EP 139: Intel Earnings, Anthropic TPUs, Challenges for AWS

(soft music) - Hello everyone, welcome to another episode of "The Circuit." I am Ben Beharan. - Greetings, programs. I am Jay Goldberg. - All right, first up on the topic list, Intel had earnings this week and it was good. I think everybody kind of walked away feeling, you know, we're not backsliding. This isn't, you know, the situation we had was about a year ago where it felt like everything was great and then the wheels came off the truck. So we're sort of back to what feels like stability. Calm before a storm, in my opinion, they are in this transition phase of you've got a ton of demand for CPU chips on client and in commercial and they don't have, they're not shipping their new products yet. So they have a ton of demand for old products whose margins aren't great because these old products are made at TSMC and not Intel, some mix of 10 and seven. But obviously they are, they are as a company trying to move everything toward 18A. And so you're in this like time period where customers need chips, years aren't ramping yet. So you're selling N minus one, maybe some N minus two. And that obviously, right, impacts some margins. But it feels like again, things are trending in the right direction and let's start from a product side and then we can talk foundry. So give me your run through on top line but then also commentary that came out around product. - It sounded really good, right? It was one of the happy, happy things I felt about the call, the product sounded good. Their products sounded good for client this quarter and they seemed to indicate that data center would be good next quarter, right? And I think that's just them regaining their footing and they have some new products and more coming down the pike. So good for them, right? I think one thing, you and I can debate this but I was a little confused today 'cause earnings were good, stock was up. But so was AMDs. And to me, I was confused 'cause yeah, the market's growing and it's bigger I guess than they're saying it's gonna be bigger than people thought for client. But doesn't, I mean, aren't they taking share back from AMD or is the market just growing so they can both win? - Well, my take is that commentary is a little bit more on the server side because there's so many data points now basically confirming, I think we talked about this even a couple of episodes ago, that the demand for just CPUs in cloud native stuff, you know, just boring workloads like we talk about is up. And I would also say like a part of this that people aren't really factoring in is when you use, when we historically looked at these models for the infrastructure that Amazon has, Google has, Microsoft, et cetera, for general CSP workloads, they're also using some of this capacity expansion to host some of that infrastructure as well. It's not just going to be AI accelerators. There's some CPU attached that's green field growth, right? So if you're again, if you're building new places to house infrastructure, you're bringing critical workloads. And I've actually had some of these conversations with some of the hyperscalers. What they haven't completely figured out yet in their minds is what workloads they wanna host in direct owned data centers and what do they wanna host in some of their co-location facilities. And to some degree, I feel like there's a tension because you could see them say, well, we're going to put some of our lower value workloads at some of these places where we can negotiate lower GPU rental deals than it costs us, right? In our infrastructure. And we're gonna road the AI workloads because that's just better for our economics. And more importantly, we want them close to home, right? 'Cause that's our baby going forward. And so, I feel like there's this tension around, sure, we need to replace boring cloud infrastructure, but where are we gonna put that? Are we going to put that in greenfield growth or are we just gonna rip out and replace where we have capacity when we do need to retrofit and spend money on some of those older data centers to host AI workloads and do liquid cooling. And I think they're working out where they want to put that stuff. So there's partially, like I said, there's a refresh and then there's to some degree a build out. And I feel there's CPU momentum happening there in some capacity and it's again, it's happening around x86 to a degree. So I feel like the point there is, well, sure, AMD should benefit also because they're doing really well in data center CPUs for kind of cloud native. So they should benefit from the demand that Intel seen outside of client, right? So server point specifically I'm making. So you think there's meaningful CPU server growth, not just AI head nodes CPU growth. Yes. And I say that for a couple of things. I'll approach this from a technical standpoint and then everybody can take this for what it is. I think in prior to the inference era right there was an assumption that we're just going to continue to decrease the numbers of CPUs per accelerators. And I have now heard over the course of what would view road maps for the next two plus years that that may not be the case. That you may actually still a meaningful number of bigger, more powerful CPUs to act as orchestration, to act as offload for some of the compute functions that a GPU is not going to do. And to handle X number of concurrent people per GPU that the CPU plays a really big role on that number. And I've even seen people try to break this down to say, let's say we've got a rack of, NVLink 72 for example, in this case, right? This is arm parts, but we could assimilate this to X86 as well. You actually know roughly what the number of concurrence you can hold per GPU per rack of humans inferencing at one time, generating and taking tokens. And I think there's architectural evidence that the CPU, the compute capabilities of that CPU meaning for you help that manage of concurrence per user on inference. And you may see, you're not going to see one to one, but my point is it's not going to be one to four, which people thought if it's two to four, that's a good constant and that's been very similar to what you've seen in accelerators before. But I think there's green field expansion because it is going to take a ton of CPUs to work with the agents coming from the GPU cluster racks, like perhaps more than people think. And I think we don't know the answer to that, which is why I've, in my own brain, I've been trying to just model like, is there a world where the CPU TAM actually grows from what it was, right? 25, 30 billion a year to more. And is that units or ASP, but either way, I think you could argue it's still a critical part of that workload and may even be more important than people thought in the inference era for rack GPUs. - I mean, we've talked about this a lot before, right? About the role of CPUs. There's going to be demand for CPUs in an AI future. - So it's, I think it's worth watching 'cause I'm not quite there yet. I agree that the ratio of CPUs to GPUs, it's not eight to, it's not eight GPUs to one CPU. It's, and it's probably better than four to one. - That's what I think too, yeah. - It's somewhere between two and four to one. Averaged out. I'm just not sure yet that we are going to see additional non-AI attached CPU growth. There'll be sort of standard refresh, but I don't think, I don't know. I mean, it's a big question. It's important for all these companies. - Yes, agree. Agree, and I just don't, go ahead. - And on top of that, on top of that, there's also the ARM versus X86 question. - Yeah, so yeah. I think where I'm sort of in the, in the, in the camp of is like, we still aren't seeing inference at the kind of scale we're going to. And I just wonder like, you know, if you look today at a general purpose, right? Cloud native workflow rack that's got two to four CPUs in it, right, depending on your config, that's to run the software today that does not have a hundred million people using agents to shop on Amazon at the same time. And I just wonder if like, does, like, does that workload, which is a CPU workload, like drastically increase in its needs to serve inference at scale for e-commerce, for, you know, general web browsing, for people's SaaS apps. And I think, we don't know, and my point is, we don't know the answer, but I think you could make an argument that that's going to be very computationally intense. And therefore you might need more, more than you do have today to run enterprise workloads that are, or commercial workloads that are engaging with hundreds of millions of people using agents at scale. So that's my only point is I don't think we know, but it feels like that's going to be a technological, you know, burst of demand. - Yeah, I, it's encouraging to hear, 'cause I've been thinking about this this week in a separate, 'cause you know, I haven't caught up on this. So I've been thinking about this separately, and I was starting to lose hope. - But I also think, I think there's so many big unanswered questions around inference. We all know there's going to be a big spike in inference, but I don't think anyone knows when it's coming or what exactly it'll look like. It makes sense that there would be CPUs in there, because a lot of the tasks are going to be involved, don't involve AI prompts and model queries. It's just basic stuff, you know, refreshing web page and pulling things together. But, yeah, and I keep going back to an example that we talked about, right? Where we, folks on myself and folks on my team, have tried to use these browser agents. In fact, I tried to do one with ChatGPT, whose newest browser Atlas the other day, where I tried to make it go do something on some research forms that we're doing, and it broke the website, and took 30 minutes. And so I'm like, all right, this is not where it needs to be, you know, you can't go work for 30 minutes and challenge, complete the job, and then have the website be insanely slow, right? Just because you're pinging it and trying to fill out a form, and I don't even know how many tokens that takes, right? So it's a wild part of a constraint. It's a constraint, you see it's a pain point. It's slow, it's not accurate. It's completely slowing down websites, you know, when this happens at scale. And I just think that's where, you know, again, if Jensen's right, and if everybody, you talk to an infrastructure, sort of like, look, what we had, what we have largely in our data centers is for the old world, is for traditional humans browsing the web. What we do not have is infrastructure for agents upon agents upon agents to be talking to each other, and clogging up usage time, and spin cycles to load a webpage, right? So I, you know, we don't know. I just, I think it's an interesting thesis to build out, the opportunity around just general purpose, compute infrastructure, trying to make sure that it can work well with AI infrastructure. And that's why I just, I think a case can be made that we'll see CPU TAM growth, which is good, right? That's good for everyone. It's good for the ARM folks. It's good for Intel, it's good for AMD, makes it more competitive. And there's green, there's, you know, some TAM expansion for CPUs that you could argue. - Okay, let's come back to the topic of agents in a little bit, we gotta get through Intel. - So, all right, so the other part, like, okay, let me just briefly talk to clients. So yes, we're seeing the client's strength as everybody who follows the drama that is always an end of life out of a Windows product we are at. Windows 10, end of life. And we have models, I've shared this publicly, so, you know, feel free to recycle my tweet if you need to be. But somewhere between 200 and 250 million enterprise PCs need to be just taken offline and repurchased or refreshed for Windows 11. So that's a lot. And that's not gonna happen in one year. Enterprises are slow. You know, sometimes you just have to pull things from a factory's workers, cold, dead hands. But this is a pretty, still a meaningful number of units that need to be refreshed. In fact, Intel gave an interesting note that they saw the client TAM number growing next year to 290 million units, which is more than all the bean corners. Like, IDC and Gartner's numbers are not that high, to be honest with you. I need to, you know, yes, that includes Mac, so that's not X Mac. So that's not a Windows number, that's just PC volume. And that's actually super positive, because that shows you, you've got back-to-back years of actual sort of growth in sales of PCs. But again, a lot of this is driven by this enterprise refresh that's happening specifically around Windows 11. So that's positive. And to be honest with you, I think anybody who tracks Intel, tracks this space knows, Intel remains very entrenched in the enterprise market. That is, that has been a challenge for AMD. They do great in share for consumer. They're gaining in gaming. But enterprise share is just so stuck to Intel. And if this commercial cycle is strong, like I said, to 26, that's gonna benefit Intel. So that's my optimism of a strong setup for the client group through next year on those structural dynamics, which again, all of it favors Intel. And again, Panther Lake looks like a very good product with very good SKUs that the OEMs are happy with. So they're going to have more Intel in their portfolio to sell to commercial. Commercial is gonna be happy. And we've got, like I said, hundreds of millions of units that need to be refreshed over the next couple of years. And I do think Intel will benefit the most from that cycle. - Okay, yeah, no argument there. - So all right, let's go foundry. - Yeah, let's go foundry. - You start, 'cause I'm gonna try to be more optimistic. I know you're not super pessimistic, but-- - Oh, no, no, no, no, no, no, no. So that to me, there were two highlights of the call for me. And top of the list was Lip Boo's comments about foundry. Remember 90 days ago, he said, we're gonna abandon 14A if we can't get an external customer. And he walked back, he didn't walk them back. He said, we are engaged with customers. 14A looks very promising. This is the first time I've ever heard him say anything good about 14A, right? - So I found that super encouraging. Whatever he needs to justify the investment in manufacturing in 14A, he's seeing that. And I found that super encouraging. So there you go. - Agreed. - I would also say-- - You doubted me. You doubted me. - Two things on sort of the back of that. So one was also, they made a really interesting point. Okay, all right, let me approach it this way. If, and I'm gonna focus on the commentary, not the analyst questions. If we didn't have the moment you talked about last quarter where he basically was like, we don't know if we're gonna go forward with 14A. If we didn't have that moment, you wouldn't know they weren't going forward with 14A from just this calls commentary in terms of how they discussed that they're ahead in the process of development, that they're early engaged with customers. Like you wouldn't know that this sounds like it's a possible no go, right? Which I thought was actually very, very interesting. Now that either means they're very encouraged by the customer engagements, which is what Lipu said. Or they're just, they know something we don't and they can't talk about that yet. So I just, again, you're right. The tone correction on 14A I thought was really, really solid. But even the tone around 18A, right? That they're ramping, that yields are improving on industry standards levels, which is between two and 5% per month. There was a comment about wafers for margins, right? In terms of yields increasing and where they need to be throughout the rest of the year, which I think a lot of people picked up on, although there's some nuance to that, given that these are Intel wafers, not TSMC wafers. But yeah, I mean, foundry sounds positive. They're also discussing, which you tell me how you interpreted this. What is the design business that they're talking about? Like they've mentioned this like, yeah, we're standing up a, you know, a design, a custom design side of this business. Like is this X86 is open for business and custom design? Do they, are they saying we want to go after Broadcom and Marvell for custom designs for TPUs? Like there was no additional commentary on that. - I think they are, I think there's a specific thing that they're working on. I think they're working on a CPU for one of the hyperscalers that's more than ARM based. It's actually designed by ARM. - Okay. - And what is more than ARM based being? - Meaning ARM, like everybody's wondering what's ARM going to start, yeah, it's ARM. Like it's ARM design chip. It's not just an ARM IP chip. It's this ARM, everybody knows ARM is designing a chip. They're moving up the stack from just IP licensing. They're designing a chip. I think it's from one of the hyperscalers and I think it's a CPU and it's not crazy to think that it is going to get fabded, could possibly get fabded Intel. 'Cause, you know, let's remember SoftBank owns ARM and now owns what, 2% of Intel? So, and maybe it's something crazy. Like it's SoftBank is the hyperscaler, right? Why shouldn't, why shouldn't SoftBank be a hyperscaler too? So, or it could be, it could be Facebook or it could be both. - So, okay, but, all right, but you don't, you didn't take from any of that that there's like, they're trying to be Broadcom, right? In terms of like, 'cause they mentioned like we've got IP and I was just like, nobody's explained in detail what this design solution service is, whatever it is. So, that's why I heard people say like, oh, they want to be an ASIC designer. I was like, I think they want to manufacture them. I don't know if they want to be the back-end designer that's to take on Broadcom, but there's confusion around it. So, you take that, you don't take that as, design services like Broadcom bring up that, but more. - I think the term ASIC designer is very vague and has permeable boundaries. And so, I think Lipoo can talk about them having an ASIC team, an ASIC design team that still fits in my framework, right? Where it's not quite competing with Broadcom. It's some subset of what Broadcom does. It's ASIC services. - Okay. - All right, so, yeah. - Okay. - That's my full-blank conspiracy theory. - Okay, well, anyway, net net on foundry, it sounds good. You know, one of the things that interests me, and you know, this has just been trying to be optimistic, is I am intrigued by this idea that, you know, Lipoo can go to Jensen. He can go to Google. He can go to all of these places and basically say, like, how could we make 14A better for you? Like, what if we just did whatever you wanted and you had the perfect foundry to make your stuff? 'Cause TSMC is not gonna do that, right? You go to TSMC, you get what they get. They're flexible, but they're not like, actually building out a new node and they're gonna take all this customer input on what can we do, right? And I also found it interesting. I don't know if they intentionally were using this language, but essentially trying to like weave in that this is a dedicated couple of foundries. So I would say 18A, but particularly 14A, two AI products. I guess all they're gonna make, 'cause obviously at some point, every chip's an AI chip, but they're custom building this. Like, it is a specific design for these big, burly, you know, chip-lit designs across the board, right? So that's why it's interesting that a customer could be, yeah, you know, actually, let's give you some input. We could use this and then they do it, which is just to make it more attractive to, you know, I would never say it's a semi-custom designed foundry for them, but you see what I'm saying. Like there is, they're taking that input for customers to build a dedicated kind of purpose-built AI foundry for these chips that they wanna make these customers' products out. So it seems like if they could have their cake and eat it too, right? They're in a position to influence what Intel does favorably for their designs. And if that's the case, like it just becomes, again, maybe more of an attractive option to make your chips there. - I think what you're describing is just what a hungry, upstart foundry would do, right? And it makes sense, it's like, we're new to this business, we wanna be successful in foundry, we'll do whatever it takes. And I said before, there are two things that really encourage me on this call. The second one was all the sort of interpretations of comments that Lipu was making along these lines. There were several of them where he was, where there's all sorts of sort of things where like, you could not imagine Intel five years ago saying any of these things, right? And the important thing here is that Intel's biggest challenge all along has been its culture. And these are indications to me that culture is changing, which is a big deal. And the fact that they're, I mean, you used to go to the, as little as nine months ago, if you wanted to do a test chip at Intel Foundry, they had to design the chip, right? It's a trivial exercise that most companies would just sign the intern to. Intel had to have had a team of like 20 or 30 test engineers who would design the chip, the test chip, right? Which is a throwaway, you're gonna make 20 of those, right? And by contrast, Lipu is now saying, oh yeah, you want whatever process, whatever thing you want, we'll do that. That's the smart commercial response that a newcomer to a market should be making. So I found that encouraging. - I think that's a good point. Yeah, I'd agree with that. It was another little snippet on the call, which as much as I like to track every facet of supply shortages that there is, it's impossible. So I'm embarrassed that I missed this, but Dave Sistner just casually threw out like, oh yeah, and there's a substrate shortage. It's like, great, so there's also a substrate shortage. Like, what's not a shortage at this point? It's like a small list of the things that we don't have shortage and capacity constraints on, man. - So, but this is where I start to part ways with the call yesterday. 'Cause I thought there was a lot of encouraging things, things that matter more, like 14A and the culture change, that's a big deal, it's huge. But there are other parts of the call that I didn't like as much. And one of them came around this topic of their gross margin guidance for next quarter and go into next year is very, it looks very weak. It's down from this quarter, which wasn't great to begin with. And they got to ask this question numerous times and their answers were kind of all over the place. And so I agree, I heard that like, oh, substrates are in shortage. That's super interesting, I didn't know that. But I also don't entirely, how do I put this? I don't have 100% confidence in the answers that they gave. I think there's lots of factors at play. And in particular, what I think is going on is they're not yielding where they wanna be. And we, like even with Intel, whatever process they're on, three, three and seven, whatever they're on. - Currently on, yes. Yeah, three, seven and 10 actually. - Three, seven and 10 is probably okay, but three and seven, I don't think they're yielding well. And then we get to 18a, where they explicitly said, 18a yields will not be at industry acceptable norms until 2027, right? Late 26, probably into 27 is what they said. And like that's 18 months away. Again, going back to the Intel of five years ago, you don't have 18 month yield learning. Like it has to be faster than that. It was, that was very eye-opening to me, right? To say nothing of the fact that there's been a lot of press reports that 18a yields have been poor and the company has consistently denied them. This was like, why is this taking so long? It seems very odd to me. And I don't know what the cause is. I mean, I think part of it is the fact that they've lost a lot of people in that organization. And partly that it's all these EUV tools that are kind of new to them. But that's kind of left me a little bit uneasy about what's going on on the fabs. - Yeah, I saw, so the margin stuff, I agree, right? They discussed that in terms of one that Altera was gross margin accretive. And so without Altera, you'd expect that to hurt. Okay. And then also right upstart costs in their Intel product is cleaning the pipes for 18a, right? And working to ramp. The positive thing they said in that was we've got enough yields are good enough to meet current demand. And current demand can't be high because these products are not shipping until early next year, at which point, they're also high end, right? Nobody launches a $600 laptop to start. And that's the scale of the industry, right? They launched $1,500 and $1,800 laptops to start, maybe $1,200. So they're gonna be higher end machines where the TAM for 1,200 plus is 60 million or less total. And Apple has a ton of that. So yeah, I agree. I'm with you on the margin in terms of industry standards. I think typically like TSMC's 12 months to 90% is where they've been. So I'd be curious if anybody clarifies, is Intel saying that's 70% or 90% in that timeline? 'Cause 70% is okay. It's not great, right? But you wanna be at 90% within 12 to 14 months, certainly not eight, 24. So let's hope that they ramp this up quicker and they figure some stuff out and it becomes better. But yes, that is the big way on margins will be their yields. 'Cause to me, when they say they're over supply constrained, I immediately went to, they're not yielding enough, right? You get into the binning and like, especially 'cause like you said, we're talking about high end parts to start with, those are the ones that you have the least of in an immature process. And so there's- - Well, but I think that to be honest with you, I think when they say they're supply constrained, like some of that is Lunar Lake and Sapphire Rapids and Granite, which are on TSMC. So maybe they just underword. - They didn't get enough. - They just didn't get enough. - And they got to commit wafers to, mix and match where they put wafers, right? Like I said, it's an unfortunate transition for them. The timelines did not work out normally. And so it's gonna be our favorite term lumpy for six months. You know, when they try to handle this transition to get customers what they need. But yeah, but I know for a fact, a good portion of that demand is from products that were made at TSMC. So take that as you will. Okay, all right. Well, anyway, it was good. Nobody's rewriting the stock. It's not a shocker. But we'll see more progress from that. Steady, it's consistent as Austin on my team said, we need, it's good for Intel to hit singles. Consistent number of singles per quarter helps. So they've hit some singles. Good job, in the right direction. All right, let's talk other compute. TPUs are in the news lately as Anthropic has officially booked, call it a million, they were very specific. A million TPUs. A million TPUs. There's a lot woven into this I wanna talk about because we have sort of said this before. This is a, I think you can tell me if you disagree, but I think this is a indictment of Amazon's infrastructure that they're doing this deal. Oh yeah. And do you think, 'cause I heard people speculate this, that this deal is what Broadcom was talking about as the fourth $10 billion customer? 'Cause I've seen four folks in your camp of sales side suggest this. And I don't, like why would you be discreet about this though? Why would you be, oh, we get this other customer? Like if it's Google, then it's a third one of yours for, not a new one. So let me take a step back and give some context here. If you think back a year ago, well, not even like, yeah, a year ago, or back to when Google started doing TPUs 10 years ago, there had always been somebody who's saying, oh, they're gonna sell this. Google's gonna go into the business of selling chips. They have their own chip. They're gonna compete with Intel, or they're gonna compete with Nvidia. And I think that is, I don't think that was ever gonna happen. Yeah. Google does not have the capabilities internally to sell chips, and I really doubt they want to. I don't think they want to, absolutely. Especially when they have a cloud computing business where you can just sort of rent the chips, at least them. So that's interesting. It is also a strong testament to the power of TPUs, right? And you gotta imagine that Anthropic could have cut a deal with Nvidia somehow, right? And they chose TPU. And I think that sort of speaks to, TPUs are really, really compelling. I keep noticing this. We have all these deals going around. Neoclouds, and hyperscalers, and everyone's scrambling for electricity. Google's not in that mix. Like Google's bought some power and building some big data centers, but they don't seem quite as, they're not hustling as hard as everyone else because they have TPUs, and I think it gives them some meaningful advantage. I mean, they have enough capacity that they can-- Agree. They can allocate a million TPUs to somebody else, who's theoretically a competitor. Not really, but sort of. I think they're a competitor. I mean, Anthropic is interesting in that they actually have a business model, and it's around their API. It's an enterprise API, and it's doing well, right? They're not, they're not trying to, I don't think they're trying to do search. I'm pretty sure they just fired their ad team. They are not trying to do search. No, they know. They're not trying to do search. They just fired their ad team, so they're not doing ads. Yeah. They are, I would say, becoming very, very enterprise-focused, for a good reason. Maybe just accepting their fate in the same way that Cohere is, and they're just an enterprise LLM, great. But there is some overlap, right? Google wants to do coding too. Anthropic's pretty much the gold standard for coding. I assume Google wants some Gemini APIs to be handled by businesses, but to be honest with you, that's a two-horse race for Anthropic and OpenAI, in my opinion. But you're right. Competitor-ish, but 100% what you said was, yeah, no, go ahead, go ahead. Yeah, Anthropic seems to be doing well, and they could have gone to Nvidia and they chose Google. Yeah, exactly. I did think it was interesting though, to your point, right? Which is like, I kept thinking, and this is where in my brain, I'm trying to work these models, most of Google, and we have this, their number of TPU installed base, somewhere between five and six million TPUs, like my assumption was that most of those, Google's running all their stuff on it, right? So like, I would have assumed they are at almost max capacity of their installed base, which is why it sounds like they're adding, they're upping their TPU order next year because of this deal. And that's where this speculation around, okay, 'cause that could line up, a million TPUs could line up with $10 billion, right, in terms of Broadcom. That's where I think the speculation came from, was that that's an increase over what they were going to do because they are adding capacity, one million chips, I think also a gigawatt, give or take of what was secured, that is additive to Google's TPU stuff, right? Just for Anthropic, as an up to build to next year. So that's kind of why I think people were speculating of that, but again, right? That's not a mystery for customers, it's Google. They're just barbed by, they're buying more for Anthropic. - Yeah, so here's, I think this could square, I think this is one of the things that's tricky is, it sounds like this is just a straightforward cloud licensing deal. Google's gonna buy more TPUs and put them in a facility dedicated to Anthropic or something like that. Which, yeah, just that deal alone would not be Broadcom's fourth customer. There is a possibility, though, that, well, let me take this to the back. When Google came out of TPU, Google's not gonna sell it. They don't have the capacity to do it. Broadcom does have the capacity to sell some of the conductors. They know a thing or two about that. Why don't they sell TPUs? And the problem there is that is IP, right? A lot of the design of the TPU comes from Google and those contracts between your ASIC provider and your hyperscaler are like the craziest and thorniest, gnarliest contracts out there. So Broadcom can't just go out and sell TPUs to anybody, but they could do it if they got a license from Google. And so, in theory, this could be a situation in which Anthropic does the deal through Broadcom. Broadcom counts that as a customer, as revenue. I see. And they share the revenue somehow with they pay a license to Google. I don't, that's not what this deal sounds like, but I could see that being... I mean, there may be some advantages to doing it that way. Yeah, that's interesting. Okay, all right, we'll table that part. All right, briefly though, on Amazon. Here's what I think everybody's gotta start to grapple with here. I mean, it certainly tells you that Tranium was not competitive. And per, again, models I've been building on who has the most AI compute infrastructure, both in ASICs and in GPUs, Amazon has the fewest GPUs of any of the hyperscaders. So, which again, if I'm repeating myself and I've said this on this podcast, I apologize, but it makes a hundred percent sense that they doubled down on Graviton and not Tranium because to them, that's what they saw, was just cloud workloads. We see CPU-based workloads growing. We need more of that. We need to do our own thing. Graviton was to them with what they had line of sight of workload to as to what TPUs are to Google who had line of sight to AI for themselves. So, I'm not faulting them. What this is though, is I think Amazon is now at a crossroads. Either Tranium 3 crushes it and can compete and can hunt both in cost, performance per watt per dollar with TPUs and to some degree GPUs or Amazon needs to just, they're just gonna be a cloud CSP. Like there's not, like they've, they're losing ground in cloud workloads and Anthropic choosing Google to basically double down on Google's compute and infrastructure is so telling, especially when Amazon has so much tied to Anthropics upside. Like a lot of their economic opportunity was on using APIs through Anthropics to run on Amazon instances. And if that's that risk, then I don't know. You have to reconcile that maybe they're not gonna be a big player in AI infrastructure. - I mean, Amazon owns five or 10% I think of Anthropic. And you have to assume investment deal like that includes some board oversight and the ability to weigh in on whether or not Anthropic is gonna use a competitor's cloud service. So absolutely, it is a very stark picture for Amazon. I think, I think Tranium, I see, I had been hearing good things about Tranium, whatever the next one is, three, four. - TR three, Tranium three, yeah. It's a new cluster, yeah. - The new ones. I heard good things and Tranium four sounded promising and they keep investing in it. I don't think they'd do that if they weren't getting decent performance. But this is such a clear like black eye for Amazon with a company that they own a chunk of going to a competitor. It really does open the question of what their plan is. And I also wanna be clear, like this is not entirely, this is not like Amazon dropped the ball. This is, you know, they are, I actually just wrote about this last week is, Amazon is not able to get the allocation they want from NVIDIA. And usually that's portrayed as like, you'll see this portrayed as like, oh, Amazon's asleep or Amazon just doesn't wanna accept NVIDIA's architecture with NVLink, like sort of technical architecture argument. But I think it's more than that. I think NVIDIA and Jensen in particular really want to blunt the power of Amazon, the hyperscalers in general and Amazon in particular. Because, you know, if you're Jensen, would you rather sell to three hyperscale customers or a hundred neocloud customers? You have much better leverage, much better pricing, much better margin, the more fragmented the market is. And so I think NVIDIA is deliberately underallocating to Amazon, which puts Amazon in a tough spot, right? So again, I'm trying to defend Amazon, like I don't think they're dumb or asleep. - Correct, agreed. But this is sort of where, again, I think this is just an interesting dynamic of who's got the GPUs and who doesn't, because let's, again, let's just carry this out. And let's just say that Tranium 3 is not gonna be it. And they've gotta have a real, you know, stare in the face moment, what are we gonna do? Then your answer is, well, then we're gonna go cut deals with neoclouds who have GPUs and secure up instances, because that's what Microsoft's even doing, right? That's what others are doing, that's what CoreWeave's doing. You know, they're expanding their fleet of access to GPUs. So if you can't get them, somebody else has them, generally those people would be willing to do these deals. The problem is, far as I can tell, everybody's GPUs are spoken for. Like we don't really have at this point, even as Grace Blackwell rolls out on the screen, it's like they're all booked. Like you've either got it or you don't at this point. And so it's not like Amazon can just go and be like, oh, hey, you got GPUs sitting around. Let me come to you, like even Lambda, right? Others, like I get that hoppers are cheaper, but regardless, right, GPUs are not sitting idle. And in fact, and you can make a case that GPU is going back to A100s and beyond that are five, six years old or whatnot are still not sitting idle. They're not the most expensive, but they're being used. So that's my conundrum for Amazon, right? The only way you get extra capacity is you've built it with your Pritchit, but if it's not competitive and you can't get customers to come to it, that's a problem. And you got to figure out how to solve your GPU deficit. And I don't see where that comes from unless Jensen just is like, yeah, sure, we'll help you out and give you more. But he'd probably say, you got a step, make a trade-in. - Well, no, no, my point is that he just won't. He just won't, right? - But even if they came and said, we're gonna stop this and we're just gonna go with you. - Yeah, I don't think he would. - Yeah, all right. - Right, I mean, and because play this forward a few years, Jensen, like the risk here is that the Neoclouds, there are too many of them, they all collapse and we just revert to the status quo a few years ago where it's three customers, right? And Jensen's not gonna let that happen, right? He has sufficient strategic foresight. There will be its big temptation 'cause Amazon is probably willing to throw all kinds of money at them. I'll bet you he's not gonna take it because it's just too complicated and it diminishes his strategic power long-term. He'd much rather have all those Neoclouds on. - Yeah, okay, all right, okay, so again, dude, then if you're right, this just, again, complicates this problem as to what happens if you just can't compete on AI infrastructure and all of those people running to run their AI workloads go away from you. And I had this discussion with somebody on Twitter, they were like, well, so what if that's their own fate? Is that they're just a cloud native CSP? I'm like, okay, I mean, it's not the end of the world. They've got a lot of good businesses. They're just not partaking in what could be a very large monetized structure of AI software and maybe that's okay. - Yeah, I mean, this was my thesis of my piece was that people, myself included, usually point to the hyperscalers, to the sentient six and say, you're throwing all this capex at AI but you don't have a business model for it. And it's true, they don't have an application for it but I think the hyperscalers, especially the CSPs are, they do have a business model. It's called the CSP business. And they see that threatened by the neoclouds. Yeah, and right, and 'cause we've already had, we've had Oracle come from nowhere to legitimate contender for third place. If you look at their own forecasts, they're gonna be second place in five years. You have the neoclouds all sort of nipping away, gaining share. So the market has shifted, it hasn't shifted a lot yet. It's still sort of low single digits move away from Azure and AWS. But it's still, it is clearly moving strongly in that direction. So in that light, the hyperscalers are throwing this money at capex to protect the business they already have. Because I think they're real scale advantages. And I think this is part of Amazon AWS's sort of core thinking about it is, they're the leader because they're the biggest. And because they're the biggest, they get the best pricing, and they can offer the best pricing to customers, and it becomes that flywheel. And if they start to lose meaningful share, then that flywheel breaks down, and it becomes much harder for them to compete. Man, what a turn of events though, dude, from the ramp that's been AWS. Well, I knew I'll be at re-invent in December. My expectation is Tranium 3 is coming. So I'll have learned more about that. But dude, I mean, it's just really a lot rides on this, and/or they pivot. So yeah. Yeah, and to be clear, I'm not saying this is all gonna play out this way. I'm just saying these are the stakes. Yeah, I agree. Tranium 3 and 4 can be very powerful. Amazon, lots of smart people there, they can figure it out, but those are the stakes. Sure, I agree. All right, last topic, agents. Their founder of Anthropic was talking to, Dorakash on his podcast, and nuggets of agents were dropped. So let's talk about agents. So, Andre Carpathi, the one of the co-founders of OpenAI, and the founder and CEO of Anthropic, was on a two hour and 45 minute episode of the Dorakash podcast. We're at 40 minutes now. So two hours, this plus two hours. And he, all right, so to be clear, this is somebody who doesn't just know about AI, he doesn't know a little bit about AI. He has defined a lot of it. And he said a lot of things, he talked for a long time, and one of the topics he covered very early on was, he said, "This is gonna be the decade of agents." And Dorakash asked him, "What does that mean?" And he said, "Oh, it's gonna take a decade for agents to work." And he, which is not a good thing, by the way, right? There's lots of companies out there that have just raised money on the idea that we're gonna have agentic AI agents doing things for us independently. And he said they don't work, right? And I think most people I talk to, myself included, don't are aware of this. Like agents are a big topic, everyone wants to see them work, and they're just not there yet. Now, in his mind, he was being very optimistic. He said, "These are very powerful things, they will get there." We just need four or five, what did he say? Four or five significant enhancements to LLMs and transformers to get there. And he didn't know what they were. He didn't, just because they haven't been invented yet, but we need sort of several big improvements to the fundamentals of AI today to get agents to work. And he said similarly non-flattering things about coding agents. He said, "Coding is good as autocomplete, but you can't..." He said, "Vibecoding is only useful for a very small subset of things." And again, this is a company that would like to, that has really good, probably the best coding co-pilots. Again, in his mind, he was being very upbeat, but to those of us on a sort of shorter time horizon, it wasn't great. And AI is gonna take a while, that's my take with this. Yeah, it makes sense, I mean, it's unfortunate 'cause one hopes it comes sooner, but we can attest because I see this regularly in coding agents that we do in particular. Maybe he was talking about broad agents, but my favorite is like Anthropic, right? You're working on some code, like some of the simulators we've been building in software, and you're like, the button's not aligned right, right? And he goes, "You're right, what a great catch." Like, "Let me go fix that, it doesn't fix it." We are now 29 revisions deep, and I can't get past like the slider. You've messed it up the wrong slider. It's saying 2025, not 2026. Oh, great catch. - Great catch. - Bro, come on. I need these things to get better faster, Jay. I need these agents to work on my behalf. - Oh, yeah, I've tried to get agents to build me a basic income statement forecast model, pull the historic data down from a publicly available database in a specifically well-documented markup language, and then put that into a spreadsheet, and then forecast out a few quarters, and I can't even get it to download the historical data. So, yeah. - You didn't download the historical data. - You're right, Jay, what a catch. Great eye. - Oh my gosh, buddy. All right, well, yes. Okay, well, good that we all have a long life of job in front of us as compute. However, these things need to start making money, man, because the dollar commits to build this out are far surpassing everyone's revenue. Anthropic and open AI is included, and so it'd be nice if people could start making money on these things. So this doesn't sound so silly, but, and I'm not like a bubble person. I just see these numbers and you're like, where's the cash coming from? Somebody explain that to me. Okay, all right. Different day, different topic. Thanks for listening, everybody. We will talk to you next time, and have a good one. - Thank you, everybody. Tell your friends, we'll be back next week, and I'll still be arm wrestling with my agents.