Ep 149: TSMC Q4 25 Earnings, OpenAI Needs more Compute and Monetization

[MUSIC] >> Hello, everyone. Welcome to another episode of The Circuit. I am Ben Baharin. [MUSIC] Greetings, programs. I'm Jake Olberg. >> So Ben, you got something new right here? >> Yes, we launched a sub-stack, as I guess everybody does these days, but ours is targeting a very particular type of audiences. We'll probably be obvious when you read it. But it's called the Diligent Stack at the, you know, can do the DiligentStack.com. You know, my approach to this sort of, it's sort of twofold, right? You know, everybody who reads my stuff knows we like to be a little bit more technical versus financial. But I did, you know, hope to tie technical realities, modes to market opportunity, you know, I don't make stock calls or have to do EPS and what not like, like you do. But I do think there's a business implication, right, to technology that says, is there upside, is there a market opportunity, and that's really what we're focusing on? I, you know, I encourage everybody to subscribe, even though there's a paid version, I'm trying to make it still valuable in the free version that's maybe 600 to 800 words. Hopefully there'll still be some insight, but we're obviously, you know, putting out more institutional grade analysis for, for that market, and that's kind of the, the direction we're taking because I've been asked to do that by a lot of people. So it's, it's Ben's return to writing a note since 2010 when I wrote about Qualcomm and Apple and mobile and telco. So there you go, that is it. Everybody should sign up, everybody should cry, go get the paid version. Yeah. And give me feedback too, and questions, there's, there's the chat, there's a way to interact with me. Also, we're still kind of using that for like a discord push to which Jason and Austin's in. That's kind of associated with just to build some community and conversation, maybe share some charts that we don't normally share publicly as is, as is happening. So, so yeah, so, so there's that. But yes, please sign up and like I said, let me know what you think, and I hope everybody finds value, whether you pay or, or you don't pay for it. But sub stacks the thing, man, everybody's, everybody's writing some, some sub stacks. So. All right, let's jump, TSMC, I think we should, we should, we should position TSMC like what are we saying for TSMC, like we used to say they were the bell weather, then it was like, well, Nvidia's the bell weather, like, what, let's start with, what does TSMC tell us, like how much of a leading indicator should we use for them versus other things before we talk about their earnings? So let's just put, let's put TSMC in context. So, so yeah, TSMC reported this week and or last week and I, I think I've, I've made this statement many times, I, I know that many people look, look to TSMC as some kind of leading indicator of what's coming next. And I, TSMC is fantastic company, but I think that, I think it's dangerous when we look to them for end-market signal. That's my take. I think so too, and, but, but I'm curious as to like, two things. So, so one, in fact, at the beginning of this podcast, because I was, I've, you know, Twitter is the worst at helping you find tweets, but like, you and I talked about this right when we were starting off, Twitter, the, the, the, the circuit. And I, and I tweeted this and I tried to figure out because TSMC was at $90 a share. I remember this distinctly because you and I were having this company and I was like, why, why are they at $90? Like, you feel like they should be higher when we know that they're going to make money, they're going to become, you know, foundational, they're going to grow with this cycle, but it was, it was, you know, at the time it was $90. But, but what I think is interesting is we, we know that they're constrained. And to some degree, in the same way when you think about the supply chain, there's just sort of an upper limit that we understand. So, so it's like, value them as they will, but their revenue is, is, is, is capped. Now, now everybody's revenue is capped, right? I'm, I'm not saying that it's just an endless stream of revenue. Like, Nvidia can't make enough GPUs, right? Their revenue is capped, right? Same with Apple, if Apple can't sell enough iPhones, their revenue is capped. But like, them like ASML, it's like, you're like, well, look, there's just limits. They can't, all of a sudden ship 100 more ASML machines a year. So it's sort of the nitty-gritty, right, of just their constraints and understanding their constraints, which is sometimes might be a little bit more financially restrictive than some other companies. And I just wonder how much that even flows into, right, people's valuations or, or thinking about the stock when they're in the weeds of, you know, a high-fixed cost business, right? So, I'm, let me preface this by saying I, I don't know what the value of TSMC is, there should be, I don't even know the stock price right now. But I do, so I'm not going to talk about the valuation of TSMC. I will say, though, that when you, we, one of the big challenges, I think, we in the semiconductor analysis business have had for 15 years is looking at all the software stuff going on. And people got very, you know, for a long time, nobody cared about semiconductors. Everybody just wanted to invest in SaaS and CRM and consumer, social, whatever. And there is excitement on that side because your marginal cost of revenue is zero, right? You just add more customers, essentially zero. And that's not true in, you know, real tangible things and hardware. And, and so there are limits here. And I know it's frustrating for people who haven't, who are coming to this from that software background, who have been hearing all about this unending internet world for the last 20 years to sit there and say, oh, wait, there's like weird constraints on that. Like I saw somebody a very well-known in some circles, highly regarded sort of pundit thinker saying, why doesn't TSMC just build its own fabs? If they're constrained by TSMC, why let that get in the way, let's go build your own fabs. And so Elon said the same thing, but Elon is talking about in a very different context. And he actually, I think, knows what he's talking about. But this person is more of a political on the political side of things. Doesn't quite get that like building a new fab is not just $30 billion of building. There's like hundreds of billions of R&D that have to go into it too. So long way saying, I think there are real constraints here. And TSMS is always the challenge with TSMC is we out here, you know, people out in the world say, oh, we need more chips. We need more chips. AI is taking off to the moon. And the team at TSMC has to sit there and say, well, how much of that is real? This is a big focus of their prepared remarks. And especially the questions, like they got a ton of questions about half the questions we're touching on their thoughts about AI, right? How real is AI? And they seem pretty bullish about it. But they admit that CC admitted, he said, how do I feel about AI? I'm nervous about it. I spent the last three months talking to all my customers and all my customers customers about this. And then he went on to say that he's, you know, after having all those conversations, he's now confident that AI demand is real. They say that AI accelerators, which I think includes GPUs, is going to, their business doing AI accelerators is going to grow 50% this year. Yeah, I think it's 50%. Which is, you know, it's a big number. TSMC also, but he also said, like if this doesn't pan out, we're, you know, this is really bad for TSMC. Yeah. They're going to spend 50, 54 billion dollars in CapEx this year. Yeah. They're going to spend, they spent over the last three years, they spent 100 billion in CapEx over the next three years. They're going to say it's going to be much higher than that. So we're already on a trend for 150. It's really about money, right? And if, if it doesn't pan out, if the man goes away, then that's, I mean, they have losses. They have massive losses. Massive losses. That's why they, that's why they've historically been risk averse. That's why they, you know, everyone is, like, egging them on to build more, more, more. And it's dangerous. And this is, this is the point I always make is that they're at the tail end of the bullwhip, right? And when the market, when the market, if, if semiconductors are ever going to be cyclical again, and the demand for AI goes away, they are not going to be the first ones to know that, right? So, yes. I am, I'm completely in tune with all of those things because inside baseball, for context, for everybody's listening, Jay and I trade a lot of charts. And there are some charts that we've been trading that show drastic cyclicality in this industry. And it's like everybody who's covered this industry for, for what, Jay, more than 10 years. I mean, we've done this for a long time. Like, you just know, you know the peaks and valleys. And you're just freaking out. You're like, this is big. It feels big. Yes, it's all up to the right. But we know it doesn't stay there, nor does it flat line and normalize. It drops, right? And that's, that's just the angst. That's the, the PTSD that we all have who's covered this industry. And that's what I think CC way was really like emphasizing. He's like, I know the cyclicality of this business. We know what happens when we over build. And we do not want fabs sitting at something less than 75 to 80% capacity, which is usually the number you use to make sure that a foundry, you know, makes, makes money on itself. And so that's, that's the concern. And so I, we, we, we, we, we move forward knowing that this is good, right? You're right. In their commentary, their margins were at an all time high by like 0.2-ish percent. I think they've hit 60 a couple times. This was a little over 60. Yes, they're increasing capex. Yes, they're on pace to, you know, grow revenue. And she'll probably be in, you know, a sizeable increase this year as they were over, you know, 100 billion for the first time in 2025. So it's the slow march. But everybody knows, right? The underlying supply chain dynamics here. And that's, that's the angst. Like everybody, like I said, everybody who's talked to his business here for a long time, they just know they're like, we can't get too optimistic because when we do, we get burned. And it's, so that's the, that's the financial angst, I think, across that we're seeing. So why so many debates center around this angst? How much bigger can it get? When will, when will the downfall come? Because it's really a win, not necessarily an if and a lot of people's minds. Yeah, certainly, certainly in my mind, it's, it's a, it's a win. But I think we're in fairly uncharted territory. It can go on. It can go on a lot longer. Yep. Yep. The, the, the very, very deep, deep there inside of me is whispering. There's a last to know. They're always going to be raising, raising capex at the very peak of it all. But that's, you know, I could be wrong. Yeah. I hope I'm wrong. I hope it goes longer. Yeah. But they're, they're very, they're very bullish, bordering on aggressive, I would say. Like this, their overall tone was beyond sort of the raw numbers. Like they, they talked about, they're pulling forward their second fab in Arizona. Yep. Right. It's, the building's done. Now they're installing all the tools. They're pulling that forward to get it up and running by the end of next year, I think. Yep. They're expanding in Japan. They're expanding in Germany. They, they bought another parcel of land in TSM in Arizona, which if you, which you've been, if you've go see that, like they already have a lot of land and now they're buying more. That's like the ultimate bullish signal to me. I mean, of course, land in the middle of that part of the desert is probably not that expensive, but still, but they were, they were saying all kinds of very, very bullish signals beyond just the big capex number. Yeah. I mean, so look, I, I, I, I, I want to preface this with saying I understand how difficult of an exercise. It, it is that I attempted, but like we, we firmly believe that the total found, let me back up what I attempted to do was say, what do, what do I then think is the absolute total way for demand, barring no constraints, right? I get the constraints. We know the constraints, but like, what, what could it be? Like it just to meet some of the demand that we're seeing in America and we think it's around 20 to 25% of a deficit, meaning that you could use that many more wafers, right? Now, the challenge here and I think everybody needs to like really, really understand the fundamental element of this is that the chips that everybody's making for AI and HPC was, I'm going to blank on this one. It was 60 something percent of revenue now. So the shift has fully moved to HPC, right? To have HPC as a mix. Not it's been growing up. We've talked about that. Every earnings call we've done this. These are gigantic chips. They take a ton of wafers. So if you just say you're at a 20 to 25% and that, that could be conservative, right? Wafers deficit, they probably need like five more foundries. Like that was sort of like what I concluded was like, you probably could use five more foundries and right now you'd keep them full, but I can't guarantee you're going to keep them full in four years and that's, right? That's that angst of just there's complexity in the challenges to manifest. And co-host is such a bottleneck. Like every time we have this conversation, we're like, people don't realize how challenging the packaging is to scale. And I think I talked about this early on in the circuit, but a lot of people are new listeners. So I toured Intel's advanced packaging foundryf in Arizona. And like I'm walking and they're walking me through all these steps to package one, you know, all the all the tiles together into one, right? Sort of package. And I was like, this is not scale. This is very like this is not printing one ship. This is a very difficult thing to scale. And so you have giant chips, recording a ton of wafers, normal wafers. You've got advanced packaging. You've got co-host wafers and Intel will face this too. This is such an interesting bottleneck. And again, like you need more hands, you need more automation. Like you need far more things that go into an advanced packaging facility than you do just, you know, printing wafers. But my point is is that you could argue that they do need, like I said, a whole lot more capacity, but what they have to balance, which again is the sensitivity we're talking about is how much of that can I guarantee will stay full? And so because of that, it might be two foundries, right? Instead of five or more, which I think they need to, again, meet this demand when these chips are so big. And you, you know, you only get a handful per wafer. I'm dressed over emphasizing or being traumatic for the point, but it's far less than 500 per wafer that you get right for a mobile SFC. These are fascinating manufacturing and scale challenges that they're up against. And that does not change any time soon. But I'm just intrigued by the capacity demand is so much higher than they can really meet, right? And again, I don't know if that means sure Intel and Samsung and others really are sitting on more of an opportunity, but the bottom line is, you know, TSMC is going to remain severely constrained for at least this year, if not the next two, at a minimum. Oh, yeah. I mean, they talked about sort of their planning and what they, what they're really saying is, I don't know, I don't know. I forget the exact term. It wasn't a majority, but a lot of the spend this year and next of CapEx is actually going towards automation rather than capacity, right? Because they have the Arizona Fab coming up next year. So that's some of it. But then there are a lot of it is just doing the things that you're talking about is automating as much as they can. Automating more as much as they can. They talked about a lot of their manufacturing capabilities. And then they talked about 28 and 29 as being even higher CapEx because that's when A16 ramps. Right? And so there's, so the next two years aren't even about capacity. They're just about, you know, capacity is part of it, but it's also about automation. Like that's not the time to worry about CapEx next. Later, the numbers get bigger. So I mean, and look, I'm being a little bit like cautious on all this and talking about the big numbers. I want to just want to emphasize that like TSMT is awesome. Like they're executing incredibly well. They are. It's really, it's great to watch. They're really solid company. I mean, and it's just wild too. Because like when I was doing sort of the margins exercise, like we were talking about, like just to see crazy spikes like in their margins, knowing what happened at that time, right? In 2012, 2013, where it just bottom, like it's such a, but if you just look at their operating margin now, like it is up into the right and it is almost the same line as Nvidia's, which I think is fascinating, which just shows you how much dollar value they're captioning. Now, I know they're raising prices and yadda, yadda, yadda. But all of that speaks to their execution, the amount of money that they're making, right? At an, at an operating profit level. And I mean, really, like you track a bunch of companies in the space. And I had done this where I just said, like show me operating margin for, you know, all the companies we're talking about. But I added Apple and Apple and video and TSMC are standouts from everyone else I put in that list who you would be the most you can guess all the other names in that list. And it's just, it's just, it's just showing you, again, the dollar value capture that they're getting because of the execution, like you said, because they're largely monopoly for now and all of this. But it wasn't just wafers. Like I said, like coos adds so much demand to that. The service is supply chain of that. Like there's just so much additional value capture that TSMC is getting, which is just not the kind of thing you're used to seeing in foundries. You know, it's just, it blows my mind. So I mean, there's a couple things, I think are important for context here. One is when they, when they talk about their gross margins, they say there's six things that play into it and price is only one of them. And what, what their CFO said was price increases last year only covered inflation. Right. Because the tools, the inputs are getting more expensive. They pass those costs on to their customers. Right. And it's, it's interesting to me, because every, every few weeks, there is some report out of Taiwan or somewhere in the supply chain saying TSMC is a positive race prices really strongly. And I think they, they framed it very differently. Now, maybe they're spinning it a little bit, but I think there is, there's room for them to increase prices further. Like this is sure. I've obviously, I watched a lot about, but it is interesting to me that they said price increases only covered inflation and that the rest of their improvement and gross margins come from other factors like improving process and mixed shift. The, the other thing is, I did analysis some point last year where I looked at revenue contribution or profit contribution to TSMC from by each node. And I had to make some guesses in it, but yeah, whenever I did, it was like, it was about a year ago, it was, it was clear that like the advanced nodes costs are called three nanometer. Yeah. Was just that break even. And so now we're at a point several quarters later where that process is, is yielding fully and is contributing heavily to margins. Yep. Because you think, you think that like, oh, the newspapers are going to be the most expensive and have the best margins. They actually, in the first couple years, they actually have much lower margins, because the yield time is right. That's right. And so like 70%. I think it was almost 70% of profits were coming from trailing edge nodes 16 and higher. And now we're at the point that we're, we're three nanometer is, is contributing fully. And that's like, that's got to be a big factor driving profitability. Huge, right? Huge. And right. And you know, that will go down in 27 when, and two, starts to rant more. But, you know, that's, that's the main nature of it. Yeah. I mean, the thing, the thing that speaks back to exactly what you said, though, is that three nanometer remain still in high demand. Like I agree with you, the shift will happen. And at some point, I want to make in that. But like, that demand still remains high when you're at this capacity challenge that they're at. And, and, and again, right? You still might need base tiles, three nanometer base tiles and some capacity for advanced chiplets at some point in time. Like that may not totally go away. But it is interesting. Like, I don't know how much you had, go ahead, go ahead. I think so. I think they're still really heavily utilized in trailing edge. I think so. I mean, yeah. A meter. Agreed. Like they said, they're, they reduce, they reduce capacity a little bit for eight and 10 inch wafers, which is like 30 year old stuff, right? Yeah. Yeah. Which, but, but like nothing about anything older than that. So they're still running those lines right? Post the full high utilization levels. Yeah. Yeah. Agreed. And, you know, and I, I remember this debate going back to, you know, pre this giga cycle of semiconductors where, you know, every, every sort of analysis kind of centered on like, could seven nanometer be bigger than 10? Like that was the debate. Like, here's all the things that we're thinking about. Like 10 was a great node. And can the next one be bigger? And then it was, okay, seven was amazing. Can five be better? Like, here's the structural dynamics where like, and the answer is yes, each node is doing far better, bigger than the last one. Like they were always like, could five be bigger than seven? Yes. Could three be bigger than five? Yes. Will two be bigger than three? Yes. Like this, this trend just keeps continuing. And, and this, and this again, I think it's at least what's insightful to me to think about this is in a disaggregated design era, you use all those processes. You don't, like this monolithic was just one thing. So when they said you go from 10 to 10 to five, that was being driven by Apple, making monolithic SOCs. That's just one chip per wafer. You know, the, the advanced packaging side for these chiplets is for sometimes five different tiles per wafer on different processes. So actually think that's my point in that the longevity of these, you know, N minus two, three, maybe even four nodes actually has a lot of life in a, in a disaggregated design era era. And, but yeah, but you know, you will, you will continue to see a draft high demand for each new node and each new node will be revenue wise larger than the next one. That's, that's the trend for, you know, ever since 10 and I don't think that goes away. True. But even here, there's a little bit of a worry, which is they also talked a lot about how CapEx intensity cost essentially goes up with each node. They, they, they, they, a lot, like, share a little bit. It was in their opening remarks. They talked a little, pretty seriously about CapEx incentive. It goes up largely around tooling costs, right? And I think what they're, I think what they're really trying to signal there is with starting with eight, with 16, eight, eight, 16, so not the next node, but the one after two nanometer, there's, that's when they transitioned to high NA, EUV systems. And those are, those are really expensive. Like those are, it's a whole new, whole new tool from the SML. It's, you know, closer to 400 and 300, like current systems. And I think they're trying to signal that. But well, yes, agreed, agreed. But I think again, if this train continues, right, if you just look at their chart of CapEx to even gross margin and revenue, like it, it's growing. Like there's a, there's a lot of structural things where that's contributing to each other. Again, gross margins. I mean, I'm not going to say they can't keep going up. I just, just to see a foundry in the 60s, still, still blows my mind. But, you know, there's a degree of, there's, it's all interrelated that I think that they're playing off of to keep, again, that investment leads to dollars, you know, drastically is impacting their operating profit because of the value of older nodes depreciated foundries like you're, like you're saying. But, you know, again, you, you, all of that to say you still have to keep that they are a foundry in your analysis. They are not in video or Apple. But it's just, it's, I guess, I just keep coming back to, it's wild to see these kind of numbers from a foundry. Like we're just not, we're just not used to that historically. Yeah. Yeah. Well, a couple of things I noticed that we're interesting is they said that they, they were talking about like, what is AI doing for customers, right? And they, they said, oh, they talked to hyperscalers. I think they talked to meta in particular. I think they gave them very strong signals that meta, you know, AI was really helping meta's business. You know, AI was generating a real return there. And so they pointed out, TSMC pointed on the call that AI boosted their gross margins by almost 2%. Right? So, yeah. Right. Using, I wonder if they're using like kulitho from Nvidia, things like that. AI tools inside their process. A two, a two increase in gross margins here is small and, you know, two percent is not a big number, but in terms of like gross margin dollars, it's, you know, two billion dollars last year, and savings just from using coding assistance. Yeah. Yeah. So I mean, not necessarily be totally kulitho, although they have talked about, I'm sure it's more. Yeah. Well, you know, but also like synopsis and cadence have, right, design agents, right, that help you. So, I don't know, it's a good example. I thought that there was, so, so just going back to CC ways, comments that like they checked with their customers, and they checked with their customers customers. He made it sound like they were having conversation with software companies as well. And that software companies are seeing the value of agentic drive value to their software, to their enterprise businesses, to which we all right acknowledge that is under penetrated, right, less than 20 percent of enterprises are deploying at scale. So I think, I think maybe that was part of it too. He was like, there's still a lot of adoption. There's positives that this is something going to keep being adopted. And that's going to, in fact, the software industry is going to grow drastically, right. You've seen some numbers where it's like, another trillion dollars of software dollars above sauce coming for agentic or more. So, but I think that's my read on that was that was really what he was getting at was that he thinks it's a build out and that AI software's early. And so there's enough demand, right, to really justify this increase in, in forward traffic. So, right. How much revenue do you think packaging contributes? Okay, hold on. A year or a quarter percent. 10, 15 percent. Yeah, yeah, a little bit less than 10, a little bit less than 10. Yeah, I know it's not a lot. Those way first or cheaper, I get that. I just, the complexity of it is the part. So, let's, okay, so let me get, let me speak to the automation point on this. When I was at Intel's foundry doing this, there were a lot of bodies moving wafers around for advanced packaging. There's a lot of humans in this loop. To be honest with you, I don't know how a machine does that until it's a human, I'd robotic, but there, yes, there might be some, I just, I was stunned at the bodies required to do advanced packaging of moving wafers from one stitching machine to another one to the substrate one to the hybrid bond. Anyway, it feels very expensive, despite the money that make, I get, I get, I get co-wafers are not the same, but I guess I'll be interesting to see if that, if that changes. Like, does that number go up a little bit each year from 10 to 15 maybe 20? You think about it like five years ago packaging was essentially zero for them. Yeah, nothing, right? Plus, but it's a, but to your point, it's a really good time to be a company that does process automation and process semiconductor robots. Yep, testing, robotication, yeah, five names I can think about that are interesting there. So, but here's, but here's, okay, here's the other thing too. Like, right now, this is co-os-ass. Maybe some L, I think, is trying to, then you got co-os-p, then you've got, you know, wafers, you know, much deeper into the substrate. Like, maybe there's, again, a price curve that comes up with that, like maybe co-os-p or it's such, it's more expensive, right? Then co-os-ass, right? And so there's a little, that's, again, that's super supply chain. Like, that's not public information. That's supply chain noodling, trying to figure out, but I am intrigued by that part, right? That this very complex process, how they might start to raise prices there. That's a good one. I like this debt. Let's keep tracking co-os-ass, contributions to revenue. So I got another one for you. So if, if packaging is a little bit less than 10%, what percentage revenue or masks? I mean, sub five is bigger than, bigger than packaging. It's 10%. Oh, it is. It is temperate. Yeah, there were a little more than 10%. Yeah, they didn't give it exact number, but they said the two of them together are about 20% and they said that packaging is less than, a little bit less than 10. So, and I say this in part, not to talk about the financials, but I know that there are a fair number of listings who work for semiconductor companies and they have to pay for masks at TSMC and just know that that's 10% of revenue. That's a serious amount of revenue for the company, right? It's very expensive to purchase some pay for masks at TSMC. Yeah, it's a lot of money. Fascinating statistic. Yeah. All right, any closing thing about? Yes. I have two things. Close with one is they got asked about Intel. Are you concerned? That's right. Yeah, let's yeah, right. It's easy said. He was he was trying to be polite and then he just said, but no, no, I'm not concerned about them. You were polite about it, but he was not concerned, nor should they be honestly planning to go around. And then my favorite thing was when he said, you know, one of the things I learned talking to all the hyperscalers was that they're really good at manage long-term planning around electricity. They'll have five and six year plans for electricity and he's like, yeah, we should do that too. So, I thought it was interesting that even even I mean electricity is a bit hassle in Taiwan, has been prolonged time, but it's clear that they're going to, you know, TSMC's going to open a nuclear plant. Who knows? That's nice. Making things up. I'm joking. Yes. Yes. Yes. No, but you're right, right? Power. I mean, that sounds like they're securing Taiwan. Right. It sounds like they're securing even more land and then obviously the power to that land. Yes. Okay. All right. Yeah. All right. So, that was a good nut on TSMC. Let's switch to open AI, had a deal with cerebris. And so, there's a couple things that are interesting to me about this. Obviously, open AI is admitting that they'll take every bit of compute wherever they can get it. So, make their own ASIC, buying video, buy AMD, partner with cerebris. Who interestingly, like other than the other than GROC, right, that, you know, kind of made an inference system, cerebris is interesting because they've been building data centers infrastructure running cerebris gear, right? To just say, hey, bring your workloads here, right? So, you know, building, building manufacturing, securing power, etc. And so, open AI, you know, of course, it's like, yeah, we, we need it. We need faster inference. We need to run because we can't get enough compute. It's not immediate, right? This is out there. I, I will say they aren't, cerebris is not going to solve this problem, but I will say the average reasoning time for me in agentic tests, I, I throw it to GPT is like an hour and a half. And I, I mean, I get like if I task a human, I'm okay with them taking a couple of hours, but like, I'm used to this being like software being faster. I'm not used to having to wait an hour and a half. So, I think everybody could agree we want faster inference. And that was one of the, you know, one of the telling points in cerebris's value is lightning fast inference. But anyway, my base read on that was, you know, hey, we'll take, we'll take a compute wherever we can get it. Here's another company that's got some infrastructure. So, let's work together on some, on some infrastructure. Thank you for reminding me. I've been running a cloud task agent in the background that I needed to give it permissions. Keep going. It's been, it's been going for days. So, yeah, open AI announced that they're going to partner with cerebris, cerebris on something. I, I will say that I'm at the point now where everybody is, the open AI now has announced one of these announcements with every major chip vendor. And now they're, they're going into some of the smaller ones. Like the, the only one missing at this point is Qualcomm. And I, you know, why wouldn't we see that soon? Have they made an announcement with Qualcomm? But I missed that. No, no, no, no, no, I was just racking my brain on it. Is there names that are that they're missing? Right. AMD Broadcom in video. Yeah. Cerebris, like, it's just everybody. I, I, I, I, I, let me just say this, I was at a, a social function last weekend with a bunch of AI people who work at some of the big AI companies. And, um, it was a running, almost a running joke, uh, like, open AI, snowsing with everybody. And I think it's important to remember that most of these deals are not finalized. Like the deals and videos as far as I know hasn't been finalized. Nobody signed anything. So these are press releases. Uh, I don't know what we call vaporware in the age of AI, but like, we're bordering on that. And so I think we have to take it all with a, with a big dose of salt. Yes. All right. So let's go into that because then the, the question, which I'm sure is what the angst you, you hear with this is, uh, with what money open AI. And so today, we're recording this on Friday, January 16th, you will all listen to this Sunday or Monday of the following week. But so this Friday, January 16th, open AI has announced, um, ads are coming. And kind of what I thought was sort of hoping for because they have this in other countries like India and a few others where there's like an eight dollar a month kind of entry level fee. Um, so they're bringing that everywhere. And, and I, and I say that point because, um, there are by different accounts and standards, 100-ish or maybe more 200, you know, ish million people who pay for iPhone storage. And I'm not going to say that that's not valuable. But I just wonder, you know, if AI has some value, is it worth five bucks a month? Is it worth eight bucks a month? Like, is there an interesting subscription model there? Again, not the $20 a month, but, but something lower. Um, so it's not all ad driven, even though the ads they are showing are not bad. They're very similar to if you see them, how Facebook does this in a post? There's like a little ad. It happens to be something you've been searching for because they know this. Um, it's not bad. It's, it's, it's relevant. It's contextual. This, this is them starting to say, let's start, let's start making some more money because we need money to afford to compute. So positive step, depending on, we'll see, we'll see how it goes, but they clearly are trying to change their revenue trajectory and both ads and kind of the eight dollar now entry level globally is the start and not direction. Yeah, I, I, we'll, we'll see. I mean, the, I do you think the context of you think ads hurt this drastically? Like, this is what I've been debating. Like, is it really hurt your experience that much where you're just like this stupid? I'm gonna try something else. It, it depends very much on what, what, what the ads look like and how, how they implement ads. And if anybody is curious about this, go back a couple of weeks ago, Ben Thompson on his library podcast, interviewed Eric Schweifert, who's a mobile, a mobile, or not a mobile, he's a digital advertising expert, right? Really good, good insight on his own. And where they came down was with was if open AI has ads, they're going to look more like Facebook's ads than they are Google ads. So you're not going to like type into search, you know, type into search into chat GPT and say, hey, what's the best, whatever. And it's going to give you sponsored results for the best, you know, restaurant to eat at. Yeah. Instead, they're going to be like Facebook where it's like they just know a lot about you. Right. They have all the all your context. They can figure out what's an ad for something you might be interested in that's not necessarily related to the thing in front of you. Right. So I don't know what their ads are going to look like. If they look, I think, I think if they do sponsored results, I think that's a disaster. And I think people will move away from that because you won't be able to trust the results. Right. But there's another model that they can adopt, which is these personalized ads, they know kind of what you're interested in. And that won't infringe on the quality of your actual task in front of the agent. Right. I completely, completely agree. And there's elements of this like, okay, Ben's going to complain for a minute. But like I still am shocked that when Ben does product research on Gemini or GPT or whatnot, like I can't meaningfully get links to go and buy that stuff. Right. Like it's still a gigantic pain. They're like, go search, a Gemini will tell you, go search this term if you want to buy these things. And I'm like, dude, just put it in freaking Gemini. Like, I'm here, your Google. Why is this so difficult? Right. So there's a lot of that underlying stuff that I think they can definitely monetize on. But I think that like exactly your point, that element of it knows something about you. It knows kind of what you've been searching lately where you're trying to get to based on the conversation with something. But it is, to be honest with you, it's such a natural fit for a conversational. Like I do this all the time. Like I'm constantly like trying to upgrade my, my B situation and like a honeybee guys. This has been talking about honeybees. And I'm like, do I need to get them, you know, a feeder? Do I need to be feeding them with things? Do I need, it's winter. Do I need to put a, a wrapper around them? Right. And then you got to go search all this stuff. Like it's just, it would be nice if it's like, yes, you do. And here are the three best products in your price range that can get, that you should go, you know, look out or buy for this need that you have. Like it's going to know that through my conversation better than my sentence search. So I think there's a lot here, but you're absolutely right. It's got to feel natural. And again, if everybody takes a look, you can see the demonstration that they showed it. It is exactly like Eric said on Ben's podcast. It feels very much like Facebook, which though Facebook can also be completely out of control and unrelated. So their recommendation and their recommender engine needs to get better. But it's improved, improved. So anyway, this is open AI because, well, yes, and it gets gotten worse because of AI at the same time. But yeah, this is them trying to monetize. So I'm not going to fault them for this man. They need to make money. They are not Google, right? Google has a huge advantage here in monetization. Yeah, I'm not going to fault this for it. But we'll see. I'm glad, I mean, honestly, like I'm glad they're making a go at this because they do obviously need money to afford all this compute. And that's not just going to come from an IPO or debt raising. They should start driving revenue, real, real revenue from the whatever it is, almost billion people that they think they've got, you know, using the service. Yeah, so I signed up this week for the pro plan on cloud. I'm in the topic because that's the, apparently, it's, everybody's using that. I'm still, I'm still figuring it out. But I know a lot of people are starting to use it pretty intensively. I just don't I just don't see it. I guess my, I guess where I'm coming to is like, it's useful. I can see how it could be incredibly useful. I don't think we're there yet. I think there's a lot of variability still here in what the ultimate business model is. I'm not convinced the way we use AI in the future is is what we think it is going to be. So I, yeah, they need, they need, I guess they need revenue. I'm not sure they. I don't know how much to do. I, I'm still, I'm still in the camp of business, business models for AI are still to be determined. Yes. I agree. I think the monetization point were, were they're talking about here is for consumers. You know, the enterprise side, I think there, there will be increasing value that makes a higher price point justified. I don't expect a consumer to pay 20 bucks a month for any of this stuff to be honest with you. So it's really what, what can they get? Because they, they have a ton of people like using this now, trying it, like using it to replace search. Like my, my father-in-law, who is one of the, the, the slowest adopters of technology I've ever seen, but, but was a high user of Google. Like, loves chat GPT. He loves the free version. He searches the, he's just curious. He's a curious dude. And so he loves GPT. He's not going to pay. He will never pay for that service. You got to monetize those people, right? So, so we'll see. It's like I said, it's a good start. This will be interesting. But, you know, they, as we've said, they, they, they need to start making money. So we'll see how this goes. And you're right. The business model may not be settled, but at least we're trying to refine what it is. And, and this is a start. All right. Last on our agenda was, we've been tracking GPU pricing for a while, which everybody loves to have this conversation. And I need to caveat this entire conversation with not all GPU instances are created equal. I think this gets vastly underappreciated when somebody shows you a Bloomberg terminal chart for H100. You're like, well, which, which eight one hundred is this? GPUs? Is this two GPUs? How much RAM? They don't tell you that. They just look at like some normalized. So not all instances are created equal, right? You can charge more for a partition that is higher GPUs and higher memory. You can charge less. If it's less memory and only a couple GPUs, which is where you see $1 to $2 spot and on demand pricing. But we've been, we track this. We have a tracker, J and I, and we look at these trends. And there's a couple interesting things. Yes, as I just said, people are charging a dollar. Similarly, are the hyperscalers charging significant premiums over everyone else. And I thought that was interesting. The pricing power of the hyperscalers. So that was my top line point. I'll let you make your top line points. And then we can dig into a couple others. Yeah, a couple of things have stood out. I mean, the data one is just how little actually stood out. Pricing is kind of where you think it would be. There is clearly, you know, black wells are more expensive than hoppers than amperes in a pretty predictable fashion. But prices are pretty stable, right? We haven't seen any big declines or any big swings. A couple categories have gone up. Pricing the neoclouds or just across the board is pretty stable for now. And we don't have a ton of history here, but yeah, so far so good. I think the premium, this is very much a glass half full glass FMT kind of thing. Because you look at the hyperscalers, AWS, GCP, and Azure, they're pricing for GPUs is sometimes double what the neoclouds are charging. And I'm yeah, like, okay, I understand they have a premium. They have the best customers. They have teams that can support those customers. There's lots of reasons why if you're a big company, you're going to go with AWS. Rather than some, you know, some neoclouds you've never heard of, that was Bitcoin mining six weeks ago. The flip side though is I look at those, I mean, it's a pretty stark gap in the pricing. It makes me a little bit nervous about what the hyper, what the long term trend is because over time, people will get more comfortable. Customers will get more comfortable at branch going further to when the savings are that big. Right. And if the really, the truly sort of common layer here isn't your AWS console, it's your it's your CUDA code that can run perfectly well on other GPUs. Maybe you start to consider, right? You know, both Core Weave and Nebius have serious efforts to get into the enterprise. Core, you've just hired a chief revenue officer who came from Amazon. You know, again, I'm not saying you go to some tiny little your Fortune 500 company, you go to some tiny little neocloud, but like there are serious alternatives there who are much, much cheaper. Yep. Now, in fairness, transitioning is hard. Like GPUs are very, very hard to work with. It's not nothing's perfectly portable. To your point, like just saying, oh, this is a hop, a H100, there are 1000 flavors of H100s and they all sort of perform differently. You have to be able to manage that. But still, I think, I don't know, it's right now, all the pricing is good. Let's go with the headlines, go with the good news. Pricing is stable. Yeah, and I think there's another thing too, like I've had this conversation with some enterprises where like I've said, hey, you know, there's a couple names and they're on our list that they're charging like $1.65 and I'd say like, you know, how much is pricing like matter to you guys? Because they're like, I wouldn't use those guys. That's like going to a Motel, you know, like I would rather go to, you know, the luxury resort for the quality that I know I'm going to get. So again, not everybody's going to be in a position to be picky right about their prices, but there is a quality of host hotel. Actually, I like this hotel analogy a lot. In the green field space of building new hotels in luxury land, you've got cheap ones and you've got luxury ones for your tenants. Yeah, yeah. And if you're, you know, it's like any company, right? The CEO and the head of sales are going to sign a big deal, they fly business class, they stay in the fancy hotel. And it comes sort of mundane, and the cost cutters get involved. It's everybody's staying at the Motel. And it's fine. Like, you know, hand in in is perfectly legitimate in a lot of markets, clean, it's safe. Yeah, close to the freeway, right? So, yeah. So the only other thing that stood out to me was like, it's very hard. Because this was one of the reasons I started this exercise was like, I want to price grace black well. Like, and it's very hard to do this without talking to a salesperson. Basically, you have to be a company willing to sign a long-term contract with the hyperscaler to even start having a conversation about price for grace black well. And like, so it's so to Jay's point, what's stood out to me was what's not on the list. And we cannot track grace black well pricing because it is not available outside of long-term leases for the vast majority of people at these hyperscalers. And I thought that was interesting, right? That this is one of the things where you put new infrastructure in, and you're predominantly going to sell that to big enterprises on long-term contracts. And, you know, again, that's great because there is a metric here that I know the hyperscalers know based on their infrastructure costs, so data center costs, which is get enough long-term contracts at a set rate and that infrastructure pays for itself in three years or less. And there's a number that they know that if it's above that for that time, they're making money on that infrastructure. And so that's why I think these long-term deals are super interesting for new infrastructure because that's driving a lot of profitability on ROI on that capex for AI workloads. Yeah, and this is, I mean, this is a super hot topic right now. When everyone looking at the neoclassage, so much debate, we get into this whole topic of depreciable lives of the servers and what they're worth. The key indicator is going to be pricing, and we don't have great signal on Blackwell. But we have some, we have some, not zero. It's some, yeah. There is a discount for Hopper. The big question I have that I tend to ask the neoclassage every time I talk to them is we know what spot pricing is. That's what we're tracking publicly, spot pricing. And we also know that contract contracted long-term pricing. We don't, we don't know what those are, but it's safe to say that it's lower than spot pricing. Right, you get volume discount if you book long-term. And so as we see shifts in, you know, eventually Hopper prices are going to come off more. My question, the question is, to what degree does contracted pricing track spot pricing? And we just don't, all of this is so new, we don't really know. But presumably over the long term, it sort of rationalizes, but I think we're we're a long way from getting clear signal on that. Yeah. Anyway, we will keep sharing those insights as we track this almost weekly, but I think not a lot changes week to week, but there are little nuggets, nuggets we're hunting for that that we'll share as well, because we know this is an economic and market opportunity conversation for for GPUs. So anyway, we'll wrap there. Thanks for listening, everybody. We always appreciate your time and your comments. We look forward to a great 2026. We've got lots of interesting stuff lined up for you guys. So take care and we will talk to you next week. Thank you, everybody. Click like. Leave us a review. Tell your friends and be sure to sign up for Dance New Substack. The Diligent Stack. Thanks, everybody. [BLANK_AUDIO]