EP 182: AI Fatigue , Meta's Neo-Cloud Rumors , and Samsung's Foundry Comeback
59m 17s
The discussion centers on recent semiconductor market volatility, driven by investor fatigue with AI stocks and an influx of inexperienced "tourists" that pushed valuations beyond fundamentals. The sell-off lacks clear catalysts, pointing to a sentiment correction rather than fundamental news. Social media platforms like Twitter and Substack are increasingly influential, shaping stock narratives and affecting prices, though they don't control them. Index-based investing compounds the issue, as fund managers must buy stocks to match benchmarks, even at overvalued levels. The conversation then shifts to Meta, following a Bloomberg report that it plans to lease excess GPU capacity to Google. This raises questions about overbuilding and monetization: Meta may have temporary spare training capacity due to model pauses, not inference, and could lease it for short-term revenue while still focusing on using AI to enhance advertising ARPU. However, the lack of a clear monetization plan for massive AI CapEx is troubling, suggesting spending without a strategy. This dynamic could signal broader industry challenges if other hyperscalers follow suit, potentially impacting the entire AI trade. The hosts emphasize the need for fundamental analysis and caution amid uncertainty.
[MUSIC] >> Hello, everyone. Welcome to another episode of The Circuit. I am Ben Beharen. Greetings, programs. I'm Jake Oldberg. >> All right, so let's start off with what's happening in the market? And I know there's like two things. Obviously, we are at the six-month time frame, so I've heard people articulate that there's been rebalancing right six months that goes on with some funds. And there's some of that. There's also continued, what do you want to say, skepticism, I guess, that just kill lingers around a lot of these markets. But that's not, I don't think, the blood bath that ensues. So like, what is the six-month read, I guess, where we're at on market sentiment? Like, what's our six-month take, I guess, and what's structurally happening at the moment? >> Yeah, so the market had a bad, semi-stocks had a bad week in the market, very bad week. And this comes off what two weeks ago, wait, it was last week, it was also not a great week. Last week was not a great week, but then, Mike Ron kinda saved it, but then Mike Ron actually ended up down. >> The week? >> Bad, yes, very bad. >> So, and now it's down a lot more. And I think there is an element of just fatigue with the AI trade. I will say that when I was at Computex, I think we touched on this before, when I was at Computex last month, I was hanging out with specialist investors, people who just focus on semi-conductors, right? Not generalists who cover the whole market, but people who have been trading semis, deep in semis for years. Thank you for listening. Right, and one comment I kept hearing over and over again is that people were a little bit nervous because there were, quote, "too many tourists in the trade," meaning people who were generalist semi-conductor, generalist investors, new to semis, also especially retail investors, very new to semis, not used to ups and downs. There was just a sense that semis had done so well this year that there were a lot of people sort of piling in in, sort of, March, April, May, that were driving valuations to places that couldn't necessarily be supported by fundamentals at the time. And the market apparently agrees with that sentiment, and things sold off this week, right? And of course, it sparked a whole set of debate about, "Oh, it's over, AI's a bubble," we'll start hearing that again. I think we can debate what's going on, but I think a lot of this just was, there was a perception that semis were going to go forever, and a lot of stocks probably got ahead of fundamentals. I think that's what's driving this, 'cause there were no other clear catalysts in the market, right? It's been trading through the war news and the interest rate news. AI stocks have done well through all that. There was no real news this week. I think it was very much a change in sentiment. Maybe you're right, it's sort of the six month rebalancing, but usually you don't see that until, it started before the end of the month. Usually you don't see that kind of stuff until after the end of the month. So, I don't know what's going on. I don't know what things are worth. I just think I think there was a lot of momentum, a lot of sentiment in the market, and now that's one of revising itself. - So I think the, you read on the tourists and the retail, I think is particularly apt. I've heard this too, and I think this comes back to what we've said for a long time, like even if it's not a fundamental bubble in terms of the build out, right? The thing that can impact anything is if capital gets spooked and retailers get easily spooked, right? So would perhaps your semiconductor tourists. And so, I think there's two ways I think about this. And there are a lot of companies who I do think are, are having their valuations or whatever's going on on that dynamic of retail and investors. And then there's others where I just wonder if, like the play is just be convinced on your fundamentals and stick with some of those, 'cause there's a lot of healthy companies out there, right? Generating free cash flow, still growth. I mean, Nvidia's an example of this. I know they've stocks done nothing, but if you just did an analysis of on their fundamentals, they'd be pretty good, right? Maybe even micron in their 4PE despite, there are a lot of tourists in that. So I think there's the, and I think this again always comes back to like, do you play the company or do you play investor mindset? And you've kind of got those two dynamics that I think are happening. With the other that I think is interesting is, I mean, maybe my read is wrong with this. And I'm over indexing what goes on in the Twitter sphere. But I do wonder if that is becoming more of a influential pocket that does impact kind of the daily of some of these names more than anticipated. Like, I know retail's heavy there, but like, are there other big funds? Are there other, I know people, you know, lurk on Twitter that might impact some of this again. I don't know, not all necessarily traders, but I guess this my point is like, does the news cycle or the things that go on right in Twitter play more of a role here? And then that's just something that, you know, you have to pay more attention to. I think that what takes place on Twitter absolutely does affect stock prices. It doesn't control them. Right, yes. I would say there's more bullish sentiment on my Twitter feed this week than bearish, even though there's a lot of people defending AI stocks as they fell. But absolutely, it's influential. And you'll see there's, I mean, there's obviously been a huge rise in new model sell side research, right? There are all kinds of new, very influential, sub-stack writers, people who are on Twitter. Those people have absolutely have a big impact on the conversation about semi-s. Again, they don't control, I don't think they control the share price but they absolutely affect what people are talking about. And to, you know, a lot of cases, it does, you can see that somebody posts something and the stock reacts pretty quickly. But even that, I think, has gotten a little bit, pushed to an extreme. There are certainly a lot of pretty dubious accounts out there right now. And so we have to, everyone has to be very, very careful but absolutely, you know, you think there's a lot of really good content that's come up, especially around semis in on sub-stacks, right? Obviously there's semi-analysis, but there's also funda and satrini who do really, really good work. And I think the investors, certainly professional investors, and probably you're probably right retail too, follow that pretty closely. - Yeah, agreed. And I think, you know, again, that's not a dynamic that's going away and obviously like you look at how much could it really swing, right? We're talking a couple of percent here and there and that's not out of the realm of possibility. But I do just wonder, like, I don't remember you saw a while ago there was a, I forget who said this. So apologies for blanking on the accreditation, but somebody was pointing out how you have a lot of kind of more legacy investor firms starting to invest in semis. From Europe and some others, right? Some of the kind of laggard's not the early adopters. And while big and kind of helping to cycle, right? They may rotate differently, I guess, when you're coming at this a little bit later and maybe not necessarily your core competency but you're seeing more people get into the space. So that in essence, impacts some volume swings, I guess. - I think there's some, I think there's a fair amount of market structure issues taking place and we're gonna get a little in the weeds here. I apologize, but like, if you're managing a sort of broad based fund, based
on growth or just a generalist fund, you're going to be benchmarked against the index. And so if you are, right, you don't necessarily make your decisions on like, oh, I'm going to really research this one stock and go heavy on this. You are saying, all right, this is the index. Where do I stand with my positions on all those stocks relative to that index, right? And so it's an underweight, overweight decision rather than, here's this incredible new thing I'm going to go buy. And as everything sort of ran this year, if you're an index benchmarked investor, you have sort of been forced to like buy things even if you're not comfortable with the valuation because you have to be close to the index. And as especially true now, there's so many ETFs and sort of passive funds out there. Those things are going to buy just because of the mechanics of how the index is constructed. And so I, right, and certainly I think there are investors out there who are buying more of certain stocks because more than they're comfortable with, but they have to because of mechanics around the index and around their benchmarks. I've heard that sentiment of fair amount among professional investors lately. And that's kind of how things got out of, there's one of the ways in which things may have gotten out of whack. Right. Well, I mean, I think again, right? It's not peak even though it feels like it's harder to move things forward, right, without sort of sort of a drastic unlock in either very large numbers of growth come in across the board. And I continue to see just, which we'll talk about with Meta in a moment, like there's just this still real sensitivity to just how expensive and this cat X is. And that's, that's warranted, right? I think that's fair. It's fair to keep an eye on that. But there's a lot of that that I think feeds into some of the concern. And again, like that's not going to go away. So that's just a dynamic, right? That's there. So it is, it is the world that we live in. But like I said, I do think there's still something to really doing a true like fundamentals analysis on these businesses. There are still very good businesses, very healthy in cash flow and margins. Again, right, are those sustainable long-term fine, but that's where the work goes in. So yeah, I'm going to withhold a opinion on the direction of anything. Just do your research. Be careful. Um, stocks can go up and down. And the world's not ending probably. It's not Jay. Probably, probably not ending. No, but to be honest with you, like, I don't know. I love the side that we're in. Like I love this cycle, even with the uncertainty, just because like there's just so much demand for like just sane thought on this, right? And we get so much inbound, both from investors and clients like around like there's just so many questions like what's happening? Explain it. Like do you know what I mean? Or, you know, if everything's crashed, they'd be like, well, why this sucks. We can't really do a lot, but like we're in this like interesting period where I don't know, there's just, there's a lot interesting going on and it makes it intellectually stimulating. And there's lots of ways you can slim this, but the demand for good information and analysis and sane thought like is like people are hungry for that. And I like those cycles. Yeah. That's probably an indicator in itself is that people have not given up on semis. There is still, uh, I'm certainly busy. You're busy. There's still people out there who are asking for insight into the industry because there's going to come a day. I don't know. Maybe I'm doing this too long. I just, I feel like I remember there, there are periods when semis are not as interesting. Right? You know, the entire 2000, the odds when nobody cared about semis, um, our lost decade. Right. And, uh, right. And so there probably a day again when things slow down, but I think semis have become too fundamentally important to be completely aware of the work. And then it's just software again that gets exciting. Well, yeah. All right. All right. All right. Let's talk. Let's jump to the, uh, the meta discussion. So this was interesting. I'll let you give the brief, uh, the brief snippet of what happened and then let's, let's digest it. So there was a Bloomberg report that meta Facebook is going to cut a deal to license. Meta, meta is building an immense amount of data center capacity. And there's, uh, there was a report Bloomberg that they are now going to rent out some of that capacity to Google, which has been very much positioned to respond as being, oh, meta is now becoming a, a neoclub. And this also follows on the wake of a couple of weeks ago, SpaceX cut a similar deal with Google, right? Where x.ai has, of course, big data center capacity in Tennessee. And they have, um, cut a deal for, what is it? It's like a billion dollars a month, two billion dollars a month, less leasing out their excess GPUs to Google. And the key problem in all of this is that word excess, right? Why does, why do these companies that proclaim to be at the cutting edge of, uh, AI? Why are they, why do they have excess capacity? Mm hmm. And that's, and then that, that, that cue, cue long debate about the subject, because that's what's been going on a week. Okay. Sorry. So let's start there. Why just better have excess capacity? I, I want your take and I'm going to give you mine. Um, all right. My, my, my take is that, uh, I actually, I'm not, I'm not 100% convinced that what's going, going on is exactly what people are. Mm hmm. I agree with this. So, so let me, let me say, I think, I think, let me present sort of a strong man argument, not necessarily what I believe, but this is a strong man argument is like the idea that is, um, meta is overbuilt and has, uh, doesn't know, clearly doesn't know what it's doing with AI. It's very confused internally. Their AI teams keep changing their names and there's lots of moving parts inside of, inside of AI structure inside of meta. They've been in the press a lot lately for some of their misdirections. I, I think. And so the, the, the, again, the strong man argument is that they've, they overbuilt, they don't know what they're doing. They're using all of that. So they, and they see, ooh, Elon's making $2 billion a month renting out his GPUs. We can do that too. And, um, that's a bad look. I mean, that was certainly a big factor in the sell-off this week. That was sort of the, the catalyst for a lot of this. Why does meta have so much capacity? That's, that can't be good. So right, because if they have too much capacity, that means they're not really, they're not gonna, that means they're going to spend less. They're going to spend less than maybe everything is an insuredge. And as, like I keep saying this, you know, forever now is there's seven companies behind the entire AI trade and meta is one of them. And if they're going to stop spending, does that mean others are going to stop spending? Yeah. Um, all right. So there's, there's two, there's two other parts of this that I'm intrigued with. And I do think I know we have folks in meta who listen to this. So hopefully we don't entirely botch or get all of this wrong. But I have talked to friendlies there who don't seem to believe that they have a whole massive amount of capacity available because they are all still struggling for compute within their internal teams. So then the question is, okay, and I, I know for a fact it's true because I've heard it from multiple different groups. So then the question is, well, is it just training GPUs that they have an excess due? Because they've stopped training right now, a model or they're done, right? And so they've got infra that's dedicated to training. I don't think those are necessarily GBs. It's probably, you know, age 100s and whatnot. And that they're willing to say, you know, the ML teams, right, the people who were building those models are in some degree of a pause or, or we're slowing that down or whatnot for now, not forever for now is where that's coming from, right? Not the people building the recommendation engines, the inference on top of their services where they're trying to increase our boost. So there's one thing. The other part that I think about is, you know, one, going back to this question of fundamentals, right? Every business that we analyze, like I always bring it back to, you know, what's really their business model? What's their North Star that's making them money? And for Meta, it's advertising. And I do think there's something, a question to be asked about when your growth metric largely is ARPU. What does AI look like in that as an increase of revenue per user tied to how much it costs to build and run that compute, right? So, so.
For example, if Metta's only adding, and I'm making this up, because I haven't done it, but a dollar or two, a month or $10 a year per person, now I get Metta's scale. We're talking billions upon billions of people using these things. What's that trade off where I got to make this much more per user in revenue from AI in order to justify my spends? That's a question. I think that's a fair way to look at this about the way that AI fuels the flywheel for them the same way it does for Azure, for Microsoft, across the board, Google, Amazon, et cetera, all their products, including their consumer-facing ones. There's that. I do think there could be a dynamic at an imbalance. The cost that it takes to build and run is not quite yet monetizing in our poo the same way. Then the other flip side of that is, "Okay, did they then also then just say, "Well, we're not giving up on that. It's certainly going to happen." But we would love to see some financial cash infusion, and we have this infrastructure sitting there, even if it's not a massive amount, let's say it's 10%. Let's go get some dollars and that's going to help our economics. Because I don't think, to be honest, as of yet, there's a pivot in the company to say, "We're going to also now be a neocloud." I don't think that's happening. But I do think they see dollars, like you said, like SpaceX 1. And again, if the ML team is like, "Hey, we just don't have demand for this right now. Let people train on some of this, not inference, but train on bare metal because that's really all meta would do." I could see it, right? It's going to help lift revenue, give them some cash to keep offset some of that costs, while they continue to stay to their North Star, which I still maintain is figure out how to use AI to increase our pool. So that's my brain dump. So if that's true, then I think we still have a little bit of a problem. We'd be in the collective semiconductor industry. Because it says that meta doesn't know how to monetize AI, and it's investing. Everybody knows we're all, all these companies are investing, investing in quotes, investing CapEx ahead of monetization. Now what you're just kind of saying is Facebook is not just investing ahead of monetization, it's investing ahead of the plan of having a plan for monetization, right? It's one thing to say, all right. We're going to throw a gazillion dollars of CapEx because we know that after four or five years, we're going to be able to monetize it in these different ways. You can build a model, you can sit down and say, "All right, yeah, it's going to be really expensive up front, but man, when we get there, it's going to be a huge, great, glorious future." What you're saying now is we don't even have a plan. We're just going to keep throwing money at it. The fact that somebody at meta is thinking about short-term monetization implies that the finance team at meta is interested in the subject in a way that they haven't been in the past. You think about it a year ago, if a company, one of these big companies cut CapEx, their stocks would fall. We're not investing to keep up with the ARIs. Now you're saying, "Oh, you make CapEx because you need to make money." You're worried about making money. That's not a good sign. Again, outside of, I get we don't know the future, I do largely struggle to see how a company whose business model is advertising solely. That's their north star. You can say we're going to charge for hardware someday. That will never be your big business. That is not your core competency. Meta's business is oriented to be a free service at scale and monetize those users. I struggle to see how you drastically monetize A. either. I will not Google into some of this, but I do think Google is in a different position because they can capitalize on commerce a different way. They can capitalize on search queries. Things might translate a little better there because A.I. does have some distinct advantages to search, for example. Most consumers are just using it as an optimized version of search. You don't do any of that in Facebook. It's summarizing things for you. It's summarizing a group comment thread that you read. They're using it for, again, the recommendation for ads, which are increasingly becoming A.I. And I don't think they're any good. It's harder when I just feel like their business model is oriented slightly differently. I think they, in particular, have a little bit more time struggling to monetize that than someone else, like a Google. But again, Google, that's the other you talked about. Meta, what not? They have other revenue stream. I'm not sorry, Meta. Microsoft, they have other businesses driving revenue. Facebook is purely, I got to monetize our eyeballs on our poo. And I just personally, I feel like that's going to be a little bit trickier within their services experience. So I don't, well, all right. So then you have the issue where Google actually appears to be monetizing A.I. very well, at least in the last quarter. And remember, all of a sudden, all the stuff changes quarter to quarter, right? So a year ago, it was the opposite. Facebook was doing, six months ago, Facebook was the one who was doing well. Google wasn't now. Now it's the opposite. Right. I personally think that the problems, whatever is going on at Meta, the real issue is that there is a disconnect between the founder CEO's big vision and how it's being run inside of Meta. It's how it's actually playing outside the company. And I think, right, I don't know if you've listened to Ben Thompson and he had Eric Sweeper on a few weeks ago. Right. Yes. So this, like, I think their analysis is correct. But like, and what they're basically saying is that Mark's vision of A.I. Mark's vision of Meta's role in A.I. is big picture, right? It's not Facebook. It's Meta technologies, right? They think of themselves as a platform. He thinks of himself as a tech visionary and he's trying to invest in the future, right? That's why we had this whole multi-tenb, decabillion spend on reality labs and virtual reality, right? And he's investing, he wants to build the A.I. Future and A.G.I. or ASI or whatever we're calling it, right? But it's actually, that's not what they make their money to your point. That's not what they make their money on. They make their money from advertising. And his, my sense, my read is that his focus is off. He's looking in the wrong direction. He has these big picture, right? And what they really need is somebody to sit down and go, oh, if we just sort of like build this A.I. model and do these tweaks, we can double our advertising revenue. It's very similar to what happened with mobile. They kind of miss mobile until a few people internally were like, hey, let's, we can easily do this thing and make a ton of money on mobile. And there was a lot of internal resistance at first. It wasn't the direction the company was headed. And then somebody at some point realized, wait a second, we can make a ton of money here. And I think that's so, so I think there is a lot of very mixed incentives and mixed directions percolating through the A.I. team. It's a very weird structure. There's a lot going on internally there. And I think it's caused a lot of confusion on the outside about what Meta really is trying to do here. And it opens them up to this critique of like, oh, now they just want to be a neocloud. Right. And again, I don't, I don't think that's actually what's happening at. I don't think it's. I don't think they're trying to be metta. I think what I think it's really, really boring is going on. I think there was a headline this week that also came out that Google cut off Meta and then Meta cut off. Yeah, I think it was Google cut off Meta because Meta engineers were using GCP, we're using Gemini too much. They basically had a deal like all the all the all the meta engineers were heavily relying on external A.I. tools. And they realized that that was not sustainable. And so now they need to get they need to get near term coding assistance, right. There's some really, really good coding assistance out there from Gemini from codex from Anthropic. And I think what's taking place is they had a bad contract in place. Nobody was clear. Everyone got cut off, right. They're now just conducting a commercial agreement so that so that metta can get access to third party coding tools, right, which arguably should have.
to build itself. There's lots of good ones out there. This is not their core competency. They need to focus on building ad stuff. Why not let them use the best tools out there? And it's just a commercial agreement that's being negotiated and one of the ways in which meta can pay for a license to use Gemini or maybe Claude is they're going to pay in like, hey, we'll give you capacity, right? It's not we're going to be a meta cloud. It's, hey, you know, we're we're conducting a contract with one of our vendors, right? And it's, you know, instead of paying the vendor cash for napkins and other paper products, we're going to pay them in compute capacity and gigawatts for coding tools. I like it. I like that. But back to your other point, I do agree with you on management/founders, perhaps misread of the big picture because it's not like this has not happened before, right? When you reorient your company around a mythological metaverse that's never going to happen for anybody, I would say that's a misread of the market as a whole. But, but, but, but you're also right, right? And in fact, when I was on Benton St. podcast, like we talked about this too, was like part of part of where I think people and meta and two, like he's a little delusional in, yeah, sure, an agent's going to do all your shopping for you. And like, Ben's point was people like shopping, like there's a fundamental human aspect that you enjoy. It's going to help make it easier. You're going to maybe discover new products, get time to look at them, but you still like shopping, like making those decisions. That was his point. And I completely agree. Like there's a human in the loop that gets, you know, this and Mark sort of something, oh yeah, you're going to have your virtual friends. No, no, you're not. I'm sorry. Like humans, like he has very little understanding of human behavior. And to sometimes, that's a fault, honestly. And somebody needs to be like, bro, you don't read human, you don't know anything about humans. Like you need, you need someone who's a behavioral psychologist on your team to just help you understand humans. But there's an element of that that does worry me. But like I said, if you just do that we're all a fan of, right, just understand a company's business model and then work backwards into where they'll have successes or failures or what will work for them. Like it dictates everything how they make money. And, and that I think comes back to, like I said, they're the only one who sold these, sorry, maybe not the only one. You can tell me if there's others. But in my brain, like they are the front facing company who will solely exist on trying to monetize AI with free services. Like that is their business model. Outside of those in China. Yeah, China's the whole different things. I'm scooch on it. But, but you know what I mean? I, I, I think, I think Google ultimately, yeah, I think Google ultimately is, is much closer to that. They are, but not entirely. They do have other revenues. They have, yeah, but the overwhelming majority of their profit comes from Apple search. It's Google search. Correct. It's Google search in YouTube. Right. That's, that's where they make the overwhelming amount of their profit. GCP, we can argue that GCP is they're, they're making a run for it now in a way they couldn't have before. They're doing well. But like, yeah, Google, I don't want to totally discount that, but I still think five years from now, 70 years from now, Google's, Google's ad business is going to still be the majority of profits. And then put some a lot closer to, to Facebook. And, and so my take is like, Google's doing something right, probably with agentic, with AI search. I think, I think, Meta will get there too. I mean, to me, this is the ultimate irony is I have, I think that Meta is one of the best position companies to have the clearest use case for AI. Right. We've talked about, yeah. Right. Like, figure out a way to do generative AI for ads that are more engaging than what any human could build. I think that's very, very possible. It really blows open their, their model in a good way. They just have to, they just have to realize that's what they want to do. And they're, I mean, I don't know, I don't, I don't cover them that closely, but my guess is they'll figure it out. It's just, it's, it's, it's been very frustrating, man. A number of fronts, the Meta's approach to AI has been just like, why are you doing this, why are you doing this hard thing? There's a really easy thing you can do. Well, okay. All right. Let's, let's, let's move off, off Meta, but we all feel that a tons of inbound this week for it. So, but yeah, that was a good, good brand. We'll see. I can't, I don't think they are completely out of the woods, out of the woods here, but I agree with you, there's a, there is more upside, but, but regardless, they are not trying to be a neocloud. They may be doing some of the things like you said on the back end and or just having a short-term path to revenue, but I don't think this changes the strategy yet that, that we hope they pull off. So, more on that. Good saga. I like a good saga. Some drama around a, around a company. All right. Let's talk about foundries. So, you start with some news that I guess is sort of undercover around in Phineon. And then let's talk about Samsung and then broadly some, some foundries stuff. Yeah, I mean, I think it's, there's, all right, in Phineon opened a new, a new fab this week. And I think Dresden in Germany, a silicon Saxony. Anyway, in Phineon opened a big new fab. It was big news because it's a big fab, it's 300 millimeter fab. It's, they spent nine, ten billion dollars on it, nine billion euros on it. They got some funding from the European government. It's the EU Chips Act. Remember the Chips Act from ancient history. So the EU Chips Act, they built a brand new fab. It's pretty big. And I just, like, I like new fabs coming online. That's, it's good news, right? Now, I can't put a life in me figured out exactly what they're going to build in that fab. They've said it's for power semiconductors. But I think it's, I think it's just sort of broad-based. Everything fab for them, it will do power, it will do analog. It's probably a fair amount of silicon carbide. And there's well, right? But it's very funny to read the press release. And all, like, not just the press release, they have a whole website on it. And nowhere does it actually mention anything technical about what they're building. Like, we know that the fab is lead certified and that it is hugely beneficial to the environment somehow. And it's built with European money. And the German government is wholeheartedly endorsing it and how many jobs it's built. But I defy you to go on their website and figure out what this thing is actually going to build and what materials are there. The only technical detail I was able to find out was that it's 300, 300 millimeter wafers, which is. Everybody does. Which is. Well, there's a whole other debate in the analog world about the 200 to 300 millimeter conversion. And. Well, true. Right. Infinity on is, you know, is on the moving forward side of that. But it makes sense, it makes sense because like all the power semis, you know, work I did, like, Infinity and keeps coming up as a beneficiary of that power content. And that, again, as we talked about, is hard to manufacture. So I could see them, I don't know, who knows what the scale is, but I could see them needing to ramp up that and something like this, you know, could be logical for a lot of the power semi-stuff. Because that is where they're seeing incredible demand. And it's very hard to scale. So, low is taking fruit. All right. So I'm intrigued about Samsung lately. And for a couple of reasons. One, I am building, mostly have built, but pretty far along on a foundry model, which has to include Samsung, obviously, and Intel. And when you just look at leading Edge, right, you continue to see them scale up leading Edge at two. They've got stuff in advanced packaging now with IQ and a handful of other things. But where Samsung gets interesting is we know, you know, I think we've talked about this, they're leveraging memory for logic, right? Come to us for logic and we'll give you memory. Or we're working on near-package memory, like we talked about with Qualcomm, right? They are working on stacking memory on logic. Okay. That to me is an interesting way, like, lever that they can pull because you need memory. And sure, right? And if it's good enough, and well, it sounds like their verticalization there between memory and logic is one of their biggest differentiators. There's enough smoke out there that they're winning some of these deals that you're, they're going to be a beneficiary of this cycle too. And not just like, because I haven't, I started like not saying Intel is just a second source. I know that they are, but I think they're
there's some things, EMYB is a good example. There are events that are leading people to choose them for that over someone else, even though like it's a first choice for something, maybe not everything, including Wages, but it's something is causing a first choice for one of those roadwaps. And I think Samsung's in the same boat, right? You'll still use TSMC for some things, but this Samsung will become your first choice for a different product, a different skew, on the basis of some of those levers that they can use, like memory, for example. But they're getting a lot more attention now. And I think that there's enough there that, when you start looking at leading edge, so three, two nanometer and below, they're gonna be right there. TSMC will still have the lead, but they will be there with some scale, along with Intel, those three. So it's just interesting, Samsung is officially back in the conversation. Compared to two years ago, where we were like, they're below the 50% line, where this can actually be economically viable. - Yeah, I mean, I've been very vocally skeptical about Samsung's ability to stay in the market, and I was wrong, they have turned it around. And I think they have always had the technical expertise inside, and what they were lacking was organizational, some issues, and they seem to have sorted those out. There's a new leadership team in place that seems to have gotten everything back on track. And then they are now in a place where the markets here, the market's ready for them. And I think, I don't think they ever fell out of a conversation entirely, but people were starting to question them. They're certainly struggling, and now they seem much better engaged. - I think you're right, they have memory, and it's easy to get foundry business when you have memory, which is the scarest thing in the world right now. I'm sure that's an incentive. But you know, you look at what Intel's done, Intel doesn't have memory. They're winning on technical merits of other things, like they're packaging. - I agree with you. - Right, but I think, my impression is that people are, again, there are lots of reasons to look at Intel, if you're an American company, you can support America, which by the way, we're recording this on July 3rd. - So, hope everyone had a good holiday. But the other, I mean, but anyway, my point is, packaging seems to have been a big part of the conversation up until now. - Yeah. - But I think if you talked to most of the people who have been evaluating Intel, my sense is that they are increasingly interested in logic process as well, right? Which is more than just a, that's more than just like, oh, I, you know, it's not, they're not just going for the packages, they're actually looking at the correct process. - But that's, that was my point like, yes, it's becoming a first source for X, Y and Z, but packaging got them in the door, the same way that memory gets them in the door for Samsung. And that's fine. Like I think there's, again, we are at a point in time where TSMC absolutely cannot fulfill capacity for demand. So we need other sources. So it's great that Samsung and Intel are there. But, right, something got them in the door on technical merit or in need, it's memory for Samsung. It was, it was e-mib, it was events packaging for Intel. But what's interesting is that what's sort of come up when kind of I've just been like digging through Samsung in particular, because I knew this about Intel, you know, the challenge Samsung had was, you know, Xinos was the thing that was supposed to help be the quote unquote pipe cleaner, right? It's the, you know, your first best customer for Intel was Intel, Intel chips. And that helped flesh out 18 A when they made that ship, that ship, and Intel does that at volume, right? Xinos volume is tiny. Like they don't put that. So like they struggled for any true scale customer. And what I've liked about like everybody talking about kind of what's going to be the fuel for both these foundries is just one external customer, like it's been talked about with Apple with Intel. Somebody comes in and really helps clean that pipe, like at a whole different level of scale. And, and the same thing I think has got to be true for Samsung. It's just who's that customer? You've heard AMD, you've heard a little bit about Qualcomm. There's been rumors this week that Google might be doing a variant of TPU, which you know, someday they could have five or six TPUs. So they're going to spread this around all foundries, but that they might be doing something again, a little bit more memory intense, so probably an infrancésic or something with, which sounds so somebody comes in and basically changes the game for them. Like one customer, just legitimizes the whole thing. That's going to happen, right? We just don't know who it is for Samsung. But I really do think, like, as long as this memory thing is an issue, which is for as far as the, I can see at this point, that they're going to be able to use that in some very unique ways. They're tied to memory and logic in their foundry that I think will be extremely differentiated over time. It's early, but I think in a few years, they'll start to see some of those wins. And we will actually see some really intriguing memory to logic packaging that perhaps Samsung does better than anyone. Not something I ever thought I would have said two years ago. I, um, okay. I mean, it makes sense. I, you know, I actually have this note. I've been writing and I don't, I don't, I don't feel very good about it. So I'm not sure if it's ever going to see the light of day. But the idea is looking at the competitor landscape, when the cycle ends, what, who's going to be left standing, what's going to be still here, right? Because right now it's, we're very much in a situation in which, what's the saying in a hurricane, even turkeys can fly, right? Yeah. And so I'm, and I'm not just looking at foundry. I'm looking at lots of other spaces too. There's, there's everybody in semiconductor land now is doing well because there's so much demand. How does that all shake out when the demand isn't quite so strong? Right? And okay, maybe we never come to that point. We just have infinite demand forever into the future. But probably, probably it's going to be cyclical again. And when does that cycle, what, what happens when that cycle takes place? And I think in, in foundry, I think the world can support three leading edge foundries. There's a nice idea. But it'll be, I guarantee at some point we're going to be at a point where there are going to be some scary headlines about somebody. So I think. No, I agree. I guess my point would be like Intel and Samsung have far less fabs in a digestion period needing to stay full than TSMC. Like this was my concern from day one. Like TSMC goes and builds five more fabs. And I think that is a minimum at what they would need to meet capacity. And again, come back to my point, right? We've been talking about what changes the semiconductor industry, you know, this idea of is it semiconductor industry 2.0 or is it just the angstrom era? You need all, all new tools, all new tools for two nanometer below and advanced packaging. So like you're going to fill all these files with brand new, you're going to go build five new files and you're going to fill them all with brand new, you know, wait for equipment. Okay. But TSMC has what, I mean, what three X, the number of green room space, then their closest competitor, like that's my concern with them as this build up. You have to do it because there's so much demand. But they're going to be the ones with the largest footprint should this, sorry, you're laughing because you're like, oh God, Ben's just basically killing killings, but that was my concern. Because their whole model had always been minimum 75% capacity. Okay. If everything plays out like you think, I don't know how you keep 75% capacity. But filled unless Samsung and Intel die and it's all TSMC. There you go folks. Ben is the TSMC bear now. Or am I the bear of the other two founders because it all comes back to his regardless. Let's not scratch it up. But it's a good point because it's, I think it helps explain why TSMC management is so reluctant to add as much capacity as everyone wants them to. They're not dummies. They're not dummies. They see this. They know. They get so much grief, especially from this area, from Silicon Valley, they get so much grief for how conservative they are. And you kind of see their point. Well, and here's the other thing. If this was just monolithic or SOC based, it's a little easier because you just get so much per wafer. You could expand a little bit and drastically
expand your output. Here you have to expand significantly with these giant chips, right? The size of these chips just makes this scale that much different. So if the demand for giant packages drops, that hits your green, your green room space far more than when these were just monolithic. And that, like, again, I just, you're right, everybody, it's not like I come not, they all know this. But that's, that's been my like concern when, way for output is so low and you're going to need three to four foundries, three to four fabs just to make as many as one did in monolithic per output. You know, it's like, "Ch." Anyway, I don't want to freak everybody out, but this is our 10 year, our 10 year. So, but, but, but I keep, so I just want to make this point because I did a piece on constraint leverage to semis. And, like, part of, part of my like philosophy here was, you know, sure, this is all fantastic. We're all pointing out that semi connect industry is growing it on present and it rates. Like, we're going to hit probably two trillion in revenue in 2030. But where this gets interesting to me is this was the sort of the stat that I came up with. So if we hold, if that forecast holds, then from 2025 to 2030, semi connect our industry revenue grows roughly 22 to 24%. But shipments will only grow 5% per year. So basically, revenue is on pace to grow 5, 4 to 5 times faster than units. And this is why ASP is becoming interesting. And this is why everybody below, if you see this chart, I put on this report in the diligence stack, there's a there's a crossover point where for the vast number of history of time, for like the entirety of semis up until 2026, basically, units were always more than ASP. Now you're at a point where that's crossed. And so everything below my kind of shipment line to revenue line where the revenue line is vastly higher and and escaling up than shipments, everything below that shipment line is pricing leverage PCBs, MLCCs, you name it, right? Components, organic substrates, waifers, everything can increase margin because the revenue is going up so much and they are at a physical constraint in units. And that dynamic stays. It's not like they all of a sudden wake up one day and was like, "Oh, we're like we talk about that a lot. I can all of a sudden manufacture more than those. The reason and finney on builds a fab like everybody under my line of supply chain is in an era of margin expansion. How deep they want to make that cut? That's up to them. Do they want to grab gouges or customers up to them? But every single one of them can increase prices and is increasing prices. In fact, I have a map in there that shows you by category what the average price increases by segment of semis. And that's all it's all going up. It's not flat. Like going down, it's all going up. So like this is the leverage state that they're in. Then bringing it all back to day one, the first conversation, capex goes up and output doesn't necessarily go up in tokens and or revenue when all the supply chain and semis goes up. And I get that concern. Like we talked about with microns. But regardless, the supply chain in semis has unprecedented leverage in costs. And I'm just fascinated to see who. What do you want to say? Price gouges is customer friendly. It's not customer friendly when you just are in unprecedented demand in for for for everything you make in the supply chain. I think corporate speak for that is who is best able to maximize their pricing opportunity. There you go. CFO. Thank you. But this is a real problem. Like, and that's the thing. Like that's why I'm interested in this is because we get you talk about it with memory. And that's one thing. And yes, memory is very expensive. But if memory move, if if bomb for set for for compute. And this will be true of custom too. Like the yes, Nvidia's prices will go up. But the costs to do custom for those is going up. Sure, it won't be as much. But it's going up. It's going up everybody. Okay, so you're again, your cop X has to increase. But but that is including an increased share of compute bomb. And I just you know, I mean, like that's we just have to all be prescient about this is is is my point. So yeah, there you go. All right. I don't think you meant to be, but it comes across as a little bit. I know it does. It does though, but because again, let's okay. So let's just let's just say this didn't happen, right? Because again, and this is why I think this point with micron as then or all the memory guys as a lens. In fact, I saw a report. So microns that what was it 83% margin? I think last quarter. Am I wrong? Is it? Low it low eight is right. Okay. But like SK and Samsung were not quite that high. So then people were like, Oh, they're going to go to 90%. Like I saw Korean analyst basically saying like Samsung and SK will go up to okay. Well, again, if everybody does that, right? That increases your costs. If memory becomes 50 to 60% of high-prescaler CapEx costs all while everybody's freaking out that they're not going to make their return on investment capital. But their costs go up. Like I just my point is like that is a real thing. And I do think that feeds some of the sensitivity to the upside here, right? Because again, your costs are going up and your token output isn't necessarily or your ability to maximize that either is going to take longer. But that's not the same sort of parallel. And I just, you know, again, I'm just being as clear as possible. That's a dynamic. Yeah. I mean, that's the that's the bear case is next year 60% of that trillion dollar CapEx 50 60% of that trillion dollar CapEx is going to go to memory. And it will, right? Yeah. But like, but if you're if you're the CFO of Google and you're like, yes, how much am I spending on memory? Right. Right. Like like the people are going to start asking these questions. And if you, you know, people are already asking these questions. Yes, they are. For the moment, everybody's a little bit of like trust us. But I think I think one of the one of the things that's also happening here is we're turning to the debt markets to finance this. And like I think you can convince an equity shareholder that hey, there's a long term story, AI, AGI agents, whatever, you can convince an equity analyst of that, right? I don't think you can convince a bond investor. Like that's just not that's not what they're they're like, well, yeah, but you're going to owe me the coupon every quarter. How are you going to pay that? Right. That's what they're going to look at. And right. And and the other thing too is if you're talking about like, oh, we're going to keep investing, then we start to have these conversations about like, well, what kind of returns do you need on that investment? Like how much profit do you need to get a return on that investment? And the numbers are getting so big that we're starting to have having to measure the return on your AI investment as a share of GDP. When we get to that point, the argument becomes very hard to me. I mean, it becomes a very 1999 argument, right? New new world order, new economy. Yeah, right. We're not there yet, but that's that's we're kind of getting. But but but but I think what we, you know, started with and are end capping with is just they're again, being as clear as possible, you can understand the fragility and that fragility just dictates a lot of what happens in the market because the things we're talking about I think are things other people talk about and are sensitive to. And so that's 100% the dynamic. And so, you know, but it's what is it they say if you if you if you think you're crazy, you're not crazy, because a crazy person wouldn't know they're not crazy. And so the fact that we're just talking about this and the lots of people having conversations is what keeps me from feeling like, oh, it's 1999 all over again, but we're adjacent to that. Yeah, but I think again, it's just good to always be able to see both sides of this, right? The bull case, the bear case, where should we be concerned? What are our signs of optimistic to be optimistic, etc, etc. So yeah, yeah. Again, short term that semi supply chain margin expansion, pick your winners accordingly. Yeah. And as bearish as I am and the cautious and skeptical as I sound, when we get done with this recording in a minute, I'm going to go spend the rest of my day managing two dozen agents. So 100% 100%. All right, everybody. Thanks for listening. We'll talk to you next week. Thank you, everybody. Hope you had a good holiday and tell your friends, click like, subscribe. [BLANK_AUDIO]
Podcast Summary
Key Points:
The semiconductor market experienced a significant sell-off, attributed to fatigue with the AI trade and an influx of "tourists" (generalist and retail investors) driving valuations above fundamentals.
Market sentiment is shifting due to overvaluation concerns, with no clear catalysts (e.g., war or interest rate news) causing the decline; it's seen as a correction from excessive momentum.
Social media, especially Twitter and Substack, increasingly influences stock prices by shaping conversations, though it doesn't control them, and professional investors follow these platforms closely.
Index-based investing and passive funds force managers to buy stocks to match benchmarks, contributing to overvaluation as they add positions even when uncomfortable with valuations.
Meta’s reported deal to lease excess GPU capacity to Google sparked debate about overbuilding and monetization challenges, raising concerns about AI spending sustainability.
Meta may have temporary excess training capacity due to model pauses, not inference, and could lease it for short-term revenue while still prioritizing AI to boost advertising ARPU.
The lack of a clear monetization plan for AI investments at Meta is worrying, as it suggests spending ahead of a strategy, potentially signaling broader industry issues.
Summary:
The discussion centers on recent semiconductor market volatility, driven by investor fatigue with AI stocks and an influx of inexperienced "tourists" that pushed valuations beyond fundamentals. The sell-off lacks clear catalysts, pointing to a sentiment correction rather than fundamental news. Social media platforms like Twitter and Substack are increasingly influential, shaping stock narratives and affecting prices, though they don't control them.
Index-based investing compounds the issue, as fund managers must buy stocks to match benchmarks, even at overvalued levels. The conversation then shifts to Meta, following a Bloomberg report that it plans to lease excess GPU capacity to Google. This raises questions about overbuilding and monetization: Meta may have temporary spare training capacity due to model pauses, not inference, and could lease it for short-term revenue while still focusing on using AI to enhance advertising ARPU.
However, the lack of a clear monetization plan for massive AI CapEx is troubling, suggesting spending without a strategy. This dynamic could signal broader industry challenges if other hyperscalers follow suit, potentially impacting the entire AI trade. The hosts emphasize the need for fundamental analysis and caution amid uncertainty.
FAQs
The sell-off was driven by fatigue with the AI trade, concerns about too many 'tourists' (generalist and retail investors) piling into semis, and a belief that valuations had outstripped fundamentals.
Social media affects the conversation and sentiment around semis, with influential substack writers and Twitter accounts driving discussion. While it doesn't control share prices, it can lead to quick stock reactions and impacts retail and professional investors.
Meta may have excess capacity due to a pause in training new AI models, allowing it to rent out GPUs for cash. This could help offset costs while Meta focuses on monetizing AI through advertising, though it raises concerns about its AI spending plans.
The argument is that Meta overbuilt capacity without a clear monetization plan, leading to excess GPUs. This suggests uncertainty in its AI direction and could signal reduced future spending, impacting the broader AI trade.
Meta's primary revenue comes from advertising, so AI investments aim to increase ARPU (average revenue per user). The challenge is balancing high CapEx costs with monetization, as AI must boost ad revenue enough to justify spending.
Index-benchmarked investors are forced to buy stocks to stay close to the index, even if valuations are high. This mechanical buying, especially via ETFs, can inflate prices and contribute to market distortions.
Chat with AI
Loading...
Pro features
Go deeper with this episode
Unlock creator-grade tools that turn any transcript into show notes and subtitle files.