris has got incredibly tight relationships there. I remember I told you I was in D.C. last minute and then you had me meeting with Biden's chip liaison. I was like in the West Wing of the White House like four hours later. So you definitely have those people really trust and respect what you have to say. But I think it's more of a collaboration

Starting point is 01:14:36 just trying to like, what does this really mean? Because I think you are probably more effective than I'm reading the tea leaves a lot of this. Yeah, the other facet, I think this is a fair statement. There are more different ecosystems that need to be brought together to understand the full picture. You need to understand the manufacturing, the IP, the end users in the data center. The supply chain is longer than a typical software company, for example. You've got to understand where all the players are to see where technology is going. That's a very interesting point, right?

Starting point is 01:15:03 Because I think you look at software that tends to be much more. siloed. And in the semiconductor ecosystems, when we touched on these different players earlier in the conversation, where like so many different folks just interlap and it's very interwoven, where everything from SemiCap, it's making the tools, to the foundry player, to the design folks. And even with design, there's these different buckets like mobile and compute and there's analog. And the EDA folks that are selling the tools that are designing the chips. And then you have the IP players like ARM, et cetera. But like everyone needs. needs to work together to push the industry forward. And so being able to see that bigger picture

Starting point is 01:15:41 and have the connectivity with those different players, I think is a really critical piece of successfully pursuing this market. And I just think it's different than what you see in software, for example. I was struck when I first started learning about the semiconductor industry. I figured there'd be someone I could call to get a view of the whole thing. And I quickly realize that there are brilliant people who know their slice and know a bit about the slice below and above, But there are so many steps in the supply chain that actually it's really hard, almost impossible to have a view of the whole picture. What are the most sensitive parts of the supply chain? The manufacturing step is sensitive, but so too are the tools that make manufacturing possible, the ASML level, if you will.

Starting point is 01:16:21 But then you can't do the manufacturing up the chemicals. And the chemicals are made in many cases by a couple of Japanese firms, which have the unique knowledge to produce chemical X with 99.99.99% precision. and that's only the manufacturing side. Then you look at design where the IP and the 88 tools are. It's hard to say one of them is more critical than the other because you need all of them. It starts this feel like a rackus or something in Dune. It's this precious compute resource that just fuels everything else. And there's a revelation going on in it that kind of we've never seen before.

Starting point is 01:16:57 How big do you think this gets? How big is this industry in five, ten years? And how big is it today? Today, the semi-country industry is probably $700 billion, give or take. Software is probably a little bit higher than that, like $800 billion. When I think about where we go, maybe this is a bold statement, I think the semi-eastern market will exceed the software market by the end of the decade, which is just given this flow of shift into compute from people into compute, as that being the primary resource

Starting point is 01:17:25 that's going to drive AI. I see the other thing that makes me optimistic about the future rate of progress is the application of compute to solving chip design and ship manufacturing problems. There's a sort of of flywheel of we make better chips, we produce more compute, and we use it first and foremost to make better chips and produce more compute. Whether it's at the chip design where the leading chip design companies have essentially been doing AI for chip design for decades now. You can't lay out a chip with 10 billion transistors that a whole lot of computing to do it for you. Or whether it's the lithography process, how do you know what shape to print? There's now extraordinary amounts

Starting point is 01:18:00 computation that go into the actual figuring out how to get the shape you want on your transistor or the shape of the transistors and your piece of silicon. There's already a huge use of computing in the design and production process. And I think there's immense excitement about ways to find even more applications for AI to make the chip design and production process more efficient. Anything we've missed that you think is critical to what's going on in this world that you spend time on that we haven't talked about? The other thing that stands up to me is that one might think the more advanced chips we need, the fewer less advanced ships we need, but it's actually the exact opposite. The more advanced ships we need, the more less advanced ships we need

Starting point is 01:18:39 to be in the systems around them. So take a new car, for example. You've got a couple of pretty high-end processors for the self-driving. But if you need self-driving, you need a ton of sensors around that. You need more chips managing the data transfer between different sensors. And so a new car will have more lower-tech chips than an old car will, in addition to having a couple of higher-end chips. I think that's an interesting dynamic for the entire industry because it means even if you're not in Vidia, if you're an analog devices, which produces a lot of the sensors, for example, AI is a really interesting transformation for you, too, because the sensors that you produce are going to be converting the data that will be necessary for AI training or inference. And you see that in basically every corner of the industry. Is that an investment opportunity?

Starting point is 01:19:19 That strikes me as something that is a far less sexy story and therefore perhaps more interesting price, but nonetheless similar long-term prospect in terms of. of market size? Yeah, absolutely. I think that kind of ties into that broader edge AI bucket. And we talk to a lot of these companies. They're more analog-based, right? And so not the leading-edge digital chips, where you try to make it very clear with them, what just happened in the data center?

Starting point is 01:19:43 That tsunami is coming for you next. And if you don't embrace change here and widen your aperture to be able to put intelligence into your historically kind of dumb chips, you're going to be on the wrong side of change and someone's going to come disrupt you, just like you saw. video disrupt the incumbent in the data center market. And so I think anyone who's on the edge should realize that that wave is coming for them next. If I were to force you both to leapfrog one or two degrees away from just core semis focus, what is going on in the world right now that most interests you?

Starting point is 01:20:16 The other place I'm spending a lot of time is in biotech genetics space, which I think not this decade, but next decade there will be a very interesting intersection of improved semiconductor technologies and biotechia. And I think if you spend time at places like leading universities and ask what are the PhD researchers working on, there's a whole lot of interesting work being done there, which I suspect will lead to company formation and new technology down the road. So I've been living and breathing semis here, in particular over the last nine months. So this is a refreshing question. I think just trying to zoom out, and this ties into the semi piece of the equation, but like how do you help create a more vibrant, healthy startup ecosystem for this.

Starting point is 01:20:58 set of companies for this industry. And how do you collaborate with folks in D.C., which I think, to their credit, the Chips Office is doing a great job, being proactive, talking to investors, talking to companies, not trying to be in that D.C. Echo Chamber. What can be done to deboddleck all these pain points that makes it so challenging for a chip company to start up? How do you motivate that very smart engineer who's at NVIDA right now or AMD or Broadcom to say, hey, I'm going to start this new company because I think the rewards of that are going to be both financially and personally massive. And so up creating that culture is something that I try to think about because I think it's really important for the U.S. broadly to being competitive on the design side where it dominates

Starting point is 01:21:43 today. But the fear is that does that slip away because you don't have a very active startup pipeline that's pushing the incumbents to innovate and keep moving this ball down the playing field? And so that's something that I think about a lot and try to figure out how can you collaborate with the different players in this ecosystem, whether it be the incumbents, whether it be the U.S. government or other governments to create a really vibrant, healthy ecosystem. Having read the book and talked to you a lot about what you're building at Maverick Silicon, I'm really excited to see the companies that start to come through here. Like you guys have put so much time and thought and attention into not just who to back, but how to help back them. So I'm sure the signal coming off of what you invest in is going to be a fascinating way. You'll be my answer to my earlier question about who to call and figure out what's going on. My traditional closing question for everyone, what is the kindest thing that anyone's ever done for each of you?

Starting point is 01:22:34 The caveat here will be this is excluding family. So this ties back to when we had lunch this summer and I was leaving for that trip, my first trip to Taiwan. And so if I go back to high school and college, the mentorship that I got from my teacher, and my coaches over that period of time. It's such a huge impact on who I am today. That probably is the kindest thing that anyone's ever done for me. And the reason why I bring up that Taiwan trip, so that ended up turning into two Taiwan trips over the course of eight days.

Starting point is 01:23:04 And I was at an airport hotel working out at the gym, trying to maintain some semblance of fitness, and got an email from my rowing coach from high school. It probably hasn't coached me on the water for over 20 years. Just, hey, Andrew, checking in. How are you doing? How's your family doing? And it just totally brighten my day.

Starting point is 01:23:22 And just individual still cares about me long after I left his program. I left that university. I left that high school. Just like really brought a smile to my face. And I think just trying to do that at Maverick where can we mentor that next generation of leaders of the firm and spend time with them and invest in them. I try to do that like people did with me. I don't think necessarily do it the best job.

Starting point is 01:23:43 But I think that's just like a really valuable lesson. I just encourage everyone take the time to do that because I think it's important. I'd mention a mentor as well, my advisor when I was doing my PhD in history, John Caddus, who often when you have a PhD advisor, your advisor wants you to study what they study. That's the standard trend. And John said, study whatever you want to study and gave me sort of an intellectual blank check and trust that I could find something interesting and accurate to say in whatever field I wanted to. And at the time, that was a niche area of Russian history, what I was studying them. But it, I think,

Starting point is 01:24:14 helped me realize that if you had the right set of tools, you could have confidence to go into a new area, understand what's happening there, make sense of it, and produce something of value at the end of it. And that initial willingness to take a bet and write that blank intellectual check is something that I value today and hopefully I've produced value from. The world's infrastructure is changing. Thank you for teaching us about what's going on. Thanks for your time. Thank you. Thanks. If you enjoy this episode, check out join colossus.com. There you'll find every episode of this podcast complete with transcripts, show notes, and resources to keep learning. You can also sign up for our letter, Colossus Weekly, where we condense episodes to the big ideas, quotations, and more,

Starting point is 01:24:53 as well as share the best content we find on the internet every week.

Invest Like the Best with Patrick O'Shaughnessy - Andrew Homan & Chris Miller - Redefining Semiconductor Progress - [Invest Like the Best, EP.396]

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.