SPECIAL: Doug O'Laughlin on the Semiconductor Industry Dynamics and Outlook

That's, yeah, thank you so much. That's very kind of you, Elliot. Definitely feel like I've known you for so long. I mean, we've chatted about names and specific things and all kinds of industries for years now. So, yeah, definitely, we've chatted quite a bit about companies along the way. So, yeah, let's just kind of dive into it, I guess. Let's kind of start with where you want to have some questions, and I'll let you direct the flow of it, because I could talk about all aspects of this space forever. So, yeah, let's kind of start there.

Yeah, so there's some context there that's really important. One, I think it's mostly because the diff... I'll briefly touch AI, I'll kind of go to analog and everything else, and I'll kind of come back. First and foremost, one of the reasons why AI is so hot and everything else is so cold is because I think, at least, let's put the winner in perspective: NVIDIA, right? NVIDIA has done an amazing job of really focusing on where it has massive differentiation and doesn't play in anything else. So, you know, just like classic competitive advantage, they are able to increase so much more price and specifically have so much more benefit to them without having massive benefit to their partners. Like, it isn't this whole ecosystem-wide thing in a lot of ways. They have many, many moats around their closed gardens, and so they specifically have had this massive outsized impact. And that's kind of percolating to a few other players, right? Namely Broadcom and Marvell. But this isn't your semiconductors of old, right? They're very custom. There are supply chains that help them, but because of the large order size, the people who are associated with them have, you know, relatively lower margins, relatively smaller orders. So let's kind of... that's the AI side, right? There's been a lot of differentiation, so there hasn't been quite as much breadth to everyone else that isn't an extremely differentiated winner.

Now, the rest of the market is in this extreme downturn, but we also have to have the context of why it's in this downturn, right? It's because there was an extreme upturn during 2020. So what happened was, and, you know, let's walk all the way back. In 2020, you know, we couldn't have cars anymore because we ran out of semiconductors, and people began to realize how critical these chips are for every aspect of their lives. I would argue most of the capacity additions in 2020, a lot of them happened at the leading edge, but a much larger percentage compared to history were at the lagging edge. And so all these chips that were pretty unsexy and hadn't had any new capacity built since, let's say, the late 1990s or the early 2000s had this giant spurt of capacity come online. And with that capacity, and demand was definitely there, but if you add a lot of capacity and anything gets a little softer than that, what happens is, as you know, the market turns. And so that's kind of the story, I think, right now is we are still kind of going through the post-COVID glut. And the reason why it's still a glut is because historically the lagging edge has had a much more stable supply additions, where most of the lagging edge capacity additions were essentially hand-me-down tools from the leading edge. And so adding de novo lagging edge capacity is what really threw the supply-demand part out of whack. And so I think it's really... it's actually kind of healthy. I do think the supply, or rather the inventory overhang, has been much longer than I would have expected at this point, but we are, I think, hopefully closer to the end than to the beginning. So, yeah, that's kind of the story, I think, between the leading edge, relatively small percentage of TSMC revenue, AI semiconductors, and then everywhere else, which is like the vast majority of, you know, wafers shipped, is in a lot sluggish, more sluggish space, mostly driven by capacity additions.

Yep.

Me, long and ugly. You bottom, right? I do think it's... and also I want to put some context in that. Let's specifically say automotive. Automotive was one of the cycles that really didn't turn until extremely late. Like, I think you can first in, first out, right? That's a really good way to think about it, because automotive was pretty much the last in, last into the cycle, and now it's taking a lot of time to roll out. Let's say the automotive downturn really only began in 2024. I remember the Mobileye pre-announcement was like the first big shot across the bow, and then everything incrementally was weaker. You know, it gets a victory lap that was a little bearish into that, mostly on the China capacity additions. But I think that's the real story, is that, look, these cycles historically are, let's say, six quarters long on average, and every single industry that went through the COVID upturn went through a much longer down cycle than history. And auto was still in an upcycle into late 2023. And so let's say the six quarters is kind of 2Q next year, or of 2025, sorry, we're in 2025, 2Q or 3Q of 2025. And so that kind of makes sense to me, because every single thing that had an elongated cycle has gone through an elongated downturn, and we've seen this, like, very consistently. Like, there was a few laggards, like FPGAs and autos were really hot, and meanwhile the rest of the market was turning. Now FPGAs and autos are, you know, in the dumps, and that just makes sense, right? Like, you can't have these inventory cycles go on forever without some kind of short, or sorry, a glut on the other side. And so that's kind of where we're at with automotive. I think industrial is a little harder because the demand story is extremely weak, right? Like, industrial does have very high correlation to global PMI, and we're seeing that industrial demand specifically is a lot weaker than we initially expected. And also, industrial supply is a lot harder to track, right? Automotive still has a few big customers out there, but industrial, the industrial customers, like, let's put it this way, like, maybe a smart sensing chip in your, you know, your hand drill, right? That has a long supply chain with thousands of distributors and inventory everywhere. And so I think the inventory and the tracking of that was a lot harder. And so during the upturn, people over-ordered a lot harder because there's, you know, thousands of different places that could go. And now we're seeing in the downturn where there is pockets of inventory everywhere. And so that's kind of where I think the, let's say, the auto and industrial cycle are. The analog, which is, you know, kind of very highly correlated to both, is maybe a little bit further along in terms of coming out of the trough. And that's like, you know, the players, ADI and Texas Instruments and stuff like that. Inventory extremely high, but we're starting to see gradual recovery there. Wanted to make sure I touched on that.

Yeah, 100%. And I think you can actually... there's a perfect case study to see where, you know, what would have been UMC and Microchip. Microchip is an IDM, an integrated device manufacturer in the United States. Microchip is actually refusing CHIPS Act money. They're just saying, actually, the capacity that gets added from this subsidy is probably not going to be worth it, even with, like, 25% off on the tool and, let's say, like a $100 million grant. It just doesn't make sense for us anymore. And I think that is the ultimate... that's the ultimate signpost where you're refusing free money because you're like, actually, I don't think I need this capacity. And UMC, for example, which is, I don't know what to think of it as the lagging edge TSMC from Taiwan, they are pretty strong in calling out a global glut in demand or in capacity. And part of this, I think, that was maybe underappreciated was the fact that China specifically is the king of lagging edge, right? China is trying to become like the rest of the world and a leading semiconductor company or country, rather. And in order to do that, what they did is they started at the beginning. So they started with much older technology, and they've been adding massive amounts of capacity. I think most of the capacity that's been added there is strategic, and from a top-down level, no one has to make any money, and they're mostly just figuring out how the processes work at all. And so they're just throwing, you know, hundreds and hundreds of thousands of wafers at the problem. But by doing so, they've created massive amounts of excess capacity. And so I think that's kind of, you know, in a capitalist consideration between supply and demand, you have this one extremely large player who's like, I don't care about this at all. I'm just going to add as much capacity as possible in order for me to learn how to make a chip better. And so that consideration for them has kind of completely blown up the dynamics, I think, on the lagging edge specifically. So, yeah, it's really ironic. Like, I remember thinking, like, oh, lagging edge is going to be so much better in 2021. And fast forward three years, I don't think that's the case. Lagging edge is a real supply issue, and I think we probably added enough lagging edge capacity to digest for the next five years. And then on top of that, if it goes too well, right, like China will really backfill into the void there. Like all Chinese domestic production has massive incentives to use Chinese semiconductors. And so that's kind of the problem, right? Like you have, like, let's say, a cartel where the bottom player is extremely willing to add capacity, and so you kind of break up the whole thing.

So, that's a really interesting point, but actually I want to point out the industry structure dynamics there. So first on the consolidation point, we've seen consolidation on the lagging edge happen secularly. Like, this has been... there's a really good chart from, I want to say, Onsemi that I don't know if they publish it anymore, but it used to be like how many semiconductor companies, I want to say it was like hundreds, and now it's like we're in the 10. Okay, it's been a long-term secular story of buying your competitor and rationalizing your supply and then, you know, pushing price and not needing as many people to design. So that was a story that already was happening. I would say this definitely accelerates it, but on the flip side, I think the United States government as well as international companies are not willing to let semiconductor companies bid for each other. That's like become a... that's also like a relatively new thing. I think since the Xilinx deal, there hasn't been a deal done, and every single deal has been blocked. Like, Tower Semi got blocked by Intel. ARM obviously got blocked. I would say the under of a deal going through, it's extremely unlikely at this point. So I think it would happen if it could, but the reality of geopolitics is making that extremely unlikely. Now, if we're talking about fabless at the lagging edge, that is a really interesting point that I actually want to point out that there really isn't many fabless at the lagging edge. A lot of the lagging edge players, let's say like a 32-bit MCU player that makes, like, you know, 90-nanometer chips or 45-nanometer chips or something like that, more often than not, they own the capacity. They own the fab, and then they also own the device. They're an IDM. So like the STMicros, the Microchips, the Texas Instruments, they more often than not, they are, you know, they do have some fab-light strategy, meaning that, like, let's say Texas Instruments, I think makes about 70% of their stuff internal. You can go look up the actual on the website. I don't know the actual off the top of my head, but over time, what they've done is they're trying to have more internal, so they're kind of integrated as is. But historically, what these companies have done is they've actually only used TSMC or UMC when they need to. And so I don't know, I would argue, yeah, that definitely helps them, but all the big consumers who are quote-unquote fabless, like, don't really exist, or they already have some amount of their capacity as internal. So there's like kind of this... that's kind of like a constraint that I think is important, is the industry structure for the old, for the lagging edge is very different than the industry structure for the leading edge because there's different constraints. Like, leading edge design costs billions of dollars. Like, it costs, like, you know, multi-billion dollars to get a 3 or 2-nanometer tape-out. Just the amount of EDA tools, the people we have to hire, all the testing, the verification, the IP that you have to use from other vendors, it's extremely costly. That dynamic does not exist on the lagging edge. You can use a relatively small team to spin up a relatively old chip. And so there, it doesn't make sense to really split the design versus the fab because the design part of the equation is so small, it doesn't... you might as well in-house it. So there's like an industry structure difference that you have to consider for, like, you know, the advantageousness of being a lagging edge fabless company.

I think it is... okay, so there's two parts of that question, right? The AI sucking every air out of everything else in the space, and then versus, like, how bad are the lagging edge. I think, okay, so AI is definitely sucking everything out of the space first and foremost, right? We have to acknowledge that, right? Like, I would say you can argue that pretty much every single, like, let's say, marginal company is trying to push for a big AI content story. Like, a good example maybe is Infineon. Infineon is historically known as this, like, ginormous European automotive semiconductor company, and they're really, really excited about, like, an AI power content story, right? So that's a good example where these lagging edge guys are, you know, who's historically been focused on something else, are very starting to focus on AI as well. You can even argue there's a lot of players who kind of fit that bill. So there definitely is some aspect where even the lagging edge players are really focused on AI because of all the incremental content growth, and that makes sense. So pretty much if you're a player that isn't participating, you're punished on a multiple basis, and if you have some kind of incremental content story, you're rewarded. And so the more you get to that, the better your stock trades. But I do think that there is, you know, I historically love to be a bottom feeder in terms of semiconductor cycles because as sure as it is in the dark, there's like this really great quote from the Microchip CEO who just stepped back in, is like, it's never as bad as you think in the darkest hours. It's obviously never as good as you think in the upturn. In the upturn, you're like, wow, we'll never hit this demand, right? It doesn't matter how much capacity we have, we will never ever be able to hit the orders that we're getting. And obviously you're being double-ordered. And then, you know, it could be literally four quarters later, you're like, wow, we are never going to fill our fab, right? Like, we are, you know, we're totally screwed. We're never going to get rid of this inventory. And so that pendulum always shifts. And right now, I think not only is the pendulum, we're never going to get rid of this inventory, but the pendulum also is just disinterest. I think that especially on the smaller side without an AI story, pretty much has been an area where I think no one even cares anymore. I would argue that I definitely am like the type of person to play in the, you know, or like care about the smaller names, and even I'm struggling to have interest. So I do think there's like a lot of opportunity. Some of these companies, I think, are getting so cheap that strategics feel like pretty warranted. And I just, yeah, I just think that right now no one cares, right? Like automotive is so dead, it's so over. Who could care about an automotive story? Meanwhile, three years ago, two years ago, everyone was talking about like an automotive 15% revenue CAGR till 2030, and it was the most secular, most advantaged part of the semiconductor cycle, right? So you have to like, you know, the narrative shifts very quickly. And I handedly can't look at the narrative and be like, wow, interest rates are so high, and, you know, there's so no one's buying cars, and the economy, you know, is it on or is it off again? Like all these kind of stuff. But I, in my heart of hearts, I know that in four, six, eight, ten, twelve quarters, at some point the cycle always turns. And when that happens, you know, from extremely low inventories, you have to have orders to backfill it. And then all of a sudden you're looking at extremely strong revenue growth for companies that trade relatively, you know, very modest, you know, multiples. And you can be like, wow, that actually is an amazing opportunity. These are, you know, obvious longs. And I think that you have to be interested and wake up and stay in the game when no one cares, because that's like the thing that will happen on the other side. So I mean, we're seeing some pretty, I mean, you know, I actually just got back from vacation, but like we're seeing some pretty drastic moves, and especially in automotive, like it's very much giving it's, you know, it's over. No one's gonna ever buy a car again, right? Like meanwhile, someone will have to buy a car eventually. So that's just kind of my intuition is especially in the automotive world, things are extremely bombed out.

So that's a good question. I hilariously enough have been spending so much time on each company, I feel like I understand the shape of what their differentiation is. So I don't usually think about like a really simplistic heuristic. Like, I mean, I think the thing that I think of over and over again is, and like especially in the semiconductor world, it's like, I know this is like kind of lame, I sort things by RSI, okay? Like what is the most oversold, what is the most overbought, right? I just look at what's the most oversold with a story that I understand is secular and their positioning is decent, and that's like historic. And like, you know, I also would love to be really timely with where they are in the cycle. And then understanding like, I think that like, let's say quality on ON Semiconductor specifically is kind of based on their products and their SKUs and who their customers are. So, and then also I would like say free cash flow. Like let's give some examples and just like, let's say automotive semi, X, like the analog guys, I would say it's like STMicro, Infineon, onsemi, I think I'm probably, oh, and NXP. So I think one of the ways you should think of quality is like who's actually made free cash flow the cycle, right? Like that's a really important thing. So for example, STM, no matter how much you, and like I really can't explain how weird of STM is as like a company sometimes, earnings great, free cash flow not so good. They really don't make that much cash flow, right? And that's like, you know, that's the bottom line of what makes a company valuable to you. ON, I would argue, has historically been in that same bucket where they really never made that much free cash flow. Now there's been this new CEO, Hassane, who's really done a really good job of turning around the company. And I cannot explain how hot of a hand he's had. He's literally like played every single hand perfectly. He moved to sell off like his lowest margin business when, you know, during the up cycle. And so now the down cycle's come and he doesn't have the lowest margin business. He sold, he bought like a silicon carbide fab for like, you know, pennies on the dollar out of bankruptcy, and it's been working really well. He's done all these really good moves. And so I think that that's an example where like historically that thing's traded with STM, and it should deserve a premium, right? Like there is a premium for people who are good at operating the cycle. Then you have like NXP and Infineon who are like the much bigger solid giants in the space, and they consistently make free cash flow. They consistently buy back shares, probably like, you know, much less risky, and they're like, you know, they don't, they're not, you know, weaving through the traffic lanes quite as much. So I think that that's kind of my framework at least, like free cash flow and how much they make over the cycle is probably like the biggest indication of quality, right? Like because everyone's going to lose money or maybe get like less profitable at some point, but if you're consistently doing a higher watermark than everyone else, it tells you that you have the right cost structure in place. And I think consistently the ones who are the worst actors in the space really don't make that much money along the way.

Yeah, still is, still is.

Like the fiber build for this round specifically?

Yeah, this is, you know, the very, very, very big question, right? I don't know. And I've actually written, I feel like I've written so dang much on this now. I don't really know. I think that there's many, many levels to this, right? Like capital cycles are important. It does seem pretty clear to me that our current data center infrastructure is just like completely not fit for what is the job to be done, right? We really are at this point where, and you can argue, and like a lot of the arguments for some of the big capex expenditures was like, yeah, I know we overlaid the fiber, but how did, you know, Facebook only work because we had so much excess capacity, right? Great. So I think the capacity that is being added is needed, but the problem is, you know, I think of it less as like, oh, it's a clear bubble versus it's a supply and demand line that are trying to find the equilibrium, right? We know demand is large, we understand that we need to invest in supply to meet said demand, but we don't understand really where the demand line is, right? Like it's far away from supply, but and it's growing, but so is supply. Supply is like, you know, as you know, like capitalism, supply and the ingenuity of people is like almost a limitless source of new innovation, right? Like pretty much in the same way the telecom boom happened, like there was supply, right? It wasn't doubling every 100 days, but people did want to use a lot more internet, and so the entire industry got really smart on how to add more and more and more supply. And that's kind of what we're going through today. Like we clearly have a lot of demand from AI, like the models are just orders and orders and orders of magnitude more compute intensive than what we needed, right? Like let's kind of take the Web 2.0 version of the market, right? Like fetching a data, like using Facebook, right, was just essentially like, hey, there's a database on the other side that says, here's your post, and if someone liked it or not, you literally just fetch that piece of data and it comes back, and then it shows if it's a like or not a like or whatever status they had on the screen. That is trivial amounts of compute. Like a CPU can do that kingdom come. Like, you know, one small web, one small core in a big CPU can host tens and tens and thousands of these requests back and forth, completely trivial. Whereas what we're doing with AI is, you know, we're running through these like billions of parameters, tons of data, all this stuff, like finding all the latent spaces and like all the little relationships between all this data. It is just inherently much more complicated and much, much, much more intensive. Like, you know, I hate to use this analogy, but it's like one is like a horse, you feed it some grass and it goes really far. The other is like actually where it's like a train, you have to do all the infrastructure, you can go tens and tens and tens of miles. It's orders of magnitude more efficient. Like you're able to move so much more power, it just costs so, so, so, so much more. But at the same time, so does the GPUs. The GPUs are, you know, 100x improvement over CPUs. The networking is, you know, four or 8x improvement in scaling out all this stuff. And so this supply-demand relationship is we blindly add supply until we hit the demand curve that we don't know or see or understand. And so at some point it crosses, and when that crosses, it's like it's all over, right? And so that's how I think about this whole process because I think it's not, I don't think it was, and like let's even look at like old capital cycles, right? Like railroads, I don't think it was unreasonable to add all the railroads onto, you know, into North America when there were frontier towns, right? Was it unreasonable how the speculative bubble that happened around them? Yes, 100%. Speculating on land value during that part in process, probably extremely unreasonable, and was one of the things that made it such a bad bubble. Same with telecom, right? Like I think the, and I literally just wrote about this, like I think the thing that is like kind of weird to me that we're missing is there's a difference between like needing capital and building a capital cycle versus a bubble. And I think one of the things that I like identify with, at least based on history, is like we don't have this like extremely fast feedback loop that we did in the telecom bubble, right? When you, when essentially WorldCom said, hey, we're adding a new wire, we're adding a new wired connection to, you know, another country, the stock would go up 30%. And so you'd be like, oh, clearly it's very valuable, right? Like there's that completely quick feedback loop that it's extremely, extremely valuable to have all this supply growth that's not happening in the markets. Like we're not seeing a company say, hey, we just got 100 megawatts and the stock goes up 30%, right? Because if you had that fast feedback loop, I think it would create like the, like that's like the important mechanism of a bubble, like it loops in on itself so that you reward capacity additions at all costs. And today I think we're not quite there, right? We really have, let's say, five extremely large players who are adding capacity. They can, they're all self-funded on this capacity. They're not raising from external markets, right? They're not raising, you know, billions and billions of dollars of debt. And yeah, maybe they're adding too much capacity. Maybe the ROIC doesn't end up being quite as good as it was. But I do think they will make a return on set capital. Now, is it going to be their historical monopoly, you know, like internet monopoly level profits where they're making, you know, high, high, you know, 50, 60% ROICs? Probably not. This is just much more capital intensive. But are they making 0% ROICs or are they just, you know, throwing away money? I don't really think so. And so that's kind of the part that I really struggle with is I do think, but maybe the part where it's like a little bit harder to understand is the adoption, right? Because you have these expensive models, they're extremely valuable. I use them every single day. I'm sure you are, or you should, candidly. There's so much data you can, like, it really does help with information processing. And the business models are not really clear to me. Like, how do you, how do you convert capacity and supply into, into, um, into a business model where you make revenue and money on the other side? Uh, it's clear that there's demand for it. ChatGPT is a top 10 website, and there's billions of, there's probably going to be, I don't think there's billions of users, there's hundreds of millions of users, there's probably billions of, there's probably a billion free users, like people use it all the time, and they're willing to pay. Um, does that support the capex involved? I don't know, right? That there is demand, I don't know if the supply is too much or too little. I think it's very, very hard for any of us to prospectively understand that. Sorry if that was extremely long ramble.

Yep.