EP 187: Inside the Future of Memory Summit & AMD’s "Boring" Bullishness

[MUSIC] >> Hello, everyone. Welcome to another episode of The Circuit. I am Ben Beharen. >> Greetings, agents. I'm Jay Goldberg. >> All right. This week was good. There was a lot of earnings. We are not going to get to them all because there was also a super interesting conference that I was at future of memory summit that we think is worth digging into. But we do want to talk about AMD. It was very interesting. There was kind of a mixed reaction despite the fact that they had some pretty good numbers. I know that there's some timing questions which we'll talk about. But obviously, and again, coming off advancing AI day where they outlined a pretty bullish forecast as well as their own belief that they will outgrow the market in accelerators. It seems to be the execution part of this story. I think is kind of what I landed on afterwards. But what was your take kind of high level? >> I thought this was very typical for what we've seen with AMD over the last few years, which is they have fundamentally, I don't say boring. They got mad at me when I said it last time, boring, but in a good way. They have a fundamentally very straightforward story, which is we're going to make some products and we're going to execute to our roadmap. And we're just going to keep doing that quarter after quarter. And that's what they did. And what a lot of, I think the street has come to expect is like blow out earnings, big massive beats. And that wasn't what they delivered. Like you said, they just had an event two weeks ago where they sort of gave all the good news. And this quarter came in, I mean, the numbers were good. Good quarter, good guidance in terms of relative expectations they beat across all sort of metrics. But people weren't super excited. I think the stock was down the day after and it's up a little bit for the week. But it's very much in the category of just an execution quarter. And I think there's a class of investors who just want more excitement than that and it wasn't exciting. It was good, which is like what AMD should be known for. And so I couldn't find, you know, I couldn't find fault with anything, any of their commentary. Some of the criticisms I saw afterwards struck me as a little esoteric, you know, data center, whatever number of people were expecting for data center numbers somehow wasn't good enough even though it was like, I think, doubled year on year. So it's a lot of little quibbling around it. But good numbers from a very solid company that's sort of just executing. That's the story. And I guess that wasn't exciting enough for people. Yeah. I mean, and I think part of it too, I don't know, I don't know, again, if this is just the boring part, but like, you know, so much strength in the business and CPU demand, even with ASP's rising, right as a part of that mix as kind of the new roadmap of CPUs, coming can command some higher pricing. But I don't know, maybe just people aren't as excited about CPUs despite, you know, the significant growth of it and their outlook into that very big number into 2027. And I mean, I get the accelerator base, but just the strength of CPUs alone, right, which I just don't think a lot of people have fully grasped in their models. Like they're pretty bullish on what's going on with CPUs going forward. And I don't know, something like I said, maybe that's just too boring. But I just thought it was interesting that it kind of feels like it's either not exciting enough or people just still aren't sure what's going on in the CPU demand. Yeah, I heard that a little bit. I sort of took that in the context of they're making as many chips as TSMC can, they're selling as many chips as these CPUs will make for them. Right. I think that's the key thing there with along with everybody else, their capacity constrained. I haven't run the numbers yet, but I think I think they lost a little bit of share in the data center to Intel, which we talked about a few weeks ago. Like it's just, the share is going to change this year, but the market is growing so fast that that's going to that will blur the share. I think share is a less relevant argument right now, just that the whole pie is growing so big. Yeah, so quickly. So I, you know, like, yeah, I mean, what was it three months ago, which seems like an eon ago, everybody's all excited about CPUs. And yeah, CPUs are still a big deal, but like maybe they're not as exciting as they work three months ago, even though their numbers keep going up for everybody. So I think it's a little bit of sentiment versus what people's interest is right now. Right. I thought it was interesting, you know, they pointed out first half CPU supply was tight because demand was unforecasted while 2027 should be better forecasted was the note I had from the earnings. So it's interesting to me that I mean, obviously, I, not a lot of people saw CPU coming and by the time we did, right, they would have secured waifers, but I kind of took that as a positive that they feel they can secure the waifers they need for what they're modeling into 27 with their forecasts, which again, I still think will be supply constrained, but the fact that it sounds like they're able to get those waifers, whether they can get the memory, that's a whole different situation, but the fact that they could at least get the waifers they need for CPU are not they're saying they forecasted that right. I thought that was a positive too because this again is an upside pool that I don't think people have really modeled into and they're saying they're going to be able to get, you know, that capacity that they need to grow into their demand forecast. Yeah. All right, I mean, AMD has probably the best access to TSMC allocation, anybody and they also are the most reliable and they're forecasting, they're not going to forecast numbers that they can't hit. So I mean, I think that's ultimately the problem is that they're very reliable and very reasonable in their sort of future commentary as opposed to some others in the industry who are perhaps forecasting things less tied to, to, you know, specific hard, hard data, right? And so the others look more exciting. AMD is like, hey, we're going to grow a reasonable amount and other people are saying, hey, we're going to grow a crazy amount. And AMD's like, yeah, but we have the capacity and we're not going to exaggerate our numbers. So I get that. I don't know how much people appreciate that when Lisa starts coming off as like, "Purty bullish signal for growth," like she doesn't do that. Like, she's pretty conservative. So it's sort of like they're like, oh, like you said, like, forget it's probably going to be bigger. But the fact that she's this empathetic on this particular growth is like you just have to appreciate that she does not do those things if there's not, that's a very bullish signal how ambitious she sounded in the growth. So take that for what it is. Man. All right. So there was a brief, I don't know if this was a big issue. It kind of got debated, so whether or what it was, but there was some early speculation that their capex spend, which is going up to, I believe, a 108 million and it was estimated at 200 something, was part of why the stock had a mixed reaction. Now, I don't know if that's totally true. I mean, I think the reality is if you need to go and secure that level of equipment, testing, et cetera, that you're doing your upspend in capex is the signal that you have a lot of demand. So I don't know if you had a take on that. It seemed like that was debatable. Like was the capex raised partially why some people were off or not, but I don't, I wasn't sure. Yeah. I, again, to me, that's like a second order criticism. I guess they're going to have to spend more, both in op-x and in capex to do this. It's not entirely clear where their capex is going. And I think there's, they gave reasonable-ish explanations for that. I think it's just cost of doing business at this point. It's getting very expensive to compete here. You need very lot more of your customers. Yeah. And that's going to apply some capex. I think partly what may have sort of triggered some people is all the financing stuff we've been seeing going around. That has, I know that that has spooked some people, like the degree of financing being required with what brought come and especially in video, putting into this. And I think there's a little bit of no trepidation among investors that AMD is going to have to start doing that as well. I mean, they, right? So they have to some degree, but not nearly as much as the other two. Right. And so when you hear capex going up, people start to get a little spooked. Yeah, I guess, again, second order quizzes, I'm not sure how important it is. I mean, there was a lot of second order criticisms that happened here. It was a good quarter. They're guiding to pretty strong growth. I guess the only other thing I heard was people don't disagree with the growth trajectory. I think they just hoped it would be sooner, like into this year versus more of a 26-27 story, but that's just what it is. That's what they're ramping to. They're ramping to meet the demand of the next two years. And so, again, critique that is what it will, but it's coming. They're posturing for very good growth is kind of, again, my takeaway. They have good growth, and they're building and ramping up for very good growth in the next two years. Yeah, I think it's as simple as that. I mean, ultimately, that seems in line with what they should be doing given the growth they're facing. The opportunity they have in front of them. Yeah. Okay. Well, yes, agree. I'm glad I summarized it. Anyway, good quarter. We'll go into Q4 where I think we'll start to see Helios start to inflect, which will be good. So good stuff there. All right. On to a future of memory summit, which I went to. So I don't know how many focus of our listeners go to that conference, but I have attended this conference at different points of time in the past 15 years or so. Obviously, memory coming up. I heard it was pretty big last year. I didn't go. I was out of town. Certainly, right? Anybody who knows the Santa Clara Convention Center, it is not the nicest convention center. It's kind of outdated. It's also not big. But I have never been able to park in the parking lot and the free parking lot that sits to the complex in between past the high it. First day I roll in. No place to park. No place to park at all. I was firstly like, okay, sign at the time. But then I'm talking to other people, executives, the friends of ours. Everybody was like, dude, I've never been able to park here. I had to park at Levi's. I had to park across the street. Everybody had the same point. This show was buzzing. It was crazy. The show floor was not big enough. Every booth was big and packed between SK and Samsung and Micron and Marvel. There's tons of like, I've never seen it. Like maybe there was an MMS where this was not. I've never seen this. And again, just make those observations to say the sign over the times of this show. Yeah. Well, I will say like this year at GTC, I couldn't get to the show. Like it was, there was no parking. The big lots were closed. It was clear it was going to take me an hour to find parking. So I just ended up not walking the show and just met people on the periphery. So, um, oh wait, that was in San Jose. Yeah. Santa Clara is. Dude, no, yeah. Yeah. Tiny. Like again, if you know Santa Clara Conventions Center, it's not big. So anyway, FMS needs to move. Apparently, there's still going to be their next year, but they need to move. It is. Anyway. Everybody in the ecosystem was there from memory to memory controllers and interconnects, which I think is an important, uh, interesting observation of why, you know, networking and interconnect companies wanted to show up. Um, but my takeaway was this was the first time where I kind of saw the ecosystem around CXL in particularly take shape. And everybody want to talk about CXL. There was a bunch of CXL announcements, both from chips, the fabric to networking. Uh, everybody in memory wanted to talk about CXL. Um, and the, the, I walked away believing that the time is now for CXL that will start to scale. I did a whole report on this at the Dillian Stack this week for, for anybody last week for anybody who cares. But, but you needed kind of the entirety of the ecosystem from the controllers to the sensors to the optical transceivers to the switches. Like everything to work so that you could disaggregate memory and now have memory B, what we all hoped it would be when CXL would scale the promise of CXL to have near memory compute, um, but disaggregate that memory off the board. Um, finally looks like it's out at a spot to, to scale. And again, everybody at the ecosystem was, was there. Um, Marvel had some excellent demos of this just showing disaggregated memory move at lower bit rate, sorry, higher bit rate, lower pico, jules of power, kind of all of the benefits across their stack. Um, and so, you know, again, I had not seen, I've seen pockets of this. You've seen some, uh, uh, proof points, but I think you walk in and you see a very mature ecosystem around the standard for true disaggregated memory. Um, it makes me confident that kind of we're on the, the path to, uh, to start to scale this. So, so let me, let me take a step back because CXL is a little obscure for some people I know. So I had to look this up. It's CXL stands for compute express link, which is, it doesn't tell you much. It is, it is a, it is a protocol for interconnecting different pieces of different chips, like chip to chip communication, right? And so think of it as, it's like a, like PCIe or USB for, but at a very, very low level connecting different parts of different, different chips or different tiles on a, in the package or whatever. And this is a technology has been around, this is a standard, I guess, that's been around for a while. I remember very, very clearly in like 2021, 2022, people were getting all excited about it. I was getting a lot of investor questions about it. I think we talked about it on the show really early on. Um, and it was clear that it was like one of these really interesting things, um, but it was going to take a long time to mature, especially because the applications that people can use it for originally were for big complicated, SOCs essentially. And, and then AI happened and everybody lost interest in CXL because all people cared about was memory bandwidth. And this seemed, this seemed unrelated. And I remember at, at, at some point, like in 2023, I think semi-analysis, I think it was Dylan wrote a, wrote a piece about CXL is, it's not dead, but it's, it's nobody, nobody's gonna care about it for a long time until it becomes useful for AI in the future. And I think we're, we're now reaching that point where it's like suddenly memory so, I mean, like, I think this is the broader theme that's going on now is like memory is so expensive and it's such a bottleneck that people are going to start doing everything they can to, to, you know, get around the memory tax, if you will. Yeah. So, CXL looks like, looks to be pretty interesting way to do that because it gives you a lot more flexibility in bringing memory close to the compute. Yeah. So there's two, there's two other parts too that I think make this interesting to inflect. I mean, obviously, right, as we talked about next week were, were, were early in the Rack Scale era. And obviously it benefits a, a medium or, or a protocol standard like CXL as we move to Racks because you're building these compute racks. It's not just like hosts of just CPU racks sitting there, but you're starting to build accelerator racks. And within that, you have optionality, right, as a, as a vendor, let's just say, as a hyperscaler at, at this point. And so you could see how as, as they start to think about Rack Scale infrastructure, they start to think about what they want to do in their, in their compute pods that gives them flexibility in TCO. So there's that part, we're early in the Rack Scale era, but Rack Scale becoming the de facto way to deploy accelerators and AI opens up this opportunity for CXL because you're just thinking about how you mix and match and, and build these things. The second one is that we were obviously waiting for inference, right? The reality is that memory makes a big deal in inference workloads. So you need a whole lot more memory in an inference cluster in order to handle concurrent users at scale, which could be many thousands of agents or hundreds of thousands of agents at any given time in an inference cluster. And so memory matters for that. So those are kind of the big ones. My concern coming into this was always like, this is great. This all sounds great, but like where are you going to get the memory? Like I don't, I don't like, you're not like memory doesn't grow on trees. You're not just all of a sudden going to be like, dude, I could now make a memory server with an unbelievable amount of memory with what memory. And so I'm like, okay, that was always sort of my thing going in. But what I came away with at this conference and kind of my biggest nugget, which I wrapped a whole report around for CXL was CXL is the only standard. And I say this because there are other standards, right? Credo is pushing Omni Connect, which I think is good. Obviously, this goes over PCIe and so people were like, well, there's trade-offs in speed, fine, you know, that can get solved. The reality is in every other solution, you have to go and secure brand new memory, DDR5, right what you want. CXL is the only solution that supports memory reuse. So reuse of old dims. So in this case, you can use DDR4. You can do MNEDDR4. You have in your data center as you retire. systems that have a truck ton of DDR4 largely in general-purpose CSP CPUs, reuse that in your CXL racks in your pod, either as its own rack, which I think happens later, but in within the rack in scale up. And so when you unpack the TCO advantage that CSL brings, CXL brings particularly to the hyperscalers who again have a unbelievable amount of DDR4 that they're sitting on that they can retire as they upgrade their CPUs to a Gen-Sek as we know they're all doing, which also incentivizes them to upgrade those because they can reuse that memory. Like it's just the TCO benefits for this make so much sense. Now again, I am isolating that to the benefits of those who have a lot of memory to reuse, which is your three hyperscalers and to some degree oracle, not others, but I also think those three/four alone can fundamentally drive the ecosystem of this across the board because again, they've actually got the memory in utils to refresh and use and use for CXN. So to me if you just said, hey look, why are you so confident that CXL is the thing? And I'm not saying other solutions do not enter the fray and we link obviously, you know, has a role to play here. What I'm saying is high memory reuse is the reason I think it makes so much sense for the hyperscalers to reuse this, which is why I think it is the one that gets the most traction first amongst others that I also think will be deployed. Okay, I think I mean CXL has always been interesting that it's very flexible, it has a lot of different things that can be used for. It was going to take time and I think now it's real, which is kind of cool. It's like a new whole new standard. And I think that's really interesting what you're saying that the hyperscalers have a whole little memory that they can reuse, which is mind-boggling to me. Like the idea that we're going to, you know, reuse memory. Like I know that's not exactly what it is, but I have this image of like data center, hyperscaler data centers with like trucks up on center blocks that they're pulling all the memory out. But okay, but there's two parts of this. So yes, the other reason I think this advantage is the hyperscalers is that like I've known this for some time. Like they do actually have as a part of their, whatever you want to say, like just the way they do TCO, they've already got qualified programs internally to reuse components as needed. So like they've got everything they need to take that memory, do what they need to qualify it and make it ready for new solutions. That's not easy. That is a hard process, but they do this oftentimes with old GPUs, with old networking performance. They have the ability to reuse that in other systems. Maybe again that they wouldn't have it in a tier one system like they can do with DDR4. But the reality is they've got everything they need to reclaim and reuse their old memory. So fine. Now again, right, I get DDR4, is not going to be as fast as DDR5. I'm sure there's trade-offs in speed and yes, it's not as low power. Again, none of that matters when the answer here is you don't need to go by a bunch of more memory. You have it. You have the memory you need to now go to Poyce DXLX scale, which again helps you keep your cogs down, helps you keep your capex down, helps you build new inference infrastructure that you can monetize with lower costs. Like it just makes so much sense. I think it's as simple as you can't get enough memory. Exactly. 100%. So you take whatever hit you can because the option is the alternative is nothing. It's fascinating. Yeah, no, agreed. And again, I'm not saying that I fully built this model and kind of projected it in memory reuse economics. But I'm not saying all of the memory can use. There is some that soldered to boards, but a good chunk of it is available to just unscrew or and re-qualify. So anyway, again, that's not an immediate thing. But that's why again, I said when you look at it in that light that they can reuse this old memory that's largely been sitting amongst their CPUs for general purpose CPU workloads. We already knew they all want to upgrade their CPUs. Like we see them upgrading CPUs are pointed about AMD seeing absurd growth in data center CPUs is only incentivizes them more to do it. So my other point is that you should actually be super bullish, hyper scale or CPUs CPUs with this line of thinking because they're incentivized to decommission those older CPUs, recommission new ones because they can reuse that memory and they get the benefit of bringing more agent to class CPUs into their data center. So it's like wins all around for the hyperscalers to do this and do this via CXL and right there's again, there's a handful of winners and I do for all of the like back and forth through the years we've gone on Morvel. I think they are very well positioned for CXL. In fact, I think they're the best position of anybody because they have the entire stack across the flexibility of everything you need to do this onboard of disaggregated memory within rack and scale up disaggregated memory and scale out in O2 rack or memory pods. They've got the whole solution versus others who have only a few. So anyway, I think I'm energized in that it's interesting. I see a ton of opportunity here, but again, going back to the beginning, the ecosystem seems mature. It feels like we're at a spot where the validity of CXL is now and we're going to start to see these things deploy at scale. So, you know, pick your winners here. Who do you think's best positioned because I do think this market's going to start to ramp? Cool. CXL. Yeah. So the other, sorry, no, go ahead. Good, good, finish this because then I'm going to. Yeah, I mean, I was like, when these old topics come back because it was so hyped up for so long and it was clearly like beyond what was reasonable. And now people have stopped forgotten about it and it's going to be actually useful. I like that dynamic. Yeah. Yeah. Okay. So the other one was I obviously had good meetings with everyone in memory. And it doesn't appear that memory costs are going down anytime soon, despite the kind of fear out there. I do think you could make you could argue that we are either at or near peak memory pricing, which to me is not a big deal, but it feels like if you just survey the landscape of Twitter, people are freaking out about this. Like, oh, it's not peaking. I'm using this term that you haven't heard me use yet that you probably would love called margin maxing. And I was like, I think the days of margin maxing are over again, for a good reason because it does sound like when I asked like what's going on with the conversations with your customers, their point was like they want flexible pricing. We understand that they need to have predictable pricing going forward, which is why there's price floors and price ceilings. We're not trying to make their life difficult. We're not trying to, you know, to some degree price gouge them. This word didn't come up, but you see what I mean. Which is why they're all doing longer term deals. And one of my takeaways, because I said like, you guys have done LTAs before, or micro and calls them S.E.A.s and SanDIS calls them NBMs, like you've you've done longer term contract deals before. And sure, the critique of that has been, oh, and LTA is great, but really they can just cancel them at will. Fine, I understand that critique. So the durability or strength of an LTA is certainly where I think people have concern today. But what they pointed out was, yes, we have done, you know, these longer term deals with customers before, but it's usually been with one or two customers and one of those two customers, or two of all of those customers that do those have a lot of scale. So Apple, for example, right, might do a longer term deal, but at a very, very low price. Their point was the two things that signal the duration of this demand is how many customers are signing long-term agreements, and how many of them are signing them at a drastically higher price and margin profile than years past. So it's sort of like those two things together, I think again, signal the duration and the durability as well as the revenue quality for these names, micron, sand, disc, high-nix, Samsung, etc. Just simply because this volume of customers would not do this, if they're supplying chain people, we're not telling them we don't see any ease to this, not even into 28 and 29, right? The number of those signing four to five year deals as average. Your supply chain professionals would be out of their job if they're like, "Nah, this is going to ease up in a couple years, like you're your host." So, anyway, that was my takeaway on where we going with memory demand. Are we at peak? The days of margin maxing are over, but focus on revenue quality and duration, which seems to be sustainable, is kind of my meta point. So we had one data point this week in sand, disc reported earnings, Dan. and monster numbers. And I thought the, a couple of things. One is the street kind of shrugged it off, just like huge numbers, huge beat. And the stock is barely up this week. One of the things that took away from their earnings was gross margins were held studies, 80, sort of low 80s, which in, you know, memory terms is incredible. Gross margins of 85% are, you know, just an incredible thing for the memory industry to see. And what you make of those though is, it's kind of, I think there's, there's lots of, there's a few takes I took on that. One is disappointment that it wasn't 90%. You see all this demand, wiring your gross margins even higher. I saw another take, which I thought was a little more reasoned, which was this is a sign that the memory companies are gonna be merciful, and that they're sort of being disciplined in this, and that if they really took prices as high as they theoretically could, that would create such badwill long term. - Yeah. - That it would accelerate all the other things we were just talking about before, before about ways that people are looking to get around the memory tax. And gross margins holding steady are a sign of reasonable discipline. And I think a lot of debate I've heard about memory pricing right now is how much higher can it go? And we, you know, we keep talking about, it's all sold out until forever. Our price is gonna go up from here. Are they gonna go up 50, 60%? Are they gonna double or triple? Are there some people out there saying it's memory prices could still triple? After they've lost track of how much memory prices have gone up, it's doubled and it doubled and it doubled again. So, and people are saying it can double or more from here. I don't write, and I don't know. I mean, all of those seem possible to me, and I think the other alternative is that pricing goes up a little bit, but I think the better debate of, you know, is it 60% or is it tripling? A better debate is how long can we sustain margins here and pricing here? - Yeah. - And I think that's the real question that people are looking at. And I certainly don't know the answer. The argument you raised, which is that there are some very serious long-term agreements in place across the numbers of customers and a very, very different nature substance of those contracts would lead some credence to think that we are in for a prolonged period of good memory, high memory prices and good structural, and I won't say structural, good margins for the industry. And if you want to get really bullish, then you say, oh, this is structural, and it's going to stay this way for a long time. Right? And I think that's the range of arguments there. I, you know. - Yeah, but again, like, okay. So I think you would be very heart-press to agree it's not structural. I agree with you that like, I mean, again, I don't think I could totally be wrong, but I don't see prices drastically increasing. I think again, the goal is to hold, hold steady, hold those margins. Now, when a new node comes on, right? When a node shrink comes and you get more capacity, maybe that leading edge your ASP goes up, but for the mature products we're in right now, I think the reality is like, let's just hope that they hold. Now, I think that will happen. I don't know if the market has fully, again, appreciating that's like Morgan Stanley had this note, you know, Sean Kim, who was kind of like, everybody blamed for the initial like memory demise. And now he's like, no, this is, we're returning now to a late stage. This is going to increase in Q4, but his point was if company, if people would accurately value LTAs the way we think they should, then that would drastically help the rating of these memory companies. And he's made the same point that I have been making for a long time, which is LTA signal quality of revenue and durability and stop focusing on margins. So again, if margins can hold, that's great. I think what people just need to route their reins around is that this is not going to crash in 28 or 29. Highly unlikely, but you still need to have, I guess, your assumptions of what percentage of share they can get in each capacity pool from HBM to, you know, NAND and DRAM and obviously storage. So this debate's not going away. Like I know what you're saying. I hear this all the time. It seems like the market just can't get their brain around this and that's fine. And maybe they won't. I don't know, but the reality is, I walked away from this show at those conversations with the memory guys, not really concerned about the structural nature of this. - Okay, yeah, I don't have a take here. I don't know where things are going to go. I know that I'm not a memory analyst and these things, there are a small number of people who really really know how to trade this and have deep, deep connections in Korea and Japan and Taiwan and Idaho and can really track this. And so I think it's very tricky to trade these. And I certainly am not going to offer trading advice. I just think from a company perspective, like that's the range of the argument is this is cyclical and it's going to end in a year when prices, when we, at some point when demand equilibrates or the other argument is no, no, no, there's something different here, we're going to have a prolonged period of heightened margins and if you want to get really crazy, you can say this is going to last indefinitely. We've restructured the industry, which, okay, I'm not like I said, I'm not going to have an opinion. Those are the range of, that's the range of the debate. - Yeah. - Yes. Okay, before we close, there was one other earnings I wanted to kind of briefly discuss, which we don't have to get in detail too, but we've talked about this company for a night. - Is it SpaceX? - It is not SpaceX. It is not SpaceX. I have yet to like put my mental energy toward SpaceX. I feel like if they keep wanting to talk about their ambitions to sell compute and be a hyperscaler and Elon's going to come out and say things he did this week about, we're dedicated 100% to Nvidia. I may have to pay more attention to them, but I'm avoiding this as much as I possibly can. But side time reported this week, who we also like to talk about. And honestly, the story just keeps getting better for them. And I just want to phrase it this way and everybody should do their new diligence. I have a wonderful report I'm working on side time that eventually will come out. But they will benefit from the Rack Scale era. Like you just look for companies whose just sheer content goes up in Rack Scale and as Rack Scale demands more compute and in this case timing and low frequency timing. So things that speed up that process and synchronization, they are going to be a big benefit. In fact, one of the things they said that that was interesting was they're seeing an uptick in demand even for their synchronization products, not just their timing products, which shows you that they're benefiting from scale up and scale out. So North, South and East, West, as a part of their MEMS based solution for both timing and synchronation at low frequency. So just to throw this out, I still this company is super interesting. They do, we've talked about this before, a very underappreciated part of the Rack, but their ASP profile is good as you move to lower frequency timing and synchronization devices. But my point is that you look for these companies who just Rack or content goes up as Rack Scale goes up and they are one of those beneficiaries. - Yeah, Scythe Time is deeply fascinating just in terms of what they do. It's like a fun technology. It is one of those frontiers of replacing old analog parts with something digital and sleek and new. That's kind of cool, MEMS, MEMS are hot. But yeah, the data center content for Scythe Time is astounding. I've lost track of how many hundreds of Scythe Time parts can start to show up in some of these racks now. There's just such a huge demand for it. And it's gonna continue to go up as networking challenges become more. If you look at some of the other things that are coming on stream, pick any field, like even memory, optical, all this stuff is gonna need more timing, more complex timing and something programmable and reliable the way that Scythe Time is a big deal. - Yep, nope, agree. I guess that the more I talk to him, the more I'm like, this is just such an interesting technology. I know it's not maybe as sexy as people like, but just very to any way. Lots of beneficiaries in the rack scale and the inference era, they among many and anyway. That's, it's Mims, how's it not sexy? I mean, it requires fafers. And TSMC loves this because it's a different packet of wafers that is driving interesting demand. So agree, but anyway, it's, check it out. All right, everybody, that's our show for this week. We appreciate your time and listenership and we will talk to you next week. - Thank you everybody. Tell your friends, click like, subscribe, get your agents to listen and summarize us, whatever it takes. - Whatever it takes. - Bye.