Ep 184: Earnings TSMC/ASML/AEHR, State of AI Semis

[MUSIC] >> Hello, everyone. Welcome to another episode of The Circuit. I am Ben Baharin. >> Greetings, agents. I'm Jay Goldberg. >> All right, so we had a good week of earnings. A lot of interesting stuff happened this week, more so than we're going to get to in our allotted time. But we do like the names of companies that we track this week. So things like TSM, CES, ML, and Air, who largely are just again giving us a lens and a view of what's happening in manufacturing and foundry and wafer equipment materials. Lots of good stuff to unpack on that. So we're going to start with TSM, C, who we've typically constantly said was, add-bell whether indicator of the space. I'll let you start the numbers and then on some packs of the commentary and the demand signals that they're seeing. >> Sure. So TSM, C, reported Q2 revenue head of expectations, EPS ahead of expectations. And they got revenue up for next quarter as well, not by a ton, but by, well, certainly above cell-side consensus, probably not so high above by-side expectations, but above. And then margins sounded very good. So strong numbers across the board. And probably most significant, they also raised their capex forecast. >> Yeah. >> Yeah. I'm glad to show everybody else we're going to talk about. >> And I think it will, I mean, I don't know what you're going to read was, most people assumed they were going to raise. I don't know if they knew how much they were going to raise. I was seeing people assume in the '60s, which is what they raised it to, but I don't know if everybody was like, for sure, we're tagging north of 60 billion as that number. But this is interesting to me because, so the other thing they said was, 70 to 80% that's going to be allocated to advanced process. Obviously part of that packaging, they have to continue to increase co-waste way first because advanced packaging is going to come hand in hand. But the observation I think is interesting, and this is where, like people look at this and they go, okay, right, when you are a company that has to increase CAPEX this way, in order to build more, in this clean room space for your investments. What makes it interesting to me is that this is a company who used to be, that are growth, their entire upside was on smartphones. And smartphones, to be honest with you, because they are little tiny chips, pack a lot of chips per wafer, just didn't necessitate. Like what you could envision is four to five additional foundries needed. Four TSMC, two meat demand because the size of these chips, plus the wafer you need for packaging. Like I say all that to say, it's just interesting, the implications on clean room space and equipment. When you move from monolithic tiny chips in a smartphone era to what we are now in the HPC era, is a drastically different understanding and shape of foundry. That's the point I wanted to make. I think of it more abstractly in the sense that Moore's Law is running out. I won't say dead. Let's see Moore's Law is running out. You could just say it. Moore's Law is running out. And so we're having, and so everything is becoming, everything is having to become more complex. We're having to do more to eGout gains in compute. And I think that's going to be true for smartphone chips as well. If we want to keep improving those, we're going to have to do similar things. I mean, we haven't quite seen that take place yet in smartphones. But we certainly are seeing in PCs, right? We're doing both Intel and AMD are doing pretty, are doing a lot more to get PC chips CPUs up to running. And I think we'll see that in smartphones too, probably in a generation or two, because we're now at the point where leading edge is probably too expensive for most chips, most phones except for maybe Apple. So I think we're seeing it first in AI and high performance, compute HPC, but it'll come for everybody else too. Cost per bit is going up. We're getting more performance, but cost per bit is going up. That's really the end of Moore's law. Yeah, yes. And again, I say, so I throw in again the advanced packaging part just again to show how much more equipment and space it needs when it's not just small model of thick dies. That's my point. I think a reframing of what goes on in foundry. And yes, advanced process nodes, all of these things, plus I don't know if I had said this before. But when I toured Intel's Fab 52, we walked through parts of the foundry that was doing 10 and 3 nanometer. The ASML machines to do 18A for Intel is at least 2 1/2 times the size of prior nodes. So it's almost like you could argue that to a degree, the machines needed, in this case, just ASML machines, to do advanced nodes is 2 times larger. So you just don't-- you don't have-- you can't fit it in the same space. So I'm just-- this necessitates a whole lot more foundry space is my point. This shift from-- and that just has square footage implications, like all sorts of implications that we just have to recognize. You got to build three more buildings to make less chips than you did in the smartphone. Fine, but these are the dynamics that necessitate that. Yeah. Yeah. I guess I'm quibbling around smartphones. We'll get all those for smartphones eventually. It's just a question of-- I think it's just a functional where we are at the physics of fabrication. And we're needing to use bigger, more complicated tools than more of them and use them more often. And that necessitates more fabs. I mean, I'm not sure, though. Like, I don't know the full reason that you're going to run out of monolithic in smartphones, for example. Like, I've wrestled with this. Like, I don't know when they would need to go to chipplets because they just might not be on the most advanced process. But I don't see-- even if we get to 1 nanometer and 0.5, you could still make monolithics on that. I don't think little chips need to go to chipplets yet in a phone. PC is yes. I think that happens. Fones, I'm not so sure. But regardless, whether you do or not, the bottom line is the space needed to support the AI buildout is just significantly larger. You need more space to do the same amount or less than you did in the monolithic era. And that's-- so when we say-- like we'll talk about this-- and when we say clearly that wafer equipment is going up and to the right, and what I think most people are now as believing as a fairly strong conviction that that could be a $300 billion market by 2030, it's because of these things. You're building brand new green room space. You need to fill that with larger equipment to do more advanced processes. You've got less space as a result to only make tens of millions of chips, not hundreds of millions of chips, like in the smartphone era. So anyway, TSMC-- and that's again-- I mean, we have talked about this forever. In fact, we've talked about TSMC before AIHPC was even really kind of a thing for them to say what they're saying. Like we talked about a few weeks. They have to believe they're going to fill all of that capex and new green field space with chip demand. That's right. And remember, they have customers who are asking them who are upset that they're not building even more. Yeah. Do you think-- here's the question. So because part of this comes back to, they could increase their margins more. Do you feel like they are maintaining discipline for the sake of not wanting to price gouge their customers, even though wafers will go up, it's a little bit. But compared to others, which we won't necessarily name names in the cartel, who are clearly pushing margins to the limit, and we all agree that changes. TSMC could 100% do this. And they're not. They are increasing, but they're not going to like 90% for example. Oh, yeah, but they are raising their prices. They're raising their prices. But my point is they're not-- I don't feel like they're gouging. This still seems within reason. I will admit that if you look at their numbers, you can back into the fact that they raised prices Call it 30% roughly, depending. And that surprised me because I don't get that sense. Like you said, I don't hear lots of people complaining about it. I feel like there's a lot of bad blood now towards Nvidia, towards micron on their prices. And I don't get that vibe from how people talk about TSMC. But they are raising prices. So, okay, but all right. So let's talk about that sentiment though. I mean, it's not like they are, oh well, sorry. I would argue that they are more strategically central than any one memory provider. Two, I think they have, again, historically been very customer friendly on their pricing. So I don't think there's bad blood historically. It's kind of what I'm leading to is like, why is that sentiment still very favorable to TSMC and they're obviously very customer-centric? We'll do what you need. What do you need to do? Like all those things is the service, maybe help soften the, or Nvidia and others are like, you should have been charging this much from beginning. You're so valuable. Like, why have you been undercharging us? I don't know. It's unlikely that point, but still. - I mean, my intuition is that they tend to be just more very, very customer-centric and very concerned about that. But I do know people who say, oh no, that's not it. I know people say, oh, it's much worse than you think. Just no one, everyone is afraid to say anything. - All right. - So I don't, again, I think that they are raising prices to your point not as much as they probably could. And they're just doing that because they have this very long-term conservative approach to things. - Yeah. Yeah, and similarly, like I threw out, what I'm hearing a lot of people project for way for equipment at, you know, in north of 330. And we're trying to figure these numbers out too. Like similarly, the Foundry business, two years ago, the best forecast you could find on Foundry was about 350 billion around 2033, 2034. That's likely to happen, are also around 2030. That Foundry has a hole, which will include Samsung and Intel and others, obviously Tower and Gloufau, doing stuff in Optical. Goes up, right? Faster than anybody thought. But again, a lot of this is, again, margin expansion. ASP is going up for all of these things, for Foundry prices, plus demand, 100%. But there's a lot of price leverage that goes into these forecasts, getting bigger and happening sooner than most people thought. - Yep, TSMC, good quarter, good numbers, raising everything, stocks down. - Stocks, stocks down. Oh man, so this is, I mean, this is our, this is literally, except for the other name, we're gonna talk about in a little bit. This is pretty much like the theme, right? - This is incredible. - Incredible quarter. You absolutely crushed it. Nobody cares. Stock market doesn't care. - Yeah. - The stock market is the honey badger at this point in time, Jay. That nice honey badger. Honey badger don't care. Meam, stock market is the honey badger. (laughs) - Oh dang, all right, ASMR. Give the top line and then let's talk about ASMR. - All right, so ASMR reported just shy of nine billion, sorry, just over nine billion euro in revenue, consensus was below nine billion. They reported very strong EPS for the quarter and guided revenue about a billion, no, billion in change above expectation. Again, margins looking good. Very strong indications that they're raising prices pretty significantly as well. And I think most important for this quarter was they're expanding capacity. They're gonna expand capacity this year and next year. - Yep. - The big, it's interesting that they said on the call tilde 85 low-any systems for 27 and a lot of investors were like, is that like the highest? Because I think people were assuming closer to 90, could be doable. So they feel like they're sandbagging, I guess. I assume there's obviously a lot into the supply chain that requires them to acquire in order to make these machines. I guess not like this is not a marvel of science fiction to make a semiconductor wafer chip. But I didn't think it was interesting that people were like, is that the most you can make or can you make more? A lot of time spent on this pulling out of how many machines do you think you can actually make next year? - Yeah. To really answer that question, we would need to really dig into their supply chain. And like, because it's not just what they can produce is like how many more units can they get out of Carl's ice and all their other fabulous supply chain. And I'm just not gonna do that. - No, I completely agree. Yeah. But yeah, they're probably can do more than they say. Just why would they tell us? - They're not. But regardless, you are a 100% right. Everybody assumes margins, mix is favorable that they will increase prices, their ASP will go up. And so will their margin expansion. As again, we have said, we are in the era of everybody in the customer, in the semiconductor supply chain, it is their era of margin expansion. Like you will not find a name who cannot expand margins if they're valuable in the semiconductor supply chain. I did find it interesting. I'm curious what you think too, that an interesting tidbit was dropped on this call about Intel using high NA EUV for some, and I'm gonna say some, I'm gonna emphasize some, because there's zero chance that there are a lot of 18 A tiles made on high NA UV. There are some, sure, maybe in high end compute, but this is a proof point that they're actually making Intel, they're using high NA UV. And then I'm optimistic that perhaps they could use that for 14 A at some scale, which has not been announced yet as an actual thing, but that would be positive for Intel, honestly, and obviously for the maturedized, maturedization? - Miniaturization? - No, not miniature, matured. - Or maturation. - Or maturation of high NA UV. - Yeah, so high NA UV is the next evolution of EUV systems. They're much more expensive than low at AEUV systems. Everyone else is using, which are already ridiculously expensive. - I'm much, much more expensive. Emphasis much. - Yeah, so I know that it was definitely held up a lot. Like, oh, this is big news, this is big milestone. People are going on to high NA UV now. This is really exciting for ASMR. I'm very confused about what Intel is doing. I strongly suspect that they're dabbling here. They're not like, to your point, - Yeah, that would agree. - Right, they're not, this is not something, I don't even know these will even sell these. I think they're trying to learn how to use high NA UV. I don't think it makes sense to do it for AT&A. I think it, you could argue that it's something they need for 14A, but again, you can argue that. There's a great, right? - I think it's necessary, but agree. Yes, exactly. It's not clear that they need it, but I think they're using what they have now to learn how to use, I don't know how many systems they have, they have more than they need. So I suspect they're not using them all, but they're running it and learning how to use it. And so they'll probably have it for 14A, but maybe not, it's complicated stuff. But whatever, it's a technical marvel. Commerciality, I don't know, but technically it's pretty impressive. - Yeah, so it's interesting though, I like this line of thinking and that I'm gonna totally steal somebody's tweet who I did, there are private accounts, so I couldn't retweet it. But it's interesting that like, in the same way that you think about a early technology design partner of a foundry, right? So like Apple was for TSMC to clean the pipe, help build scale in order to mature an advanced node. Like this same kind of thing exists here with the UV or high in AUV, right? So EV has been around for a long time. It was mature and now Intel is kind of the earliest adopter of EVs. and this person's tweet was, the irony of this is that they were not the first adopter of EUV. But they're going to be the first of helping ASMR mature the next evolution of EUV. That's right. That's right. They were last to the game for EUV and they're now the first to the game with high and EUV. It was the unbelievable strategic blunder run run again by the finance guys I didn't tell. I just you know hindsight's 2020 but like what would have happened if they would have actually just adopted EUV instead of their own stupid thing that did not work. So five years late was it where they were longer was it longer? Yeah. What did that answer that was? Yeah. So let's just I just want to point out that at the time there were plenty of people who didn't need hindsight to say that was a stupid idea. But let's look for the future. That's what they're doing and that's what they're doing with high and a EUV. So yeah, I think that's good signals from ASMR. I mean again, right? Like I said, the entirety of wafer equipment is going bananas. In fact, another person who I did retweet and I'm also going to again steal this basically said like, hey, for memory prices to go down, don't you need a whole lot more wafer equipment material asking for a friend? It's like yes, you are 100% true. Right. In fact, for any of this to go down, you're going to need a lot more like wafer equipment material. So yes, wafer WFE. If you're applied materials, if you're K-Lac, if you're a lamb up until the right guys. There's talks are both down for the week to all down. Nobody cares. Nobody cares. Y'all go make a bunch of money. Your industry is growing and exponentials, but nothing to see here. All right, let's move to air. Who Jay and for me is how you how you say it? I would not have said that this morning if you woke up and said, then how do you pronounce a company, a HR air? I would not have said that. However, that is the name. Really good quarter and the stock was actually up, even though it's not today. It was up, but give us the air down. So air in line revenue, modest speed, depending on how you want to call it, modest revenue beat, but very strong earnings beat, 11 cents versus a consensus of one share a lot of one cent loss. But then they got it, they got a revenue of the street was expecting 85 million. They got it revenue next quarter to 140 million dollars. So that's that's a very strong number. I yeah, so air is you ever go on like Wikipedia and you just like. What's waiting for you? You go on Wikipedia and you're like you're like you're reading something and like click on a link to go to the next thing and then you next thing you know your five hours and you're learning about, you know, Italian battleships and World War two. I just wanted to know anyway. So so I have last night I found that like a like Claude is even worse for that because I spent yes a lot of time last night. I am now very deep in the weeds on the technical merits of air solution for those not familiar with air A E HR they are a test equipment they do burn in test, which is a form of testing of of of wafers and or package chips that basically stress tests them across temperature and usage and voltage. And testing that chips are come out of the fab they're going to work and the world needs more burn in test because of AI because of reasons like it's AI right and air is demonstrating that had a very very strong quarters of wealth because they're. This is this is something we were talking about a minute ago is like as we're moving away from monolithic single die chips to something more complicated. This kind of test becomes much more important right because one you're running at high power all the time as opposed to a smartphone chip which really doesn't max out very often. These you know AI chips are being run 100% for as long as they can and so you need testing becomes more important also very important because everything now is getting packaged together so you're packaging multiple chips together you're bonding them on top of each other. You start with a chip that's already expensive you bond it to a couple more chips that are just as expensive and you package them all up all up that's that's a lot of costs in that you need to know before you start bonding them together you need to know if any of them are broken. Yeah because you're not just like you have a broken part in a smartphone chip you throw that away you take a little hit you throw away one of the one of these breaks and you're throwing away the whole module that's very expensive loss and so you want to test there's a lot more testing taking place in AI chips because because the they're all getting so much more. They're all getting so much more complicated. Yep and air has a good solution for one aspect of that so I made a tactical error and that I had a report on air ready to go two months ago because I've been tracking them for a while and while I don't want to come out and necessarily say that. They represent a fundamental inflection point for some semiconductor testing you could make a very strong case that that's the case for things like advanced packaging and optical for example optical is an area where as we shift to silicon wafers and I had a similar report at this time that again I wish there's only so much time in the world to to write reports and publish on the diligence deck but I had another one showcasing the trends in wafer production. Specifically looking at silicon wafers and a how difficult that's going to be and be scaling that's going to be a real problem but part part of that is because there is a lot of waste it takes a really long time in your learnings that your silicon foundry whether that's TSMC coupe or with sour global foundries you could say six to seven months in responds air helps solve that like air helps solve your your efficiency rate your known good die in a timeline that's reasonable and it's and it manifests itself mostly in brand new kind of processes and so they called out optical and my point from day one when I was coming you're going to see air dollars start to scale and ramp prior to seeing silicon photonics scale and ramp which is a lumpy market and remains so now but I've always liked them within that view but I also think it's true of some deep advanced packaging and oh by the way not immediate but next generation memory as well like they called out memory there's a lot of questions about memory do you think you guys will be relevant for memory and it's not necessarily with hpm now although they're trying to have hpm but like 3d stacking a whole new new parts of memory transistor like once you start to have to do things that are disruptive to your process and those take a lot of time to learn and you need to make sure this is where way for burn in starts to become very very interesting so all of that to say I've largely viewed them as a very early indicator for some of the more complicated things coming in semis and that's what you're starting to see manifest again we're not saying this next year but on these timelines I think they're the best beneficiary a mini a mini inflection point to next generation packaging process materials etc. I agree I said this last week or two weeks ago is we're going to have to do a test episode soon so actually let me put that out to listeners if you have anyone you think we should have on as a guest this seems like a really good guest we should have someone come in and get nerdy with us about test because there's a lot going on there's lots of different kinds of test there's all kinds of different ways to do it one of the one of the questions I asked in my my sprawl last night across AI across Claude was you know are there other ways to do this and Claude let me know that was a really good question. Congrats Jay what a great question. Yeah and there are lots of ways to do test and you know I know from experience is test is painful like if you're if you're very it's very painful it's it's kind of manual ish and that and not you know and every it's hard to do right you just think about the complexity of how to test it it's very hard and then what you're looking at and right all kinds of things and that's part of what makes air interesting is that they have they have like basically a consumable of razor blades. component to their revenue, which people like, is it sort of a recurring revenue. But it's all very complicated and it's painful to do, but it's very clear we're gonna have to do more of it. And I think I think that's like air is a good example of that. There are others companies like TeraDine or you know - Yep. - Advanced or all other - TeraDine and Advanced are the two others I recommend people check out if they're co-hus and other ones COHU. - They're a whole bunch. - Yeah, there's a lot. - Agreed. It's in the weeds, but 100%. Yes, there's other ways. All that to say, the hunt, the needle in the haystack for anybody who wants to get in the weeds of this, is some of these test companies for the forward indicator they are, too much more complicated semiconductor designs and process and technologies that are upcoming. That's my pitch by T's. - Yep, more test. - More test. Anyway, just to just to just to find my own strategic error of not producing this note, which would be very good and I and I will get it out. I just was like there's gotta be a line that I draw on how deep we need to go and do I really want to get into testing because that could then open up all these other Pandora's boxes like you said and so I was like, I'll have to do it at some point, but in retrospect, it would have been very prescient, but so be it. Is what it is. All right, let's talk a little bit more Nvidia and stay. No, no, IBM. I forgot about it, Jay, for the entire reason that we wanted to bring up this point, that we were going to talk about the train wreck that is IBM. So, I said, who I've felt for a long time is like the most disruptible company there is on the planet right now, but so be it. So apparently there's this company called IBM that still makes mainframes. They had a very bad quarter. Again, because of AI, it's gonna be a answer for everything. What that happened because of AI. And personally, I think that's all we need to say about them because it is interesting. I mean, the most interesting thing to me about IBM is to the extent of how one interesting they are and how little they matter to the main conversations everyone's having. I mean, for two quarters, I've sort of danced around this like sure, they've got a consulting slash ISV business. People have talked about their model. Like nobody talks about Watson anymore. Like, it's not going to get used in enterprise. Like, oh, they're still relevant in big enterprise. I, every bit of their business is going to be hurt by AI. 100%. And to be honest with you, Jay, like this is one of those, how do I, how do I put this? No one likes working with this company to begin with. So like, if you're looking for who you're going to replace, start with the vendors that you really dislike. And IBM's one of those. I'm, yeah, maybe I didn't, maybe I didn't put that as gently as I did. Yeah, I can't, I'm going to be more diplomatic and just stop talking. I think we said all we need to say about IBM. That's okay. Because there's other stuff. There's, there are other interesting things taken. And it's not just us, our friend Ben Thompson also had a very negative note on them. He spoke, he spoke truth. So if you haven't seen that. But I have, I have not seen, even in some of the post notes, I have not seen any kind take away on this. Like the emperor has no close situation. It's, it's probably going to be bad over time. Leaky bucket. Leaky. I mean, yeah, whatever. Use your, use your favorite economic framework. Okay. Yes. Let's move on. All right. Let's talk about Nvidia state. Oh, market. Why does the world, why do all of the stocks hate semiconductors again, Jay? This is, this is unbelievable. So, so I, I have this, this thesis that I'm kind of tickling around the back of my head, which is every cell. All right. So obviously, there's a lot of demand for AI, but still unproven. We can, we can talk about ROI, but like it's, we're still, the world's still trying to figure out what, what we're going to do with AI, how much money people can make from it. So there's a little bit of unease out there. Certainly a lot of people asking these questions more vocally. And I think one of the things that's concerned me for a while is the extent to which Nvidia is propping up the, or at least supporting the, the Neoclado ecosystem. Sure. And I came right, was last week or earlier this week, they announced a new partnership program where they're basically financing, directly financing data centers, right? And they've had it in place for a while. I've talked about it before. They have now $30 billion of backstops in place. They're not on the books. It's off-balance sheet financing, but it's in the footnotes. So you can find it. But $30 billion of backstops for Neoclado's, what they've now done is formalize that program and are going to start doing it more aggressively. I think they have two, two customer, two Neoclado's already, two customers who are going to be participating in it. And Nvidia will give them backstops when they buy Nvidia gear for their data center. Nvidia will backstop a certain amount of demand, right? If you can't sell these GPUs, Nvidia will buy them. And then Nvidia, I think, now, to share. Red them for their use. Red them, excuse me. We'll now buy that capacity, I mean. And if not, if you just, if the Neoclado is selling them, Nvidia will participate and rep through revenue share. Which to me is, again, we're bumping up against vendor financing. Like, are they supporting demand? Are they sort of enabling things? Or are they actually creating that demand? Because why couldn't these Neoclados get financing on their own terms? Why do they need Nvidia support? I know some people, like Semi Analysis, had a good piece on this, basically saying they're just a failure in the capital markets. And so Nvidia stepping in, because others don't know how to do this business yet. The other side of that argument is that Nvidia is creating demand and that the reason capital markets won't invest in it is because it's not investable. I'm somewhere between those two. It's still, it's very messy and complicated, but it is concerning. It is, like, why does Nvidia have to work? If AI is so fantastic, why is Nvidia having to work so hard to get these deals done? And then on top of that, we have SpaceX. And then Meta, we talked about last week who are sort of leasing out capacity. SpaceX is, we don't really know what Meta is doing exactly, but it kind of looks like that. And I think that has catalyzed a lot of the sort of more bearish takes on AI, that we've seen a great AI. And I think even more than that, even more than sort of fundamental reason, I think there is just a sentiment. I've I've I've sense from the buy side, which is it's just time to look at other things. Like there's like that's why every and so what happens is in the last two weeks pretty much every semiconductor stock is down. Good results, bad results, everybody's down. And it has nothing to do with fundamentals. It's just the the buy side is rotating into other areas. Not because they think anything's like there's a crash happening. There's nothing doomsday about it. It's just correct. There has been a lot of exposure. Everyone's made a lot of money on AI trade this year. They're taking money. They're taking profits and lowering their exposure. I mean, because I think just mechanically a lot of you know a lot of funds are very heavily over indexed to semis because semis have done so well. And so I mean, just out of that. Yep. So I don't I don't think we're I don't think the bubbles over. I don't think there's doom and gloom, but I do think it's just there's a lot less enthusiasm for semis in the investment class than there used to be. 100% and I think that's why like I was just interesting. I got a lot of inquiries around air after the fact. And I think this is signals of early optical trade too is that people were just kind of looking for the smaller companies who didn't have a huge baked in valuation. We're up and coming. This is also why I feel like so many people are chasing down the native gritty of the supply chain. And even not so here like a lot of investors that we talk to look at Asia things like PCBs, obviously MLCCs have come into that like these are they're like trying to find those little gems who have a run, but everybody's not into yet like they non obvious names. In fact, I think this is interesting. And I know you had a weird exchange with somebody on Twitter around your one of your positions for Nvidia. But but I think it was maybe this time last year, maybe a little bit before where we asked is Nvidia kind of at peak trade. And from that time, because I looked it up, the stock really hasn't moved. Like it stayed the same. From that exact time that we asked is Nvidia at kind of peak. It's gone up some, right? But it's not been 120% like like the year before. But they've said Again, because those are your obvious beneficiaries. And that's not to say, I can really agree with you. That's not to say that people aren't like, oh, and Vitya's gonna continue to make an unbelievable amount of money the next two to three years. That's agreed upon. It's a lot of people are in it. That's the obvious. Exactly what you said. It's time to look at other things who might be the, and to some degree, I'm curious if you agree with this. Investors largely also have kind of like, sometimes short-term memory, not all, sometimes short-term memory. People are used to like 80% gains, 90% gains, if you look at a stock, you're like, oh, it's up 10%. That sucks. I'm used to like an unbelievable amount of money. And so sometimes I think that color's there. Who's the next 100% growth stock? Like fine, do the exercise, but we're coming off a period of a pretty unprecedented returns. Yes, I think that's right. Is who's the next 100% stock? Maybe there isn't one. Is the possible answer, right? There is. Sure, agree. Yep. And I'm not talking about any specific stock. I'm not gonna talk about my call on Vitya or any other, but I'm just saying like, the markets had an incredible run. It just doesn't mean it's gonna keep doing that. No, entirely. So, you know, the NeoCloud stuff is interesting. I'm gonna make a point and then you can tell you might completely disagree with me. And that's totally fine. But like we released this week, a model slash report that I'm naming gigawattonomics. And what's interesting is, and this is an NVIDIA point. So while NVIDIA is the more expensive CapEx play, it costs you more to buy NVIDIA than it does other things. The economics of the token rates that they give you is still very, very good, if not better or on par with any cost you'd have for custom ASICs. And why I say that is because to the degree that the NeoClouds are dependent on NVIDIA and they are, I do think that the demand justifies their build out. What I think is hard is exactly what you said is, who's the long tail customers for the NeoClouds? So the point I was gonna make is, let's just say, and this came out because of Kime K3 this week, which we just added to our benchmark at csbench.com, if you wanna see it, it's a good model. It's very strong and reasoning. It is not quite taking over, fable or clod yet for knowledge work. There's a debate about this for coding and that's fine. But people were like, oh, so does that mean we should be bearish overall compute cap X. And I think the take is no, and I'll explain why, but you could raise some questions around Anthropic and OpenAI and that's fair. But even if open source model, let's just say that we're on the timeline and an open source model becomes the predominant way that stuff gets inferenced and at scale by every enterprise. And that's the model that somebody's running and compute costs drastically lower. All that does is change, that none of that changes the demand for compute by the hyperscalers. And I would argue, this is where you can tell me if you disagree, that if only the three main Cloud CSPs and let's add meta to this, if they were the only ones spending on this Cap X cycle, and I get that there's more, but they're the largest spenders. Meaning that they were all that's left. Perhaps they used the Neo Clouds for spillover, perhaps they weren't even spending OpenAI and Anthropic weren't spending the same kind of money. If it was just the four hyperscalers, I think that alone, you would still have more demand than supply for compute infrastructure. - I agree. I think the issue though, as everybody says that old saying is the stock market is a discounting machine. And you look at the market right now, you say, all right. The big seven spenders are gonna spend 800 billion this year on Cap X, one point something trillion dollars next year. Yep, we've known that for six months now. What everyone is looking for, at least what investors are looking for is, okay, what comes next? 'Cause we know all this, right? And so that's already factored in. What comes next? And we're at a moment now where there's lots of questions about the models. We had a lot of excitement over Opus and then Fable, mythos, all the new models coming out that can do a lot more reaching a broader audience. But that sort of, that excitement has been there for a while, it's tapered off. People are starting to, again, ask these questions about, like, all right, where is the value of curling? Who's capturing this value? Is it all going to Nvidia and the memory companies? Because at some point, those big seven have to show a return on their investment. And just on the numbers alone, it's not clear that it's there. Right, let me be more blunt, it's not there. Right, we are seeing Anthropic and OpenAI, well, we're not seeing it, but certainly the, all the press reports of leaked financials are, they're, they're seeing immense revenue gains. But we're not seeing that from the other five people who were spending on all this CapEx. And I think what we would need to get people more excited about AI is some interesting new things to do with it, some whatever the next big advance in the model is. Right, and this is what Kimmy, it's been such a big discussion point this week is because Kimmy has seemingly caught up with Frontier models. And so that tells me that the Frontier model labs, Anthropic and Cloud in particular need to keep spending to stay ahead because they're correct. There's not a great mode on the model. And yeah, they're both probably going to go public this year, maybe, so they're going to have money to keep it going. But they're, they're not profitable. They're not in close to profitable. And so it's, you know, at some point we're going to run out of money or, or AI is going to have to become so useful that this will all seem like a silly conversation. Yeah. And my, my thesis all along has been, AI will be that profitable and do that meaningful, but it may take longer than is currently factored into the market. I, I agree. And I think this is why, I mean, I could be totally wrong, right? But I think, well, one, I think the hyper scaler's report next week, right? Or at least a couple of them. At least a couple, right? Maybe not all. I can't, I can't, maybe they all do on the same day. And it's hyper scaler, I get in again, like it was last year. Last quarter. But I think everybody just again, and I don't think, I don't know if this changes the market trade rate. But I think everybody wants to know like, where do you see CapEx going? Like, are you still confident that you'll see returns on this capital? Are you still confident, right? That you're not going to cut CapEx anytime soon? My assumption is that, yes, they're still confident and they're going to present that tone and they're still going to increase CapEx to the degree that they increase CapEx. I don't know. Maybe there's a whisper number of like, well, if they don't increase it this much, we're freaking out. But all of that to say, and this is what I'm just going to give this data. This is what our gigawattonomics model basically showed out. If you, for, let's just use it for an Nvidia Rack scale cost. So the cost that they would spend-- and I used a gigawatt facility because that is A, the unicorn. There are not a lot of gigawatt facilities out there, most are vastly smaller. But let's just use a gigawatt facility. Or you could just do this on a gigawatt of compute. We know how much Amazon, Microsoft, and whatnot have of gigawatts. At $6 an hour and average and 75% utilization, they can pay back their Nvidia CapEx on just compute CapEx in a little over two years. Now, the hyperscalers get, spoiler alert, a whole lot more than $6 an hour on average in long-term contracts for Grace Blackwell. It's double digits. And which would tell you then that they can pay that back in less than amount of time. Now, again, they do not have a massive amount of Grace Blackwells yet that will be installed this year. If you believe that everybody's number, that Nvidia will probably ship north of 10 gigawattts of compute this year. So roughly 65 to 70,000 racks. They have a lot of H-100s. who I actually think are pretty profitable at fairly low. you know, good rates. I think what everybody wants to see though is for that to then transition itself into a free cash flow again and my hunch is again Ben could be wrong this trough for seeing in free cash flow is very very short-lived We'll return very quickly If we're right about inference margins cap X costs dollar per hour Now again, I say that's to that's to pay back the Compute infrastructure costs But however at that metric I just gave you at $6 an hour average 75% utilization it pays back a full giggle off facility including power and shell in 7.1 years So do your math depreciation on whatever you want to do on life cycle of shell is it gonna depreciate of 10 years? Yadda yadda yadda, but I focus more anyway all that to say It's possible it seems likely They'll return to capital how long fine question averages yes as long as demand keeps going up And we believe that every enterprise will pay premiums for this as they will their employees will benefit and you'll somehow monetize a consumer that every human on the planet somehow one of these people monetize a consumer with a computer using AI I don't think this is a problem, but I get that that's a timeline that we don't know the answer to In my dark moments Usually when I'm driving my dark moments I I question that because We really need is to see broad-based enthusiastic adoption and I'm like I'm one of those I'm an enthusiastic Adoptor I use it all the time I think enterprises are kind of getting there, but they haven't totally figured out yeah correct and then you know It's gonna take them a while because they have to for their compliance. It's curing them. Exactly. Absolutely um There's probably something there I just don't know I think the the real question is how How much are they willing to pay and that's why for the last two weeks we've had this sort of broad debate everywhere about People putting companies putting limits on AI usage by employees. Yeah, and and to me I think I think it's overblown. I think companies are just trying to figure it out. It's not like they're gonna stop spending absolutely But they're they But I also think there are legitimate questions to be asked about what people are gonna do with this AI in the workplace and really does it does it extend beyond You know, I mean a A lot of corporate work in this in this world certainly in this country, but probably in the world a lot of corporate work is Preparing is is going back and forth between excel and PowerPoint Right, let's have a meeting to talk about the big meeting right and A lot of what happens in corporate America today. This will shock you is not predict not terribly productive And I know there are people who that's I'm hitting a nerve here And so is AI going to make that easier? Are we going to get to a world in which oh we have I have to go to all these dumb meetings this week And I have to have PowerPoint for each one of them I just did all that work in 10 minutes and now my life is much better or we're going to world in which Everything's gonna be creating more and more PowerPoint Well, but well the PowerPoint AI arms race and we're not gonna get anything any more productive. Yeah so I think you're you're absolutely right. I mean from our research on this right. I think I've said this before we know that really the The top 20% of more aggressive and ambitious enterprises. So you're early adopting enterprises are the ones Trying to figure this out right now deploying trying to solve problems seeing what their token spend will be it has not diffused But I but I will say there's an interesting parallel and y'all can just make this what you will Uh, to my own experience in the mid 90s being a High schooler and a very early adopter of the internet in my computer science class Who ended up starting a business my junior year? um Convincing every company I could to have a website because you know what we made websites and we made them quick And we understood that like I don't care about your business, but if it's on the web I might care about your business. So it was like early, you know, forward deployed engineers at that time trying to go sell I find it interesting That there are these pockets of companies like one got funded and I don't remember who it was who was basically doing the same thing We're gonna send young people into your organization To basically evaluate your entire enterprise and then figure out where agentics gonna make it better And I make these parallels because Jay you you will I'm sure you can understand this but I would go have a conversation with a local business Because I was also selling this right as a junior in high school and I would say It I'm trying to convince this person your world will be better with the internet people can discover you And some people just didn't want to do it. Just didn't get it nah I'd rather just call people up. I'd rather than find me in the white papers Like bro you don't understand what's coming and I feel like we're in that right now with agentic People it's going to take time for every business to adopt this and and maybe you're right maybe again our assumptions are that we need You know does a does a does a you know gigawatt shell pay for itself in four years seven years or ten Yes, it does depend on The adoption of large enterprises and everybody I don't know how long that's gonna take I just know that It's going to happen because this is transformative the way that the internet was transformative But it could take it some of these people are very backwards and you're exactly right like I talked to friends So wait, you know on both sides of the spectrum I talked to companies and friends at these companies where They're just using an obscene amount of tokens and their IT is like what the crap's happening stop using so many tokens And then I have other friends who are like IT's telling us to use more tokens Because we're not using enough and you're like Just just the parallels of understanding how to use this workflow like That's it's gonna take time diffusion technology of diffusion takes time and I can't tell you the timeline. I can just tell you it's happening It's diffusing and maybe it takes long maybe it doesn't I don't know Right, and I think the fear is that the timeline of diffusion that the street has priced in is very different than the Timeline of diffusion that will actually happen. Yeah, and I think that's fair totally fair Right, we don't know if we if we knew a man we'd be able to make these bets and we'd we'd build we'd be billionaires But we don't know but adoption takes time and it's happening and This is again why I spend a lot of time studying the customer at this point, which is the enterprise To understand how they will deploy and what that looks like because that to me if we can just get a handle on that We get a little closer to some semblance of a Insight as to is this stalled is it happening like where we at any adoption cycle so There you have it Hey, you know, I actually have a good metaphor for AI I've been you know, you know how we're always on an allergy. We're always trying to think of like oh this is like This is like the internet or this is like the railroad. I always say electricity All right, all right, but when I've kind of been using for a while like this is like telecom is like 3g build up now This is like gyms Right you got to you got to put a lot of money upfront you got to buy the building you got to fix it up You got to buy all this fancy fancy equipment Right You got to have all that and you get people to sign up But then what you what you really want is for them to not use it That's right That's right so that's my it's you know AI is the 24 hour fitness of the modern era Well 24 hour fitness is are going out of business so maybe not but gyms exactly right and then you know people are like Oh, I got to go to the gym. I got to sign up for the gym membership. Yeah, and then never go that's that's the business model all the AI companies one Yeah Joy Joy All right everybody Thanks for listening we will talk to you next week As we will cover the kiperscaler capex build out news and see hopefully not a lot of drama But we're okay with some drama makes for some good things to talk about so talk to you next time Thank you everybody. I got to go to the gym now same Have a good one bye