EP 181: Cerebras Earnings, QCOM Investor Day, Micron Earnings and the Memory Mafia

Hello everyone. Welcome to another episode of the circuit. I am Ben Beharen. Greetings, programs. I'm Jay Goldberg. Well, this week was an up and down week. One of the downs was the tech industry learned that O'Mallock, a long time staple in the tech industry, arguably one of the earliest bloggers and probably the most influential of bloggers. I think the whole blogging thing happened largely because of O'Mallock and his popularity. He passed away on Wednesday of heart complications. And everybody's been sharing some of their stories on Twitter. I've known O'Mallock for a while. So I'm every year at multiple Apple events. He was always extremely kind, very, very generous. I know he mentored countless of top tier tech media. And was just an overall, overall great dude. I just want to send, we want to send our consulances out to his family and loved ones. O'M was an important part of the industry. I read his site, giga O'M religiously every day for years and tried to think and write like him and he'll be missed. And he was important. One of the, he was one of the first and one of the most important influences on how we conduct tech journalism tech, tech writing today. Yeah, and if you aren't familiar with them, I'd encourage you to take a, do a search for him on Twitter. You'll see kind of everybody sharing stories. And even some of the blog stuff, he's been pretty active off Twitter, but it's written some good essays. He was also a columnist in the New Yorker, which I always was like, how do you help me get a column in the New Yorker? I could put your name in, but he was like the only tech actual journals with the column in New Yorkers. It was pretty epic. So anyway, sad news, but great legacy that he left. I know everybody, I'm so appreciated. The impact he's made on their life and the tech industry. So, and as many thoughts go out to his family. Okay, this week earnings wise, we're kind of going to go in chronological order. There was a very keen interest on Sarah, Briss's first as a, as a public company, I will say I was on the road. I saw snippets of obviously the earnings release, but really the conference call. But I'll let you unpack it, Jay, and we can talk a little bit about the position that Sarah Briss is in. They got a little bit of good news with the launch or the tease of opening I 5.6, which they're running a just shy of two trillion parameter model on Sarah Briss at higher token costs. Premium tokens, but that's happening. So anyway, I'll let you give the spiel, the quick spiel. Yeah, so they reported an earning and earnings. This is the first earnings as a public company. And it was one of those prints that was if you just, if you just read it as it came across the wire, everything looked fine. But fine is not enough in today's world. And I think the the the chief concern was around their margins. Lots of questions about their margins. Sarah Briss, of course, has this very, I don't say very unique architecture like they don't produce chips. They produce entire wafers that operate as a single chip wafer scale engine they call it. And there's lots of questions about how that works and how you package it up. And it it looks like it is very expensive to package because their gross margins were not where they should be. A lot of complexity there, not quite clear what happened. But the street didn't like their margins. They of course just went public a few months ago and it was very well received. I will say that your first two quarters as a public company are incredibly important to get right because you've brought in a new group of shareholders who don't know you the way your venture backers do even if some of your venture backers were themselves public funds you usually have a different set of fund managers you're working with. And you need to work really hard to establish credibility. And they something something got lost in communication there. And so it's going to be a little bit of a struggle for them. They need to sort of amp up their investor relations work in the next quarter they can turn around. But there's a lot of there's a lot of questions about this right. There's a lot of moving parts there. They I think they felt they they did really well they have big deals with open AI and AWS those are coming on stream. They say that they're having so much demand that they're having to go back to their previous customers to rent out capacity that they rented to somebody else or they sold to somebody else and they're having to rent it back. And so that's part of the problem with their gross margins. And again, like they're there you know their only real customer is a series of entities of the UAE government. And so and the UA government those same entities are also the larger shareholders. So it's a weird it's a weird mix. But right remember the thesis that a lot of these companies are trading on right now is around this idea that inference is hot. It's in demand. There's not enough compute to handle all the inference that people want to do. And Sarah brush to be you know top line numbers were pretty good. So there's. They just I think the problem is they have to work out how to communicate. How their economics work. I think it's unclear to a lot of people. It's very very different than traditional chip companies. They also run a cloud service business. So they need to sort that out and how to communicate that. They you know I don't want to say that they should be doing better because they were doing pretty good. But I think there are a lot of questions about the the like how this model is actually going to work in the real world. Yeah and I think they guided you know obviously they they guided margins down slightly which again range of factors like you said. Having to rent back economics. I mean obviously it's been rumored you know that their waifers aren't or their yields aren't great. They need to get yields up. But where where this interests me is you know they are either. They are the proof case or they're not for this idea of premium tokens, which is what you know in video wants to have happened. And I think other vendors want to have happen. If you can accelerate inference even more so. Then kind of standardized chips right at a higher level because of your architecture and people are willing to pay for that in mass. Then that's kind of a validation also for you know the business model and video wants to use with GROC and enabling meal clouds and others to sell. You know dedicated premium token inference generators. So it you know again I think their success kind of validates this idea because as we see right now these are not cheap tokens running on on server yet so. So anyway I'm interested to see if it's just validation for that space because we talk about it a lot and we don't really have a whole lot of evidence that companies are going to play premiums for just faster inference yet. Yeah again I think I I push back on that idea I think what Sarah boss is trying to say is that they have very fast inference which is very good for certain use cases think speech recognition and conversational chat bots. And they're you know you could you it's pretty clear from how their architecture works that that's true like they have a latency advantage here in inference. And whether or not that's actually going to be priced at a premium or how how that price is going to get sort of set by the market. Is to me unclear I ultimately I think that the provider the hyperscalers and the meal clouds who sell that service will segment it differently. We'll see how this all you know it's it's obviously a lot of moving parts here a lot of things constantly changing. Yeah no agreed and I agree with you there there comms the way they tell this story also if anybody from server is listening. You know drop your earnings release on time things little things you know yeah it's the little things that do matter so anyway we'll see and and you're exactly right it's now they're going to need a couple more quarters than normal to keep executing and validate that they know what they're doing when it comes to earnings releases earnings calls questions with investors etc. Okay well it's exciting nonetheless I we've been waiting for that one so okay also this week was qualcomm's investor day I was in attendance there had some meetings with management afterwards got to meet Tony Pialas for the first time who was the CEO and founder of alpha wave. It was it was interesting to me and I'm actually very interested you know for your take because I knew there was a lot of skepticism going into this there's obviously a lot of kind of concern. are concerned around Qualcomm, their focus. They've been in this position of trying to have to convince everybody that the handset business is not the most important business that you should focus on, but the problem was up until now, the handset business was the largest business, but not really growing. And so as great as auto is and IoT, those are not 25-ish billion dollar businesses a year. So the adjacency story has been kind of where they landed on and their investor day last year was again, really trying to just talk about the growth and the adjacencies. But they had not been in kind of the largest pool of money that anybody's going after right now, which is data center in some capacity of accelerators and custom A6. To some degree, they also make CPUs. So they're kind of stating that they're going to be there as emergency CPU vendor as well, kind of like ARM. So anyway, I heard everybody's skepticism, and I think that's certainly fair. At the same time, you well know, this wasn't their first attempt at data center, even if everybody who was in that group is no longer there. You and others have told me that what they did have was pretty good, and ARM was just early, nobody's going to dispute that. They do have a good custom CPU architecture in Orion under their own ISO. So there's technical merit there that has proven itself. But they talked about custom A6. So there is a, they've hinted at this. We teased this a couple of times talking about all right. We want to hear more about it. They didn't name a customer, but they have a custom A6 business that they say they have two customers for that are multi generational programs. And then they launched their kind of accelerator roadmap with A100, which they have, A200, which is coming. And a new technology they call high bandwidth compute that is basically just near memory compute for everybody who follows the industry. And that's really just 3D stacking of memory on top of logic. And in this case, it is just DRAM, not HBM, and not SRAM, stacked on top of logic, which again, in itself is not new. This has been R&D forever and Nvidia has talked about it. AMD has talked about it. I was at a LAN event with their CEO. They had startups there all talking about it. D matrix, this is also their approach. So not new, just the first time I guess we see a commercial roadmap for it, and potentially, right? Making use of kind of this new approach to memory. And then again, they're focusing it purely on the decode part of the flow. So they were very clear. This does not replace your accelerator. It sits as the companion to the accelerator as a part of that decode pipeline, which, though, however, requires a lift and shift in software, which is why they also announce that they are purchasing modular. And the modular guys were on stage, both co-founders talking about their excitement to be at Qualcomm, what they really want to do, and just making software agnostic, architectures agnostic, so that anybody can port workloads, do set lift and shift. So really, the underlying kind of strategy there was, hey, we've got this new thing. It is a new architectural approach. It's going to help speed up your inference pipeline, but it's going to require some software work to do so. And here we now have this asset that should hopefully make that easier. I heard a lot of positives from the industry, kind of inciders around modular. Again, we'll see if that works. But we can go into it. I'm curious, you're taking-- I have some deeper points to make on the economics, but I'll pass it to you now. So let me pause for a moment to initiate diplomatic mode. I think on the positive side, I'm encouraged to see they have a complete roadmap. I'm encouraged to see that they will have a CPU, obviously, in big fan of ARM-based server CPUs. I particularly like the fact that they are naming the CPU Dragonfly, because here's a little insider based by the original Centric CPUs way back in the day, all the code names were Dragonflies. So I like that they're keeping something there. It's a callback. I was also impressed that they had a roadmap beyond just compute. They also have some sort of basic networking products, essentially active electrical controllers, like Astera does. So that's right, and they're getting in the market. And they have customers, and they guided to fiscal 29 revenue of $15 billion for data center. So that's a big number. It's $1 billion this year, $5 billion next year, and then growing to $15 billion in revenue by fiscal 29. So that's all the positives. I think the concerns that I have are first and foremost, I don't think they fully grasp the software lift that's ahead of them. I can say with the experience that this is something that is-- it's outside of Qualcomm's DNA almost-- I want to say DNA because they have software teams. It is outside their comfort zone. If you were designing a CPU, you need to have a very significant amount of software engineers, a third, at least, of your engineering force, asked to be software engineers. And I don't think that's fully grasped by Qualcomm. It's just very new market for them, and I don't think they've fully gotten there yet. And they kind of software, they've really done best at around telephony, around mobile phones. That's all standardized. And so it's a very complicated multi-party solution, but it's finite. I think AI software is very, very different. Look at the markets where Qualcomm has done software well, which is telephony, cell phones, and arguably around automotive. Those are very closed, very tight ecosystems. They're very controlled by the participants in it. And the AI software ecosystem is not. And I think, yes, they have modular. Modular team is a fantastic pedigree team. But this is a hard problem. This is a really, really hard problem. It's going to take a lot of work. And just throwing-- you can't just sort of say, hey, we have chips, and we have modular on top of it. That is not how this works. And you look at-- Nvidia's been working on CUDA for 20 plus years. AMD has had Rockham for call it five, six years now. And they're still like-- they've gotten a lot better. But everybody knows they still have a long ways to go. And I just think just buying modular isn't enough. And then you sort of throw on top of that the fact that Qualcomm doesn't have a great track record when it comes to integrating acquired companies. Notable that the week before the event, the last senior new via employee left Qualcomm to go do something that's competing with Qualcomm. So right. And on the networking side, like let's remember Qualcomm bought a Theros 10 years ago, more than 10 years ago. And they're just now starting to revamp their networking efforts. So they have a good-- a bunch of products. They look encouraging. But it's a long road map. The CPU is available until 28. And then they have a lot of other-- they have a lot of hard work ahead of them. So I think a lot of this is still to be proven. Yeah. And I think that's the fair take. I kind of brought that exact same sentiment to Cristiano and a cost. And they were both basically like, we're happy to prove it. So they-- so here's the question I have for you. They seem very confident in that number. OK. You guide to a number. I have to imagine they're pretty-- you don't do that if you're not like, pretty sure you're going to hit a number, right? I mean, just give me CFO speak. What's the risk in it being horribly wrong? It's a lot, in my opinion, right? Because again, they were very confident. They were extremely confident. And I listened in on the financial analyst QA. They kind of said the same thing. They were pounding the table. And then I'll tell you, I'll tell you while I'll explain why I think the mechanism for that number is because it's not the accelerator or the merchant CPU. It's something else that we'll talk about. But just give me the read. I'm like, you say something like that? I mean, there has to be something, right? To guide that much. So-- Okay, a couple of things. One is it's a very long-term forecast. It's four years out, three years out. Right, 'cause the exciting part was that 15 billion and fiscal 29. That's two, right? And that's two, three years away. Nobody's modeling that out yet, right? That's one. The, let's not forget that three years ago they gave a forecast for PCs and they didn't hit that. They didn't get close. They didn't say this obviously, they invested it, but they effectively backed off that number. I won't bore you with the math, but I backed it out. They've lowered that forecast. Their internal forecast for PCs. So you can put out a number and be wrong. Right, that's why they have safe harbor statements at the beginning of everything saying, we don't have to update these estimates. These are just our best guess. The other thing is when pushed, somebody asked him this question, like how certain are you of this? And when pushed, the CFO Akash said, these aren't, he conceded that these are not signed finalized agreements. And that gives me some pause. They said that they have two customers. They only named one. I think everyone at this point knows who the second one is and it's curious that they didn't announce that one. Like what, you know, a lot of people know who it is now. They didn't say it. They didn't announce it. So that's like, was it just a timing thing or was there, was the contracts, you know, was there still a lot of big gap in the agreement? I think what they have is the near term stuff, the billion dollars this year and the custom ASIC in quotes work that they're doing this year and next. That's probably locked down, but the bigger numbers aren't certain yet. And so, I, you know, if I were to model out to fiscal 29, I would, you know, just mechanically, I wouldn't give them full credit for that. I'd give them something, but not the full number. And I think it's important to compare this to what, what arm set, right? Because that's what I was going to ask. Arm also gave out even longer, like a five year forecast, fiscal 30, I think. And they're going to another big number on that timeframe. But they said that these are all finalized agreements. This is all in numbers. Like these are finalized agreements. They said that. And when pushed, they could walk you through those numbers, right? Between what they said and what I've, some stuff I picked up, I know how they're getting there. Right? I know I can, you know, they obviously didn't give units a pricing, but I think I've been able to back into those. And, you know, I can model that out. I don't know what the pricing is. I don't know what the units are. I don't know how I would get to 15 billion dollars for Qualcomm. Yes. Okay. I don't disagree with you. And I think you're right around 29 because no one's modeled that long and a fair amount of their assumption where they said non-handset revenue would be 40 billion. And then, you know, to analyze their multiple, you basically need to figure out what you think. Total revenues would be me, which means you kind of have to take a stab at handset in that timeframe and kind of what it's growth is. So I'm with, most people haven't done that. Like I could ask anybody on the street and be like, hey, do you have a idea of what Qualcomm's 29 mobile revs are going to be yet? And they're probably like, no, because they just haven't done it. So I just, I want to talk about the custom stuff because to me, that is the closest upside in revenue. And I'm coming off of that from the viewpoint of what's happening with MediaTek with their big customer in that if no one's done this or seen these numbers yet, MediaTek is at this point in time, by most people who've modeled them, forecast to double the size of their business entirely in around 2028. So like double, the ASIC business will be larger than the sum of the entirety of their business this year. Basically, just because Google's scaling TPUs with them and they're getting a lot of value on the TPU side. Okay. I don't know if Qualcomm is doing much on the accelerator front for the two unnamed customers, for custom ASIC. However, I think they have engagements for what I loosely call custom custom ARM CPUs. And why do I call them custom custom CPUs? Because custom CPUs is basically what Graviton and what Axion and Cobalt is, which is an ARM CSS license. And you go work with mostly GUC, who's the backend designer for that. And they go and make your chip. And you just tell them how many cores, you know, you kind of optimize it yourself, but it's, you know, standard ARM CSS. So that's custom ARM CPUs. Qualcomm being an architecture licensee and being willing to use that architecture license, their custom ISA to make a custom custom ARM CPU. I think it's very interesting in this moment where CPUs are not just hot again, essential to custom hyperscaler ASICs workloads. And so here's the my logic is this. If you're designing, if you're going out of your way, your Google, your Amazon, your Microsoft, you're meta, your opening eye, whoever, right? If you're going out of your way to deeply customize a ASIC, a compute ASIC, why would you not also might want to do the same thing with the CPU? Why just go standard CPU? There's reasons I'm not disagreeing with that. There's reasons that ARM CSS has done what it's done today. But if you think that CPU plays such an important role and could use a degree of either performance or low latency or speeding up of course or optimization, Qualcomm I think is right there willing to say we can customize a CPU in lots of different ways that's not just standard ARM CSS and we are willing to go do that for you. And so there's nowhere else you're going to go to do that. There's no other vendor who does that. And that I just that to me seems, again, not proven. I'm not modeling that all I did in my model. I built for that was like, here's how much goes into custom CPUs today. So here's the dollar that's potential if they take some of that business and then you build your own scenario. But that's the thing I think they have readily available of interest in a rapidly hot and growing market in CPUs that no one else has, that I think would be very attractive to those hyper scalers that want to make that are already making and planning multi-generational custom CPU programs to do custom custom ARM CPUs. Why would they just get an architectural license themselves? Well, I think it's, ARM doesn't want people to really have architecture license anymore. Their business model kind of discourages that. I'm not saying it's impossible, but it's not their preferred. They are trying to make CSS. All better, flexible, kind of again, all the things I'm saying. But I don't think anybody disagrees, right at this point. You can benchmark Qualcomm CPUs. They're pretty good. Like they are a performance, they're low power. Prior to Nvidia, they were the closest thing to Apple in performance per watt. And that architecture is scalable. It goes from small to large. These are clever designs that even Apple does, right? When you have an architecture AMD does this too, from small to large. That is a technical merit design. So again, my only point is, and I also, to your question, it's hard to go and do your own custom custom CPU. Do a ISO. There is a lot of work there. And there's a reason why most of them offload a lot of that work to their design partners because they need help. They don't have all the IP. There's things that need to be done. So again, I think that has merit. I think that will become more interesting over time. We'll have to wait and see. I think if they come out and kind of prove or somebody figures out, because there's already rumors around this, that they're working with one of the big hyperscalers on said programs. I think people will take note of that. Because again, I do think there's technical reasons why a hyperscaler might want to customize an ARM CPU, even more than they can today. And like I said, Qualcomm's kind of the only place you can go to to do that. I think the opposite. I think that they are providing design services the way that L-Chip or GUC provides design services, which is fairly low margin. And I think the hyperscalers who I think are their customers have pretty good design chops. And there's obviously all this debate about customer own tooling and how much they need, someone like Google or say Qualcomm. And what's the difference? But Qualcomm has an abundance right now more than anything else is capacity at TSMC. >> Wafers. >> Yeah. >> Right, because they're not selling as many smartphones as they thought they would need it a year ago. >> Absolutely. >> Booking the capacity. So they're shifting wafers here. And I think I suspect that a lot of their near term custom, what they call custom silicon business, looks a lot more like an outshap or a GUC design service business than say what Marvell or Qualcomm or Broadcom does. Right. Now in fairness, in fairness, one way that I'm one thing that is different is that Qualcomm has alpha wave surges which are not as good as Broadcom's but are supposed to be pretty good. >> Yeah. >> I think that has led that's led the design. That's what's gotten them the foot in the door plus their capacity at TSMC. >> Agreed. >> But what I'm curious though about is, and this is why I think the numbers are important to dig through this for anybody who really cares about Qualcomm's strategy here is, you could also argue that alpha wave got them deals on the periphery. So kind of like Marvell, right, that they were on the IO periphery of some of those designs. But I guess the question is, could we figure that out, was it periphery or was it compute tiles and or CPU tiles when they report or will we not be able to figure that out? >> Well, unless they tell us. >> So they've said that at least in the near term, the design services businesses they've warned will be lower than corporate cost margins. And so, who can get 65% like Broadcom in that business, probably nobody. >> Yeah, you know, nobody. It's not even clear how long Broadcom can do it. So I guess the question will be how much lower are they? And we won't, well, yeah, they probably won't give us enough information to totally figure it out. But we'll see if they have a big gross margin hit, that's going to be the first suspicion is it's going to be that this is a little margin. But yeah, I mean, it's going to be horrible. >> Yeah, you're probably right. I guess it would be a matter of, if it's in the 20s, then I would agree if you, if it's in the 30s, then they're getting more portion of that IP. So, but who knows if we can figure that out? >> Yeah. >> Or 40s. >> Yeah. >> Well, anyway, I like I said, I have been curious on whether or not it's with Qualcomm. You're already hearing so much chatter in the supply chain of all the names I mentioned, drastically ramping their CPU programs, which is great for arm, regardless. All just to say, the priority is, I don't know what you want to say, moving to 50/50, they're ASIC and their custom CPU design, scaling like crazy. And not just again, validates a lot of the CPU thesis we've talked about. And I just think the beneficiary map is widening some in that space, including potential wafer is going to Intel with whoever that comes through to manufacture CPUs. So. >> And arm, and arm too, let's not forget. Let's not forget. Everything Qualcomm talked about at their analyst day was arm based. They're obviously not going to mention that, but arm was not shy about tweeting that out to remind everybody. >> That's correct. >> So they're asked to drag and fly from arm based on arm, arm IP. >> And it's true? >> It's true. >> Yeah. >> So, okay, anyway, interesting to watch. Like I said, the conversation was good. I talked to 15 or so investors. There, a lot of their sentiment was similar to yours. So just to everybody, it is a, it is a proof it. Let's prove it and watch it scenario over the next few years. But to be honest with you, Jay, I'm going to be so much more excited to talk about Qualcomm earnings now that there's data center than before when we talked about it. So at least like I got more to look forward to because there's more in this, talking about handsets. I get it. Again, they raised auto's guidance by 2 billion. That's great, right? That's great. But this potentially drastically changes their business. So now there's a giant spotlight on it. So every quarter, this one's going to be the, let's hear about what's going on in Qualcomm's data center. So I'm excited for that because we like talking about data center, not necessarily devices these days. As cool as devices are. It's not a gross story at the moment. Making this tough for me. Yes, I too will be excited to talk about Qualcomm in the data center. That is something I would like to see. Okay. Okay, good. We're on the same page on that. All right. Let's move to my cron. This is a strange week. I assume you believe like others did that there was some profit taking slash rebalance, mid-year rebalancing that happened, which was why the market pooped all over itself. And then my cron came out and absolutely shattered earnings. And there's a lot of goodness to get into this. Despite all of that, I don't know. Everybody still is confused about memory or unconvinced about, you know, the long term fine. We can talk about that. But it was an interesting reaction. Popped, corrected, closed this week, I think, kind of up for the week, but not necessarily where it was when it was at 12, 12, 12, 12, 20 to 12, 30. But yeah, there was good earnings call. I'll let you talk about it and then we'll get into the nitty-gritty. It wasn't just a good earnings call. It was a monster. I know. I mean, it was, you know, it's one for the record books. Micron reported very strong numbers and guided just crazy strong guidance, which may sound familiar because it's almost exactly what they did last quarter and to a lesser degree the quarter before that. The statistic that stood out for me most was their prices increased 60% quarter on quarter. Like you were paying a buck last month. Now you're paying a buck 60. I just staggering, right? And remember a lot of their businesses on sort of longer term contracts now. So that means people, you know, six months ago a year ago said, hey, yeah, we'll pay, you've already, prices have already doubled. We'll pay 60% above that. So I've lost track of how much memory prices are up in the last year. Is it quadrupled yet? It's just an incredible, incredible run. It's over 300% in categories. HBM is at least, but yes, it's, that's the minimum. It could be much higher. I can't, I didn't remember last time I checked, but it's over 300%. Yeah, it's up a lot. So once again, once again, micron has sort of saved the industry. Because up until micron reported, everything was looking very gloomy, right? I was on TV two or three times this week, I think. And every, every question I got was, is it over? Is it over? Is the AI trade over? Right? It was, you know, right? And like everyone can have their own take on whether or not the AI trade is over, but certainly micron for a moment reversed that narrative. Or a moment. Yeah, or though. Yeah, so it's, I just looked at a stock price up. It's, it's down on the week, down over the last five days, right? So it started, it was, it fell a ton on Monday, shot back up on Wednesday after earnings, and then kind of trick tapered down after that still, still above its lows for the week, but down for the week, even on a monster quarter. So okay, so let's talk about, yeah, let's talk about though that, so the nature of the contracts are changing and Sanjay went out of his way to talk about it. I think we had talked about the, what we had been hearing in the, the last couple quarters where customers were agreeing to capacity, but the prices were still a little fluid quarter on quarter, which we kind of read was, well, that means that they're going to go up. It sounds like now they are starting to add actual long term pricing to that capacity, which I think is interesting because, you know, put yourself in micron shoes, right? By the, if you were sitting there saying, hey, you know, we're happy to have you tell us how much capacity you want, but we're going to, that, that price is going to change quarter over quarter. They know it's going up, right? That's their leverage, right? Everybody sort of has that, right? And not that that's to change. But I think the fact that they are now saying, we'll add some long-term contract and that long-term pricing means that that vendor also believes it's not going to crash anytime soon. Because if they thought that the market would drop, next year, they wouldn't go, I'll pay you for three, four years, right? So, microns kinda like cool, let's agree to some and maybe there's a floor, maybe there's a top, I don't know, but I've heard a lot about like floors and caps fine. But for both parties to do that, my point again, the customer has to know there's no chance they're coming down or I would not agree to a three, four or longer year contract because I would rather those prices be better for me in a year and they probably won't be. And so the fact that I guess both parties are doing this just kind of in my mind, justifies the durability. But it also means, I think you can tell me as a true gross margins probably will come down some and I think that's okay as long as total dollars go up. I mean mid 80s is, and I looked this up, like I did the historical average. The historical average for micron was, low teams and margins and the historical average for SK and Samsung was high single digits, historical average and now they are both, they're all averaging between 75 and 85. So just to see like what's going on, an industry who historically rarely had strong double digit margins for any length of time is now probably going to have mid 70s at the very minimum for foreseeable future. - Yeah, I mean they have high gross margins than anybody right now in semis, right? They're high gross margins and Nvidia, high gross margins and TSMC, right? It's pretty staggering. I think they're going to make more in operating profit this quarter than Nvidia will, which is again hard to fathom. - I guess my question is, I saw there's a few things. One is sort of, I don't know, I mean the big question is what is gross margins, what are the gross margins going to be like for the next few, next year call it? - Right. - And I was reasonable, like that's a question too. What's reasonable? - Reasonable, left the room three or four quarters ago. - No, but no, you're absolutely right. What I was going to say actually was, I think what they should, what I would like to see them do is say have another good really strong quarter, really strong gross margin and then say, we think the industry is going to find us equilibrium at gross margins of X percent, right? And we're going to now target that and it's not going to be 80% it should be like 65 or 70% way higher than average, way higher than normal, but lower than where it is today. And say we're going to target that. And I think that would send a really clear signal to their competitors, right? Because a lot of what they say on their call is messaging to high-necks and Samsung, right? So they want to send that message that, hey, there's, we're going to fight on price but there's a floor to that, this is our floor. You guys good with that? I think that's one of the dynamics that would help. So I, but I don't know, I don't know that they can do that, but I would love to do that. And then they can say, all right, this is we, that's sort of proven what we've been saying is there's a new normal here. And we're going to reestablish our baseline gross margins, peak gross margins at the sum high number. And I'm sure they have lots of incentives to just make as much money as they can while they can. So I don't know if they're going to do that. But I absolutely, yeah, I just don't, I don't know. So here's what I think, here's what I think is interesting about my, about my ground. And again, the reason I stated it the way that I did is, you know, could we hypothetically think that what we're seeing is to some degree, the sports analogy of a team friendly deal, not team meeting the other two competitors, because I'm going to give you the context there in a bit. But that they recognize exactly what I think they got pressure for and people are freaking out about that. The, the capex for hyperscalers memory can't, just can't go to 60, 70%. Like the memory guys could absolutely make that happen. I think again, you have to reconcile with the reality that they have all of the pricing leverage, like literally all of it. And if they all can sort it on price, everybody pays it zero chance that that, so, so, okay, so we know that. I feel like to your, exactly to your point, like the reason they would do something like that, and the reason I think these floors and kind of, you know, ceilings are being talked about now, it is to one, help them plan, help them, you know, do, you know, fund the R&D, expand capex, all the things that at the end of the day equals more capacity, while at the same time, as you say, have a reasonable, you know, reasonable margins and still making money, but not price gouge. I guess is kind of where I'm like, 'cause again, they could all do this, 100%. And I think we've, everybody's been like, you know, cool, it's like we get it, shareholder value, but at the same time, if this thing crashes, it hurts everybody and there's just a limit to this, I guess, right? So anyway, so my point is like, is kind of, I guess what we're saying, micron could take kind of like a team friendly structure for their customers. And I say that because like, are they in an interesting position to do this, because they are the first of those three to have a lot of capacity come on in 2028. In fact, I have seen models from the street that have them doing almost as many wafers as SK hindex in 2028, 2029, which would completely shift the bounce of power, because right now, micron is around a 20% share of memory player. If they start doing, again, even double that, not more than double that, which is what SK is in that timeline, they all of a sudden have a lot of price leverage to even come down some, right, against those competitors and really hurt SK and Samsung. And so that's kind of where, you know, I'm thinking like, because they have this coming capacity, like how do you start to plan for that right now and in some of these when you know, sure, you're gonna have that capacity come on, you're gonna possibly more than double your way for output and probably sell all of it at the same time. That becomes a very different dynamic that microns never had before in leverage. - I think that's a huge dilemma for them because they're, like I said, there's a huge temptation to do that, but they have to be very careful because both high-nix and Samsung, or members of large-table, have infinitely more capital to bring to bear. And we'll use that and have done that historically many times. They are very sensitive to share. And if if my grant made an aggressive bid for share right now, I can't, I think that would, I think that would lead the war. I think that would be really, that would be really ugly. >> I can take it ugly, ugly, ugly. >> So, hold on. So what you're telling me is the memory industry is basically game of thrones. That's basically, they're the mafia, fighting blood battles for share and like, you stay off my turf, I'll kind of say off yours, but we're going to have an agreement like because it's the game of thrones of the semiconductor industry. >> Yeah, yeah, I mean, there are no dragons involved, but kind of wish there were. >> Yeah, it is, right, I mean, we've done, that's what these cycles are about, right? >> I know, I know, but historically, I think it was a fairly polite game. Now it's a very different world in terms of again, their control and leverage on this industry. So I just wonder, like, you know what I mean? Like sure, sure, they're going to get along and you're out there by the chair. But like, again, I'm not saying that they would like crater pricing, you just be like screw the both of you, but it's just this dynamic, I think, is a little bit more interesting when micron, all of a sudden, it starts to have a lot more capacity. >> That's what I was saying. I think what they should do is signal a floor, right? De-risk the gross margin concerns that investors are worried about, but also send a signal to the other two saying, this is how low we're going to, we're not going to, we're not going to crater pricing, we're going to take it to this level and the lower and then make use of their capacity and make all the money they can until the other bring capacity online just a little bit further out. >> Yeah, I think that's the smart move, but there is gonna be a huge temptation to pick up share. Right, and I don't know if, I think this is part of the dynamic too, was it last week? At some point, Samsung leaked a story about how much money that they were preparing to invest in memory capacity. And it was, I forget how much it was, it was hundreds of billions of dollars. It was just a huge amount of money. And like that wasn't an accidental leak. That wasn't like some engineer got some Excel spreadsheet he wasn't supposed to and leaked it to his friend in the press. That was like deliberate, deliberate leak from Samsung to say, oh, just a reminder, if we want to, we can, you know, we can swamp you with capacity. Yes. So yeah, this is a brutal game. And I think what's up ending things, like they've had a very polite stable equilibrium for a decade now, where they've had established communication pattern and you can sort of track it out. What they're saying, right? You can decode a lot of it. What's disabilizing now is one, there's just so much money on the table. Everyone's gonna scramble for that. And then you also have the dynamic of China, of CXMT. Sort of looming out there, potentially long-term, you know, fully destabilizing that equilibrium permanently. And so it's fun to read this. And it's absolutely, it's absolutely court intrigue, game of thrones level. Maconations. Memories game of thrones. Yeah. We're gonna have to work, we're gonna have to work those parallels in now in little nuggets of the analysis whenever each quarter when the memory guys come. Yeah. Who's trying to come on? Who's turf? Be careful, tread carefully, everyone. Yeah. I think it's, I mean, my big question right now is, prices keep going up. Who is the marginal buyer? Who's the incremental buyer here? It's coming in after memory prices have quadrupled. Saying, oh yeah, I want more. Like, who is that? That's one, that's like, I know there's been all kinds of like speculation on the street and all kinds of stories going back and forth about what's really going on and the memory tax. And like, but to me, I'm really curious, like, who is coming into the doors now trying to sign a new contract? Right, because that's me is what makes us all a little bit fragile is, are there new customers coming in? Is it just somebody who suddenly really, really desperate for memory and is buying it? Because if you think about it, the story we've been talking about most of this episode, everyone's been talking about a lot lately is the rise of inference and the need for CPUs to do a genetic inference. All right, those are not necessarily memory stories, right? Their CPUs are short. And yeah, you need some memory, but you don't need the HBM for most inference, right? You can use it, but you don't need it. And so why is it suddenly prices are still going up so sharply? That to me is, I'm very curious who that is. Because one theory is just like one buyer was late to the game and is now scrambling and is ratcheting up the price as a result. But it could just be everybody's just buying a lot and long term, like, when they signed those long term agreements over the year ago when Sam Altman went to Korea and signed a long term agreement and Jensen went to Korea and signed a long term agreement. That they just said, yeah, in 2026, we'll start paying double what we're paying now. So. - Yeah. And Anthropic did that deal with micron. You know, maybe it's a mix of both. Maybe it is a couple of these players who are outside your top five in scale and just have to, you know, pay the piper. I also have a theory that and I would say it's, there's some smoke here in supply chain checks, but I have not like absolutely 100% like confirmed that's across the board. That your big customers are your masters of scale, as we like to say, are securing that memory in advance for some of their customers in order to win some deals and maybe have some bundle pricing, especially those in your custom A6 area, right? If you can get that memory, then somebody is like, I need to make a whatever in a memory. Cool, I'll do it all for you and I got the memory that you need. And so I think there's some, whatever you want to call it, like hoarding in the, because you know you're going to need it, it's valuable, you're going to sit on it and you're going to be able to float that to help get customer wins or design wins. I think some of that is happening with some of your bigger, deep pocketed people. You know another word for that is double order. It's double ordering is what that is. Yeah, I mean, but again, I think it's, it's going to get used. It's just, hey, you need this? Come in the air, open the shirt. I got your memory inside of my coat. Now I'll still buy an A6. So, so let me ask you the really dangerous question and all this is in video, right? In videos locked up, all kinds of memory, they've signed all kinds of long-term agreements. Way ahead, everyone else. Yeah, they currently charge full price. They buy memory from, from Pynix and, right? Pass it through. They don't pass through. They mark it up full. They mark it up to their full 79 corporate cross margin, which is actually, which doubles the price of memory, right? Can in video keep doing that because no one else is doing that. A and B doesn't do that. Qualcomm can't do that. Qualcomm's not even using HBM, right? Qualcomm doesn't do that. Can in video keep charging that premium? Well, I would say, in some cases, yes, but it depends how long. I think they still have a ton of leverage. I also think, you know, they have the NeoClouds deeply committed to them in a range of ways. And they don't have a choice. And I think the hyperscalers, you know, this is what's interesting. You still hear the internal teams the big hyperscalers still wanting to use some GPUs. It's not like full, we're 100% on TPUs. Like some of the stuff they do, they want to use for GPUs. So you still have internal use there demand. And you also have third party use demand, right? Your wider market is still. In fact, I was having a conversation with an investor, a couple of investors actually around this where, you know, they were like, what happens to the CUDA mode? And I was like, you know, at the end of the day, like CUDA may not actually be like the real mode here. I think the real mode is their install base. The mode is the friends we made along the way. The friends we made along the way. That Andy installed base. Like I think that when the developers go, the friends we made along, and I think that's the hardest one to topple. Like you have such compounding value when you've got what it might be now, 30 million GPUs out there in the world. And the closest person's less than 10 maybe. And that's a custom asick. So I don't know. That's a hard one to get over. And I guess, you know, so it's a question of longevity. I absolutely think they can keep doing that for the next few cycles. Demand is too strong for their products. Your point is, taking, could that go on longer? And I don't know. But I don't see that happening any time soon. Yeah. I mean, I was just, I was thinking about that because a lot of work I've done this week is looking at the casualties of memory. And, right. And there will be many of which there are, right, we had Apple this week announced they're raising prices. I mean, that's the one that really gets me is like, Apple was like, Apple was surprised by this. Like Apple's been getting, getting asked about memory for a year now, right? Every single earnings call they've been getting asked about it. And the last call they said, no, no, we have long term agreements in place where we're set. And now they're raising prices, which means they were surprised by this, right? They're surprised by the increase in prices. And yeah, there's been all kinds of chatter about like a legit bad blood between Mike Ron and Apple. It doesn't, like, let's not even go down that path because it doesn't matter. Apple's pretty good at buying things. And the fact that this is taking them by that much surprise is, is tells me that, like, this is like, there's a lot of strange things going on. >> Well, I actually, I don't think they had as long terms of contracts as people thought. I think you're right that what they were surprised by was, you know, Apple historically would never really agree to a too long of a contract because their assumption would always be, they're going to get a better deal in six months. And for them, they would all be. always use their leverage and their scale to negotiate better terms. And if it's commodity, which they treat the memory players as a commodity, then to them, it's going to be, I'm just going to constantly get better pricing from you. That leverage went away. And so what they should have done was two years ago, just committed the crap out of memory, like whatever Jensen's doing, like just balance, you know, putting it on the balance sheet years in advance, just like I said, hoarding in the Apple, I just don't, I don't think they, I don't think they saw that coming. And by nature of their procurement, that's not how they do things because they're just constantly convinced they're going to get a better deal next time around. They can use that pressure. Neither is, paints them a good light, right? Neither, neither of the scenarios. That's, you know, I mean, you, you, you know, everybody, you know, in the supply chain world is always like, look, Apple's both hard and great. It's guaranteed money, it's guaranteed scale. No other but first it can do that, but they are harsh when it comes to negotiating. And you and I talked about this. I don't know. What was it a year ago when this started to happen when we were like, Samsung, who is their largest vendor? Like we'll come back at you at a vengeance. They will remember what you did to them. Yeah. However, for, for 20 something years, including when they fabbed with you and made a one through seven, two through seven, whatever it was of chips, they will remember and they'll be like, Hey, buddy, you either pay this. If not, I got to learn 10 people long. Who will? So, yeah. That's the, that's the times. It's the memory game of terms, the memory mafia, the mm, the memory mafia. I don't, all right. Anyway, let's just, let's close with a, let's close with a quote from the Godfather. Okay. It's not personal. It's business. There you go. It's just all business. Everybody. It's just business. I don't know if it's personal. All right. Thanks for listening. Everybody. We'll be back next week with more drama and the tracking of the mafia and other things that happen. And we will talk to you later. Leave travels all. Tell your friends, click like, review, leave a, leave a review. Tell your friends, subscribe, all that online stuff. Thank you. We appreciate you all. Thank you. All right.