EP 185: AMD Advancing AI, Intel Earnings and WFE Demand

0:05Hello, everyone. Welcome to another episode of The Circuit. I am Ben Bajarin.

0:11Jay Goldberg: Greetings, programs. I'm Jay Goldberg. Well, we had another action-packed week across the industry. This week, with a couple of moments, we had AMD's Advancing AI Day that we were boots on the ground with. We had Intel Earnings, which then sent a whole slew of conversations in the industry. So we're going to cover all those topics and more and some of the little fingers that they extend into. But let's start with AMD. You know, we obviously, Keynote was good.

0:46Ben Bajarin: For us, they had a couple technical deep dives that I think are super interesting that I want to talk about. but let's talk let's talk broad strokes firstly I want to make the point and while we are all massive fans of Lisa Sue and have been forever I think this was her best keynote presentation wise like candor wise style wise like I don't know I thought it was really good like she's been great in the past but I think sometimes like I've been like get a little bit more excited Lisa like really like really feel like we're we're having a friendly chat you know and kind of like like and i thought this time she did a great job so so presentation style tone uh candor all that stuff i thought this was her best her best keynote i so just i want to set the record straight neither of us were actually wearing boots we were on the ground nobody nobody's wearing boots except for one

1:48Jay Goldberg: camera grip guy I saw, he was wearing boots and a skirt. So it was, it was, it was, it's 2026. It was, it was a good event. Actually, no, I'm lying. It was, it was a, it was a boring event, but in a good way. Right. It, I think I've said this before. The struggle with AMD is that for them, the best thing they can do is just execute really well and just do what they're supposed to do yeah and and that's what they did and it's it's not exciting or flashy i uh uh this i'll i'll grant you lisa is is getting very good at communicating in a more animated fashion um she was more comfortable on stage than i've seen her before absolutely she's she's very good uh Communication has never been, I think, her favorite thing to do.

2:46Jay Goldberg: But she's, you know, I think she looked at it like a task, like a skill. And she just honed that skill like she's honed every other skill imaginable. She's so capable. I noticed something about the keynote. And I was trying to think about this. They had, you know, as with every keynote, they have guests come on stage and talk. Right. And it was the co-founder of Anthropic and Sachin, the hardware lead from OpenAI. We're both on stage. And she had actual conversations with them. And they were kind of scripted. Like I think that they knew what the questions were going to be. But she was actually conversing with them.

3:28Yeah, right.

3:28Jay Goldberg: And it didn't feel canned the way that usually somebody comes up. Oh, let's have this, you know, dignity from some other company come up and talk. And they'll give a speech. And then the CEO will come back. And Lisa actually stood on stage with them and talked to these guys. And I think there was a little bit of off-the-cuff commentary she threw in that's hard to do. It's really, really hard to do. It's high stakes to do at a keynote. And she pulled it off pretty well. Yeah. So let's talk about the product and positioning.

3:58Ben Bajarin: There's two things that I caught that I think were interesting. And, I mean, you're right in the sense that a lot of people just kind of, I mean, they had already teased Helios. So Helios wasn't news. It was more technical details. What is this? We obviously believed they were going to do more in CPUs. That seemed logical. So I don't think there was anything people didn't expect. So it's sort of back to, like, what's the story? Where are you positioning this? How are you thinking about your differentiators? differentiators obviously the anthropic deal you know is interesting in fact i went back to um some of the commentary we had in like in november sorry in august of 2025 we actually talked about um amd being a good fit for anthropic where uh they could actually do some customization to clod on amd hardware which is exactly what's what's happening so um good on us for making a good guess, I guess.

4:57Ben Bajarin: But I think the question here that gets interesting, and there were some specific words that Lisa used, right? Undisputed leadership, right? In CPU, like powerful things, right? And I grant that, like, this is true, right? That is like, yes, what you've built and deployed does look very very good at the same time like the industry is largely absorbed by other players and so while amd is making good strides i keep coming back to like like how much share right is left like everybody's projecting growth nobody disagrees that they're growing and talisa's commentary and our qa outgrowing the market for accelerators in what is now being raised to a 1.4 trillion dollar TAM, which our model is a little bit lower than that.

5:56But, you know, so be it. TAMs are TAMs, guesses are guesses. They're going to be wrong anyway.

6:04Ben Bajarin: But yeah, it's just interesting the posture of how good the product is, which is 100 % true in a market that's just regardless very hard to get share because of nvidia nvidia's ramp supply chain custom asics like all of those things so um i don't just my my take in like superior products are great while at the same time this market is just structurally constrained

6:33Jay Goldberg: so so let me let me break that into a couple pieces because i think there's two things going on one is she made superlative claims about not just like we're the best not just about cpus but about gpus too and i thought that was that was telling like on on paper 455 mi455 she's claiming is better than whatever certainly than i think vera rubin i guess she's comparing it to which is of course a debatable claim but the the fact that she would make a strong statement is important because um amd and lisa in particular do not um tend to embellish let's let's just say exaggerate no embellish the way that other companies tend to get a little bit more exuberant in their language and i'm talking about a video i don't know why i'm playing coy right and so they're not going to make a claim that's completely made up or just like their guesses or their feel is this is better she's she has specific data that i i'll bet that you can she could point to to say hey we're better than vera rubin which is which is bold and then you you've then you move into the conversation of how big is the market the tam and they're they they say they're going to grow faster than the market, which means that they're going to gain share.

8:02Jay Goldberg: And that one's trickier because I don't think anyone has a real feel for exactly how big the market's going to be. It's also difficult to count right now because you have things that don't fit into market share boxes like TPUs. And so I get a little, one, I always get nervous about these TAM conversations just because I think the numbers, there's not a lot of rigor in some companies' methods.

8:33Jay Goldberg: And the other issue, too, is I feel like everybody's saying we're going to have 50 % share. All seven of the vendors are going to have 50 % share. And so I don't know what to make of share gain comments. Like I said a second ago, AMD has a lot of credibility on this front, but everybody says they're gaining share, which is mathematically impossible. That's my concern for the broader market. And I'll say you absolutely deserve credit. You said this a long time ago, a long time ago, AMD and Anthropic make a good fit. I'll tell you, leading up to the AMD event, I got a lot of questions from investors about who's going to be their announcement.

9:20Jay Goldberg: They're going to announce somebody. Who's it going to be? And so I drew up a list. I did this in a note a couple of weeks ago. I drew up a list of all the accelerator vendors and CPU vendors. And I made a matrix with all of the seven big customers. And it's a little bit of free love in the data center because everybody is working with everybody. and if you look at the chart the only people that haven't announced with amd is google anthropic and google anthropic and amazon and i think we all assume that amd has anthropic excuse me has amazon already they're just never going to press release it so they have they have something there google's not going to do it because they already have tpu and so that left anthropic is the only one left yeah right and now right and so every everybody loves everybody what does love even mean anymore

10:21yeah yeah agreed you know and one of the things that you know was interesting we we had some time

10:26Ben Bajarin: with some of the execs so i had a a chat with vomcy on this too around anthropic and like one of my questions was, you know, how, you know, and I think this is true in inference, right? And so I'm making a broad point, like for the market, like in the training era, when you thought about what infrastructure, what cloud you were going to use, I think they were very particular. It was NVIDIA. I want to train on NVIDIA, unless you're Google doing this in TPUs, but I'm going to train on NVIDIA and I'm consciously making that decision. In inference, however, it doesn't matter as much, What matters is the best tokens, the best margin.

11:05Ben Bajarin: And so I had said, how's Anthropic thinking about this when it comes to the broader deployment? They're going to have your infrastructure. They're going to have TPUs. They're going to have NVIDIA. And my suspicion was Anthropic routes that. Anthropic chooses when you do an API text for code or whatever. They choose where it goes. and they're going to choose the right infrastructure for the right workload that gets them the most margins for that task. And this is a API slash cloud point in general that's going to happen for inference. When I'm just coming saying, I just want the best tokens or they just want the cheapest tokens or they just want the best margins on their model via routing, it's transparent to the end customer, but they will route those accordingly, which again feeds the, if you've got great TCO, it's attractive, right?

12:00Ben Bajarin: They'll deploy it. You're going to give them better margins because your hardware runs it more effectively, more efficiently at better revenue or better tokens per watt, whatever that is, right? You, the end customer, will never see it, but that's helping their margins, their revenue profile for those runs. So I just think that's interesting, right? Again, and like we make a big deal that all this infrastructure is going there, but it will be very heterogeneous in an inference world. and all that matters is that that person who cares about the money, who has that capex, wants the best return profile on those tokens, and they will make hardware decisions at scale that they deploy based on that metric.

12:41Yeah, it's a very different math equation than training,

12:46Jay Goldberg: and it's going to be heterogeneous and mix and match, and that's opened up the field significantly. To the point even that other accelerator vendors are working with AMD. Because that's the other thing that got announced was they have a partnership with Cerebros. Yeah. I guess for CPUs, which makes sense in some regards. You've got to admit, things are getting kind of weird, right? Not only is everybody buying from everybody else, but everyone's working with everybody else, right? NVIDIA is working with Intel for CPUs. AMD and Cerebros are working with it. Like all these companies are competing on one side and cooperating on the other.

13:30Jay Goldberg: It's all kind of strange. It's all kind of strange.

13:35Ben Bajarin: Yeah. I mean, again, right, gaps in portfolio, which I think, again, is interesting, right? You are recognizing that you have a need. You maybe don't want to do it yourself, right? Again, right? NVIDIA bought Grok for this reason. The response to that is going to be AMD and Cerebris, right, for premium inference tokens, the latency inference tokens. Okay. But yeah, I mean, it's, again, like people filling the gaps. Like, you know, we're going to talk about Intel, but Intel is going to need to do the same thing. Like, what's their XLRS strategy? Are you home-growing this? Is it SambaNova? Is it you could also work with Cerberus, right?

14:09Ben Bajarin: I mean, they have options. It's just they need a complete profile. And that's where I think comes back to, like, we've talked about this before, and this is, again, just further evidence, and then I'm going to make a broader point. But we're moving to these compute clusters being rack scale integrated systems and systems that need to work with other systems in a unified fabric way. That's the goal. We're not there yet, but that's the goal is that it's tied together and functions as a unified fabric of compute. I like that word. I just invented it, but unified fabric of compute. and from that vantage point yeah like if you don't have it you got to work with somebody like that's just because these things are full systems that need to compete with other players right who might be more vertical or have more of those pieces so you got to fill those gaps right in in your in your system um but the the the broader point that i wanted to make was i did i do think it's interesting and i'm not sure you know more than just the observation what the implications are but i just remind everybody like this concept of rack scale being proliferated throughout data centers is still very new like there's not a ton of rack scale systems grace back will being the first one out there so like we're like early innings of rack scale so so then my question like this is the this is the hypothesis i'm throwing out okay in a few years you got a bunch of rack scale well then what do models look like then what does inference look like then like the profile of all of this the capability of the models the ability to serve into like all that changes because we're basically still running on hoppers you know like just just the world could look very

16:01Jay Goldberg: different in two or three years yeah i i think that's that's fair i mean i certainly have no idea how far the models can progress i assume a lot um but i know that from from the silicon side of things it's it's changing that the move to rack scale is a is a big deal and like i i was tagged in some somebody was taking pictures of the helios rack um and it's it's it's immense so just let me let me back up for people who don't know helios is the the rack scale design is the actual physical rack the server in which you put the the boards with cpus and gpus in it right so it's a helios rack as um and it is it's a monster it's huge um it's what it's like eight feet tall something like that it's really tall and it's very wide and it's immensely complicated and i know when when blackwell came out there was a there were a lot of problems with it right just manufacturing it putting the system together can like set standing it up inside the data center you need a crane to put the the back door on like there are all these sort of complex things that emerged after the fact that i think nvidia learned from that it sounds like vera rubin is very different um they they fixed a lot of the sort of manufacturing assembly pieces and it's to me an open question how amd is going to do this they bought they mean they bought zt systems zs systems for uh for exactly this this purpose is so that they could design these these racks um and i i asked them about this uh and they've and a lot of people asked them about this like how hard is this going to be to assemble and i got the sense that they are aware they're aware of this they've planned for it but i think some things you just have to learn through experience and so they'll get them up and running but even you know amd is closest to nvidia and even and both i would say both companies level ways to go into getting these up and running and so what do you do if you're somebody else in this field right intel for instance like intel's gonna have to eventually move to rack scale they're way behind this so market's changing but it's you know it's not evenly distributed lots of people are still further behind yeah so i there's two other things i wanted to to poke on so i i i do think that they actually have a lot of differentiation and strategic merit around the CPUs.

18:54Ben Bajarin: And I say that because, like, to be honest with you, this is where I think a rack scale, side-by-side competition with NVIDIA is actually very interesting. Because, you know, NVIDIA, right, people will argue, like, has the best GPUs, and the game who disagrees with that. But I don't think anybody would argue that AMD CPUs are better than NVIDIA's, except for NVIDIA. But regardless, there are a ton of just general purpose x86 Epics out there in the world, and a lot of the cloud runs on AMD. That's just the bottom line. That's not changing anytime soon. So what's interesting is, okay, so you take now this era where CPUs are much more important than anybody's realized.

19:46Ben Bajarin: Everybody acknowledges that, like we've all got CPU TAMs that are up and to the right. In that acknowledgement, I think that levels the playing field some. That's what I'm trying to say, is that I think for AMD, their CPU roadmap, their strength of CPU, In fact, we should also mention the paper that came out from NVIDIA earlier in the week, because I think there's some things worth saying CPU-wise, competitive architecture-wise. But I think that NVIDIA or AMD has a little bit of a level playing field there because of the heritage they have in CPUs. And because they've already got now around with Epic with the architectures, the ability to have a family variance, an agentic one, a cloud CSP one, a head node one.

20:32And where I think this is really interesting, again, we've sort of talked about this, but Ben's love for the DPU remains deep. Deep love for DPU.

20:44Ben Bajarin: And their DPU is actually super efficient and does a whole lot more than the things we've talked about, security, noisy neighbor. They actually gave a statistic in one of our technical breakouts that the way they built their DPU can actually drastically limit the amount of resources on the CPU head node that it could eliminate the need for, I think they gave it like 15 to 20 % less CPUs, head node CPUs, because of how good their DPU is. and like that's again that's just an efficiency that's an arcade that's a rack scale efficiency but my point is is like they're making these decisions because they control some of these architectures and because now they're rack scale with their mature cpus that's actually it is going to yield tco benefits exactly to the way that that lisa positioned it and so so anyway i think the competitive playing field in terms of specs perform tco at rack scale they've leveled the playing field and i just think that's that's an interesting bit because like this renaissance and cpus is helping them level that playing field is my broader point i'm getting to i okay i accept

22:03Jay Goldberg: that i think the one area though where i still have i still see amd as being somewhat i don't know vulnerable in all this is their approach towards networking um because because dpus or smart nicks are one piece of networking but it's and for those who don't know dpu or smart nick is is the interface card on the server that then attaches to the the network or sort of manages the networking traffic out of the server right but that's just one piece of it if you look at what broadcom and nvidia are both doing these you know amd's biggest competitors they both have very very deep very very deep networking stacks um and i think there is i don't know i could go i i think it's important i think they need to explore it more there is an argument to be made that they don't need networking and that they can be open to anybody's networking i know that matters a lot to the hyperscalers they don't want to be locked into nvidia networking because it's very expensive uh but i also wonder if they don't have to do a little bit more to get there um right i mean you have broadcom makes its own lasers that's how much they care about networking in nvidia just you know invested two billion dollars in lumentum and coherent and is spending money all over the place locking up all kinds of networking components does you know when everything is tight we're going to have you know we're going to have shortages and networking components pretty soon does that sort of is that a blocker for and these share ambitions when their two biggest competitors have this networking angle that they can they can provide yeah right because because what what they announced i think with i I think, I can't remember if it was Anthropic, I think it was Anthropic, they basically, or with Helios, it's basically going to be a Broadcom switch inside.

24:05Jay Goldberg: Right, or it's going to get a Broadcom switch. Correct. And so if you're Broadcom and you have limited capacity, do you allocate your networking components to something that's going to run on AMD Silicon, or do you sell it to the customer who's buying your custom ASICs? Yeah. um okay so that leads me to the sort of another point that i wanted to make but but but on on

24:28Ben Bajarin: your point uh fair i i also think it's very interesting like they walked us through the architecture for helios and one of the things that they pointed out was the way in which it works with the memory fabric is that the sort of like chiplet design on top of the memory fabric and that memory fabric is designed in a way so that it works very cleanly with all the other GPUs in the inference pod. So again, coming back to this one compute fabric point, which I thought was super interesting, which is why I think they're getting some of the memory advantages that they're getting for inference. But I also asked Forrest Nerod in our session, because I was curious about this, like why why is she wide why is it a double wide like clearly you you you've made a distinct decision that this form factor yields you benefits that the normal thin racks don't that typically are pretty standard and i went into this thinking like oh it's taking a compute per square footage uh hit and it was interesting because his comment was like you know when we talk to customers like we don't see the compute per square foot as a problem like concrete cheap um yes you know getting the power to it's fine but like they can make it bigger the the the problem from customers they heard was one serviceability and for whatever reason they heard that having a wider rack and you be able to you you you the operator being able to get your whole arms or shoulders in or around this thing was easier to service.

26:10Ben Bajarin: The other was, and they pointed this out a little bit in the keynote, but the way they did the cabling in that tight sort of cluster, I tweeted a picture of this from the show floor, but it actually reduced the latency, the distance for cables was shorter. And so there's some networking efficiencies that go into that, some latency and interconnect that go into those designs. And so I thought that was interesting too. And then there's also, this was interesting also i'm saying interesting way too much but um the top and bottom of the rack is is where the customization comes in for the customers so like the middle is you know networking compute trays and at the top there's some other things and so like their point was a lot of the work they've done with meta and some of the work they're doing with microsoft is that's where the customer is customizing those bits at the top that they really like now to go against nvidia it's interesting because i didn't fully appreciate this so this is on ben but when you look at one grace blackwell rack uh that is not all you need to run a grace blackwell rack you need a rack that sits right next to it and then a cooling rack that sits right next to that so it's actually so one grace blackwell system is actually pretty wide it's not it's not as thin as i thought when you just look at one grace blackwell right rack so so those as a thing take up size and amd actually did all that in fact you could probably argue that it's short it's smaller than all of what you need to run a grace blackware whack just in terms of width so that was what i learned i thought that was interesting that is that is interesting but i think you said it right

27:42Jay Goldberg: concrete is cheaper than silicon right and you know the other thing too is you think about it when we get to the fineman era in 28 and whatever that is mi 550 whatever comes in 2028 600 it'll be 600 it'll be 600 all right yeah so so fineman we're talking about a megawatt per rack right and so if you think about that you have a gigawatt a gigawatt scale data center it's only gonna have a thousand racks in it good lord right are you kidding me yeah it's it's just it's like that's a great stat i know they won't build them this way but i just picture this giant giant building Just like a few lonely clusters in the middle.

28:25Jay Goldberg: You know, I got to make up. The future is going to be just like one rack in the middle of a power plant. One little compute brain that sits there. I got to say, like this is, Amazon is building a 100 megawatt data center not too far from my house.

28:44Ben Bajarin: In fact, they're now building three. One of the land specs that they had for a fulfillment center over here in Gilroy is now actually also going to be a data center. So they're technically Building 3. Anyway, I drive by this thing on my way to Costco and I take pictures of it every week because we go to Costco every week, sometimes twice a week. And it's gigantic. And I'm actually – like how is the – I didn't think a 100-megawatt data center was so big. I don't know what I was thinking, to be honest with you. I'm just like, oh, I'm thinking of data. So if that's how big a 100-megawatt data center is, dude, I don't even want to know how large a gigawatt data center is to then again only have 1 ,000 racks.

29:30That's right. Oh, dude, fascinating stuff. Okay, well, let's move off AMD. Yeah, we don't deep enough. Yeah, lightly encouraged, and honestly, it seemed like the market,

29:44Ben Bajarin: believes their execution story, thinks they're well positioned, they're executing everything you said. So good stuff for AMD. Okay. Intel had earnings on the same day. Thank you. I really appreciate that. Making all of our lives more difficult. And everything was right except for one thing. Everything was right. and to our not shock, because this is what happens, the stock round tripped overnight and is actually down today. I have my opinion on that one thing, and then there's some other things I want to tease out on this, but you start with the highlights and then we'll unpack it.

30:35Jay Goldberg: So Intel reported, we're recording this on Friday, July 24th, Intel reported yesterday on the 23rd, and i i think they they crushed it uh revenue was a billion dollars ahead of estimates eps was i think double almost double what people were expecting guidance was similarly strong um gross margins got back into the 40s for the first time in three years operating margins were the highest they've been in four years five years they have traction for 18a they have customers they they confirm for real for real that they're going to pursue 14a which means they have customers for that too across the board everything sounded very positive but and and yeah to your point like the stock was up almost 15 in the after hours but here we are today less than 12 hours later and the stock is trading down three percent along with the rest of the market but but still it's pretty pretty shocking reversal yeah so i i had a hard time finding a flaw what is your one thing okay so you are right that there's two things you're you're you're right that every signal they sent it's exactly what

31:52Ben Bajarin: you said that there are customers for foundry that's wafer wafer customers so not packaging which we know not asic which we know and they they they highlight it on something i do want to pick up on because I think people are kind of underappreciating the ASIC story. I think they wanted them to not necessarily name a customer, but I think they wanted to officially say, we have a wafer scale customer, something to that degree for 14A or 18AP because they're signaling it. Every sign is there. CapEx go up. They said that's, but they didn't say we've landed our customer for 14a like do you know what i mean like they've it's clear it's coming there's no reason they don't but but they didn't say it and so that that this is going to happen it didn't happen this quarter but that to me was like the one thing i feel like you looked at all the all the questions in the qa afterwards and and a good portion of them were like really poking on like you wanted to hear them name, like not name a customer, but officially say, you know, we, we have a customer for 14 a and it's clear in the signals.

33:05Ben Bajarin: It just wasn't outright stated.

33:09Jay Goldberg: All right. Um, I, I guess that's it. I mean, I certainly, I saw, I, I heard the same things. people were asking that trying to get them to name a customer in the after call callback they did with a bunch of financial analysts it was the same thing everyone was trying to get them to name a customer they were very clearly not going to say it um and they have a lot of good reasons for that right tsmc doesn't tsmc doesn't really name its customers 100 right tsmc names its customers only when like the customer has said that and everybody known it and everybody's known for years do they finally use the customer's name so i'm not surprised intel doesn't want to do this customers and i i think the the key sensitivity here is that whoever those customers are they don't want tsmc to hear about it absolutely and so long as nobody says it officially it's fine for tsmc i guess i don't know it's a mess i think there's another dynamic at play here too which is a lot of the people who i think this is true on the south side and i'm gonna be careful here i want to be polite but like there are a lot of people who have there are a lot of semis analysts who have been doing this a long time and have been around intel for a long time and so for much of their time covering the space intel was sort of the core of their coverage universe yeah i agree and they know intel really well they know intel far better than i do and they're they're they're like okay we're out of survival mode and they're reverting back to questions of from the old intel yeah correct back on intel and and so they're trying to get into all kinds of nitty-gritty of of all the different dynamics and play in a in a foundry or in a fab and i i think i think they're sort of i think there's a lot of overthinking taking place here um and and also you know not for nothing in fairness like the stock has run incredibly strongly this year yeah agree on not a lot of like real substance and i thought this quarter was interesting because they actually showed that the product stuff the product stuff long before we get to foundry being real material revenue the the products are doing really well and so it's it's confusing what's going on with intel and i think people are overthinking a little bit but I don't know the whole thing confuses me because yeah now that you say that actually I I can see that in that they're kind of like reverting to mental mode Intel was dominant

35:48Ben Bajarin: Intel went and spent 30 20 15 billion dollars on a foundry why are you only spending two right extra right I actually I can see that but I think my takeaway from this is that this management team which i i'm actually okay with but i understand the disconnect is really just going to kind of give you the bare minimum of information like they're just kind of then i'm not saying that they're sandbagging it but i think they just want to be a little bit more conservative like that's the tone that's the posture that they're taking and you go it is you know that that's just going to be it but but i think you're right because my other take was i felt like it was like why aren't you going to spend more like this opportunity is yours go get your bags you know like take it take it and and and that's not what they're getting from management like we're gonna go spend blah blah blah like other companies so i so i sense that all of that to say intel has moved into a structural growth period.

36:58Everything is there for it. You could argue how much. I'm not going to bicker or try to pick that up. But this is a growth period. And this is a growth period where for them to come out and for them to basically signal, margins will

37:10Ben Bajarin: keep improving. It's not slowing down. It's going to continue to get better. Literally means that they have solved so many of these functional problems and they're scaling at wafers with good yields. So it's like every sign

37:24Jay Goldberg: is there right every sign is there right there's one honest and i won't name him but he asked a question on the call and he asked a similar question in the follow-up which is why aren't you spending more and i was just like i was pulling my hair out i was like my brother in silicon this company was going bankrupt it was on the verge of like you know real serious problems less than a year ago right and you all collectively on wall street hated the last CEO because he was spending too much. And now you're complaining that these guys aren't spending enough? Like, nine months ago. It wasn't, like, years ago.

38:00Jay Goldberg: It's, like... So, maybe let's all just take a breath.

38:06Ben Bajarin: But that emphasizes your point, which I think is an excellent observation. That they, like, this is the hearkening to the intel of old, and this is not the intel of old. You really do need to change your mental model. And, again, that's fine but you're totally right like you're you're coming off people who still even though it was 15 years away i don't know ago where intel was just the undisputed dominant champion killed everybody like that's just not this intel anymore and that's right you know so anyway i i i think all of that is interesting like like i said there's nothing was wrong like they did nothing wrong everything they're doing everything right um i mean again i'm not that i expected the stock to go back to 200 all of a sudden um but i was just surprised that it was like man like you guys really just want a whole lot a whole lot more that's pretty much an unreasonable amount of more you uh you this said in investor community so um anyway yeah that was that was sort of the the take was there anything are we missing anything well i think there's one

39:18Jay Goldberg: There's one last important element of their call yesterday, which is that they're increasing their CapEx forecast. Again, not as much as people would like, but by a fair amount. And they didn't guide CapEx for next year, but it's sort of implicit in everything that they're saying that CapEx will go up next year and it'll be a big number. And that's very important for semi-cap companies. Yeah. On that point, then let's talk semi-cap. That's definitely called transition in podcasting. That was a transition. That was pretty neat, huh? It's a great transition if we could be radio hosts.

39:52Ben Bajarin: I think the CapEx number is larger than most people think as well for next year. Which when people see it, they're going to be like, well, you're gearing up for a whole lot more than yourself, buddy. Anyway, it should all be clear. All right, let's talk about something. So, yes, so I didn't look today, but I remember we were talking yesterday. So AMAT and I know Amcor was up on NVIDIA's news, but that's also relative to Intel as well because they benefit from silicon as an OSAT.

40:33But I think the wafer equipment market is going up a whole lot more in dollars than most people have forecasted and modeled for this. And I say that because I don't know if a square footage analysis has been done in the factories that will be coming online between Intel and TSMC and even Samsung before 2030, which they would all be placing orders

41:03Ben Bajarin: for when they're ready. But that's a lot of space that's going to come online. It's going to need fresh WFE.

41:16Ben Bajarin: but I think it is I mean I've seen numbers that are like 300 I've tried to build some forecasts on this to the degree that forecasts are just fun educated guesses that's bigger than that but you get the sense that wafer equipment is going to be constrained everyone they'll sell and they're going to start charging more which means that those ASPs just continue to rise and WFE goes continually up and to the right for some sustained long period of time.

41:52Jay Goldberg: I, yeah. So, of course, all the semi-cap names are down today too, which, okay. Yeah, there's every fab in existence. everybody who owns a fab is expanding capacity that means and like we were saying before the cost of the fab is really not the concrete it's all the equipment that goes in it and you know maybe you could argue that the the semi-cap names semi-cap companies don't have enough capacity to meet all this demand but it's tough to see a world in which semi-cap numbers don't continue to go up this year. I mean, now they've already run a lot. And so I'm not going to make a, I don't know, valuation is going to look like, but certainly in terms of fundamentals, it's going to be a good year for semi-cap and it looks very likely to continue into next year.

42:52Jay Goldberg: And then, you know, if you want to argue that, then we're arguing like, what is the memory outlook for 2028 look like? And who knows, right? But there's a lot of capacity going to get built in the next two years or 18 months.

43:05Ben Bajarin: Yep. Absolutely. And that's why I think like, again, right. You had, like we talked about last week with ASML, they're like, yeah, we think we're, we're, we're going to make 85 machines, but everybody was like, you got to make more than those. Like they're being conservative, but nobody actually knows. Like you just don't know how much more they can go and output. Right. And we talked about the size of these things is, is so big. It's not like this is, you know, just go outside and plant a vegetable and get tomatoes and, you know, in 60, in 90 days. And that's just across the board, right? With lamb, with K-LAC, with applied materials.

43:41Ben Bajarin: This just takes time to ramp. So there is that lag. But I think, you know, again, right, the difference is, and I was doing this analysis actually just on the back of Google's earnings. things um you know capex always has historically lagged this i mean this just makes perfect sense right from account lagged the revenue that you see from that anywhere from like 18 months to two years sometimes three years and so you see that inflection in revenue after at which the time that the capex is is is spent so that's why it's kind of like you'll see so much of this happen with wafer equipment like it's going to compound very very quickly because those orders are being placed, they're not recognizing that revenue yet, then they will add a uptick that I think is faster than people are getting.

44:29Ben Bajarin: And even just if Intel goes and does north of 20 billion, right, in CapEx next year, which I think it is probably a good chunk, a bit more than that. Okay, like that's 20 billion. I don't know if I said 20 billion. That's a lot more way for equipment, you know, That's a good chunk adding to the base and greenfield capacity foundry space that they've got. So anyway, brought all that to say, I still think we're under calling the size of the market for WFE by 2030. I think it's bigger. But every sign, again, just points up to the right at a faster uptick than we were a year ago or even two years ago.

45:11Jay Goldberg: Yeah, and I think Intel is an important part of this because if you think about it, There's really five major customers at the leading edge, right? The three memory makers, TSMC, and Intel, who are the biggest customers for semi-cap equipment at the leading edge, which are the most profitable products, right? These are the five most important customers. And you had the three memory makers adding capacity like crazy. You have TSMC adding some capacity, but not enough. And then you had Intel as the laggard who had cut CapEx. And now they've signaled that they're increasing CapEx, which means you have that fifth, the fifth leg.

45:58Jay Goldberg: I mean, it's pretty meaningful. You had, you had, you only have like five major customers. The fifth one is now back in business. And I think that that matters a lot when you look at sort of how the spend is concentrated among those five. so and i i don't i don't have a i i don't like i'm not gonna pretend to know what the what the cycle looks like and how it plays out i understand people are concerned it's cyclical semi-cap has is is very tricky to cover because it it will when it when it turns it will turn very ugly very quickly but it's just hard to see that coming anytime soon and which is why i said Like, are we really debating what's going to happen in 2028?

46:39Jay Goldberg: You know, that seems very far away right now. And there are plenty of signs that it could go up more than, you know, maybe it goes down, but also equally, maybe it goes up to all the points you just laid out. Yeah. Oh, and then there's TerraFab. There's TerraFab. We didn't even talk about TerraFab. Like, I mean, it's hard to talk about TerraFab because nobody knows anything about it. But, like, it certainly all signs indicate that it's real. absolutely they have the potential to be the sixth customer and all of that um absolutely right i'm just gonna say it's it's happening and i think they are i think they are trying to fire on all cylinders yeah and and i thought it was like i know that i don't think it's on the call but afterwards they everyone tried to get more information about terra fab out of intel and they weren't saying anything other than to say it was real.

47:36Jay Goldberg: I think he actually said, Lipu and Elon speak every day. Right? And so... BFFs. Maybe they hot tub. Yeah, I was going to say, I don't think that's a social call. Like, two extremely different people. We hanging out today? That's right. Going to get some... Is the cognac excellent, Jay? Did you watch the game last night? Reference to the... Yeah. So I so so TerraFab's real. I thought it was the other thing that actually didn't come up much on the call with Intel was the was it was TerraFab was one. And the other one was this something's going on with their facility in Ohio. They had sold like there was a rumor that Hynex was going to buy the Ohio Fab.

48:24Jay Goldberg: Right. Which got disputed, which got the shot down because it just didn't make sense. But what could still happen is that there is some kind of joint venture between Intel and Hynix, which would be super interesting because Intel is not in the memory business anymore. And so if they actually had some play there around Intel Foundry, that could be super interesting because there's a memory shortage. Yeah. Well, then let's leave this on a cliffhanger, tying these two pieces together. this came up on the call about memory and Lipboo basically just said yeah we've got a nice heritage in memory and left it at that so there you go thanks for the cliffhanger buddy we can all now speculate JV-ish stuff but yeah they're going to bring back Optane would that be awesome?

49:14I'm Sites 2020 for Intel alright last one so Texas Instruments You know, we all love talking about analog for all the things we've done on this podcast forever. But sort of similarly, stock was down. Expectations were really high. Maybe analog's still a little lumpy. Like not like digestion period cyclical.

49:43Ben Bajarin: Like there's parts that you just can't get, right, in power and microprocessors, controllers, etc. etc but ti did what it did market wise so yeah so ti reported their quarter very good quarter from ti

50:01Jay Goldberg: uh very solid beat decent raise to guidance and again this is a company that is very conservative and how they message to the street and the numbers were good uh the the keys things on the call were that auto industrial all their sort of end markets were inflecting in the in the good way they've all sort of bottoming out and are turning up. So we're very early in the cycle. And then there's their AI business almost doubled again this quarter. It was all good news on the call. There was no clear hook for anyone bearish to sort of hang their hat on. I think the stock sold off the next day, just in part because it had run so much, people were sort of taking profits.

50:46Jay Goldberg: I don't think there's a strong I didn't hear any stronger criticism than that it's down again today but everything is down today because of you know the world but I think for TI the story sounded pretty good and remember what they're seeing now is just sort of like good industrial and in the data center it's just sort of good attach rate for lots of power stuff in the data center we haven't even gotten close to 800 volt that's all next year at the earliest and so i i was a little worried that like their ai their their data center stuff would start to sort of taper down a little bit um that wasn't the case it's still growing uh i think maybe it still gets a little bit lumpy next year but at some point 800 volt kicks in and becomes more meaningful so analog yeah analog is cool yeah and this is this is it's interesting too because i like i like your point on 800 v because we did talk about that uh a while back on on one of the models I built for power semis, but like it is highly dependent on that timing.

51:48Ben Bajarin: That timing appears to be, at least people are waking up to the fact that it's going to take a little bit of a longer transition to get to 800 volt. Um, it's not going to be instantaneous, but, but when that ramp happens, it's, it's going to be a pretty strong pull for, for power semis and analog kind of across the board. And that shift should be relatively sustainable for some time. because that's a very big changeover infrastructure-wise, even on board. Like, you know, I'm a fan of doing these things. I like to go and somebody would have laughed if they saw me, but, like, there was nobody staring at the Helios racks yesterday.

52:27Ben Bajarin: They were behind plexiglass. But I was, like, kneeling on the floor with my face against the plexiglass trying to see all of the chips that were there, including the power semis and some of the voltage regulators. the size of the voltage regulators continue to get bigger, which I think is fun. But regardless, just looking at, again, how much more power content is coming and these rack scale solutions is significant. But that has to ramp. That has to come with a ramp. And I'm not sure we're seeing this true inflection ramp yet in analog. I think we will, but we're not seeing yet. And I think people were hoping that was happening sooner than later.

53:08Jay Goldberg: i i guess i don't i don't think that's applies to ti's quarter because again they doubled their data center business uh i just i looked at that helios rack and i saw i think it was six no it was more than that it was 12 u so 12 slots of just 12 slots reserved just for power semiconductors not the upc not all the the sort of bulky electronics passives just the voltage regulation and power management chips in that server. I mean, it's not quite a quarter of the height of the server is just power semiconductors, which would suggest to me that there are going to need a lot of power semis in the AI data center of the future.

53:58So I don't know. On all of that, we agree. Yeah. Up and to the right. That's right. Oh, man. All right, everybody. Thanks for listening. We appreciate your time as always. And we will talk to you next week with more and more hyperscaler earnings, CapEx raises, return on invested capital, all those things we'll discuss.

54:23Jay Goldberg: Plus a few more. Thank you for listening, everybody. Tell your friends. Click likes.