Speaker 1Hello everyone, welcome to another episode of The Circuit. I am Ben Beharin.
Speaker 2Greetings programs, I am Jay Goldberg. And you can tell that I am not an AI because I'm very tired, and AI does not get tired.
Speaker 1Okay, what a freaking week. There's so much to talk about.
Speaker 2Warning ahead of time, people, this is going to run long.
Speaker 1We got so much to get into, we don't want to jip anybody of the goodies, of the good, bad, ugly, I don't think there's any indifferent, it's either good, bad, or ugly. There's no nominal. Oh man. All right, well, we start the week with, you know, what was somewhat expected. We started in Anthropix S1, and man, did that bring out the Warshack test of your convictions or your opinion on everything that's happening in NAI. We had Fudsters making a big deal about things that weren't a big deal, like missing numbers that are going to get updated before the IPO. It's like the numbers weren't there because they're like terrible and they're hiding them. It's time to finish the quarter, so need to update the freaking numbers. Like that's just common sense. But then, you know, for others, holy cow, they're spending a bunch of money, or holy cow, they're getting so much subsidized compute, that's why their margins are better. Anyway, lots of reading, it was very long. The risk statements were exceptionally fun. Existential risk to humanity. In fact, I hope, I seriously hope when OpenAI does theirs, like they take it to the next level. Like how do you, how do you take that statement to the next level? But they should. Like the risk statement battle. Who could make the existential risks higher, Anthropic or OpenAI? This is a new game.
Speaker 2I will admit that this week, I literally wrote the line, Anthropic may develop products which could wipe out all of humanity, which would cause earnings to decline.
Speaker 1You think? You think? We'd all get the needed break. We need, you know, on a similar week where like governments are out saying, hey, we should prepare for a wipe out of the power grid. Make sure you have food and water for three days to 90 days in a military context. Oh, geez. Governments all over the world prepare, prepare for disaster, existential risk to humanity everywhere. Okay. Well, not to get too much into. Your analysis of it, but Jay initiated on Anthropic. Let's just talk sort of high level. Obviously, you know, their revenue continues to rise. I think no one at least should expect, right, triple digit annual growth for three to four years or more. They are undoubtedly by a large margin, the most profitable. Enterprise software company in the time that it took them to get there, like fastest growing enterprise software company in history. They obviously continue to see their demand increase. There was positives. There were certainly some question marks, but let's get into it. What was your, your read?
Speaker 2So first off, I read the whole thing. Cover to cover. I may have skipped some of the boilerplate securities law language, but I read the whole thing. There were parts of it that I absolutely needed Claude to explain to me. And we'll get into some of those. It was a very dense document. But I think, like, let's, I think if we just sort of strip aside all of the extreme language, the hyperbole on both positive and negative. I thought that if you could just sort of. I think if we just sort of strip aside all of the extreme language, the hyperbole on both positive and negative. If you had to sum it all up, like what is anthropic, they actually said anthropic was founded on the empirical conviction that scaling laws are the primary engine of AI progress, which is, you know, like that's very clean language. For an S1, that's very, very clean language. I thought it was worth pointing out. They think that building ever bigger LLMs is the way to advance AI. And if you are a company that sells semiconductors, that's very good to hear. Very good to hear. Right. Bigger models. I don't know quite how they can say empirical conviction. That's a kind of seems not quite linguistically correct, but I'll take it. This is a company that believes fundamentally in AI. They believe a lot of other things, some of which are a little stranger. But they see the future as building bigger and better. Or bigger, bigger AI systems gets to better AI results. And then from there, everything gets crazy.
Speaker 1So the other part. Right. Again, I don't think anybody. Well, sorry, I'm not going to say anybody. Undoubtedly. The argument continues to remain that the demand for compute is high demand for their services are high, which, you know, is quantifiable. And I think that's a good point. They had some points about their customers, and I thought that was important, right? I think that's a good point. Because in the debate of circular economy and that everybody's just paying themselves, right, we continue to try to make the point that, you know, look, the long tail of the enterprise is what's ultimately going to fund this. And so they went out of their way to talk about, you know, the customers that they have in the Fortune 500 and then again how small of budget they get as a port of that, those operating costs or revenues from those enterprise companies. Really, I think to make the point, which is something that I had said on Twitter and others said as well, you know, we're just still very early in the enterprise adoption cycle. So as long as we're early in that growth profile of as enterprises go and diffuse and scale production AI in their enterprises, certainly Anthropic and OpenAI, right, perhaps others, but certainly they are well positioned to get a share of that budget. That is very, very large. This budget is significantly large. And so I think that was an interesting point. I think a lot of people took away from that like, yeah, sure, we're early. We get it, right? You're in a part of a growth TAM for AI as a spend of both IT budget or as a part of revenues from enterprises. That certainly goes up, right? And so now it's a matter of how do you continue to compete and create new products that go and compete with your peers? And get that budget or the vast majority of that budget. But I thought that was a helpful, you know, sort of framing. Like I said, a lot of people in the bullish camp pointed that out, how early it was. But that's a true statement. And I think it was good that they pointed that out to just show, you know, this is a growth wave and we're early in it. And we expect to get more of that budget, right, over the next couple of years.
Speaker 2Right. And I think this goes to what I was saying before, which is if you just sort of strip out the use of the word AI and you look at this strictly as a software company, it is very interesting in not just the sort of top-line growth, which is obviously very quick, but also the way in which they're building up the sort of software ecosystem of partners and distribution and go-to-market. right this is this is a they figured they figured out the right growth hacks to get this to grow and if you look at sort of the the penetration of to your point of enterprise and how much enterprises are spending there it's still a very small number for large enterprises and it's certainly they have demonstrated at least anecdotally that they have lots of proof points It's that once one employee starts using it, they start evangelizing it to their co-workers, and then those co-workers start using it, and then everyone starts upgrading. They have figured out that growth pattern there, which is, I think, very important for the company, right? That's certainly, for me, the headline. I think you tweeted this, your, your, your wife had sort of started evangelizing use of Claude to her, her coworkers or AI to her coworkers. And I think, I think that's true in a lot of other cases. Yep. Right. Yep.
Speaker 1Yeah. So let me, let me add some helpful, just color on this. So everybody could kind of level set on, on where we're at. Um, we just finished and I published yesterday round two of our CAO CTO survey, uh, which we call the, the EA index. Um, obviously there was a whole slew of questions I wanted to get into, like, you know, how far are you out in, in production? Are you seeing ROI? Because the ROI question is there, um, what's going on with token dollars? Are you going to spend more? Are you going to spend less open models versus frontier, all of those things. And, and this - This is interesting because exactly the point that you made about how my wife has viewed this as well as, as, uh, as her peers is like, it wasn't, it wasn't like AI wasn't good or cloud or anthropic wasn't good for them earlier in the year was that they didn't know how to use it. They didn't know how to tie it to their workflows and get the most out of it. And it wasn't until I sat down with her and was like, here's how you can do this for lesson planning. Look, you can create interactive games, those interactive games, like as kids play on their compute, their, their, their computers can actually track their progress. This gives you a way to like track them over time. Like these are all things that most people just don't know. In fact, and you have this conversation, we had some qualitative conversations with a host of these, these enterprises we talked to some who were like bleeding edge. You've got all these different groups who are using it to the, to the, to the max. And then you have others who are just like, I don't know what to do. Like, I'm not that far. And it's, again, it's not a, not that it's capable. It is a, it is a, a education or a skillset. A skill gap within AI. And so that has to close like, like whether that's like we talked about with Benedict, whether that's UI or others. But the other part that to me was fascinating and I, I get this, I guess I also don't get it, but like, all right. So, so top level of our survey was, um, only 6% of CIOs over 200, we surveyed are seeing a return on investment. Now at a top line, you're like, that's terrible. Like dude, 6%, that's awful. However, the vast majority aren't even tracking ROI. Like they, it's not because they're not getting it. It's because they don't know. And so you're like, okay, well, first of all, you all need to figure out how to get some tracking in place so that that can justify your spend. But the color, the, the picture gets very good of the percent, the 33 of 200 that we're actually tracking it 67% saw distinct ROI. And that distinct ROI in that decision led them to have more it budget or more budget to AI because they saw it. And the other, the other important part of this, which I think is probably the big nugget of insight from the survey was ROI scales goes up as production scales. So the more places that you scale AI, the more you see ROI. And so if it's just one BU, you see it, you bring it to two BUs or three BUs, you see more. And so it compounds as it diffuses. In the enterprise. And so that was clear as day, everybody who was tracking it directly saw line of sight to ROI with scaled production across the enterprise. Now, again, logical, I get it, right? You're maximizing, you're compounding your productivity gains, your cost gains, whatever those are, the more it goes into your business, but, but people, the vast majority of enterprises just don't know this yet. And that to me, I think is, is an interesting part. Now, the evidence for that. This point that I'm making, and this was super interesting was that we also discovered that there are a lot of companies who the BUs themselves are funding AI. That's not a part of it budget. That is the BU saying, this is great. This is amazing. This is helping my people. Everybody in my department's coming here saying I need more AI. And so the BUs are putting that as a line item. And that to me was super because that's not necessarily going to show up in an it budget that BUs themselves. Now, I guess someday it's probably going to have to. I would argue like you probably need to make sure you're using approved models and processes with your CIO, CTO and CSO. And so that should get rolled in. But the fact that BUs might contribute revenue to the overall growth of this, not just an IT spend that every employee gets is actually very interesting. And to some degree, to just use a very different analogy here, there are enterprises who as a part of their hardware program have employees. Employees who want Macs and IT says no, and the BU says will help fund those Macs. And this is not an uncommon thing, but it's also the kind of thing that is shadow AI turning into official AI as a part of BUs. And I, and I think that's very high signal, super interesting, almost impossible to track. I would tell you that right off the bat, but the fact that that's happening in all of these cases that we track shows you there's something there. Again, how much I get the timeline of this, we don't know, but I walk away. All of this to say, I walked away from the second version of this survey. I was pretty confident in this version one that there is a ton of ROI. They're willing to spend on that. We just have to get over this hump of education and scaling it to every department. And what does governed production look like and how do you put security controls and all that? We need to get over. Once this happens, once that happens, man, it's, it's, it's, it's slowly than suddenly that it budget ramps here.
Speaker 2I buy that. I, I, I think, I think actually talk about shadow it in the, in the S one somewhere where they, they talk about it. Yeah, they, they, it's a, it's a real thing. And I think it's what you say is fascinating because the, the B because typically we think of shadow it as just being the loan.
Speaker 1I am my own decision maker. So that's all fine. But you see what I'm saying?
Speaker 2What remains? Yeah, I think it is for, for people, it is, it is a little tricky to get it started. And then you, once you figure it out and you realize it's really not that hard to get started, it is, it can be very productive. And I think that speaks to what I, what I call the growth engine inside of Anthropic. They figured out how to do that in part because it's just so simple to use.
Speaker 1So, yes, yes. Anyway, so I'll, all of that to say directionally I'm with the trend timing, right? Tricky. Again, a lot of this is dependent upon enterprise scale of AI. Everybody like our infrastructure, the hyperscaler conversations that we're going to have in a bit. This is all related to when enterprises scale. Which is why, as I've said many times on this podcast, we are laser focused on studying the enterprise on this adoption cycle, because that will dictate all of this going forward. But like I said, I walked away encouraged that, that there are very positive signs for this. We just need more enterprises to start seeing those returns, track those returns, scale within their enterprise. And I, like I said, I, I think it will become a more sudden inflection point. This is again, just. Just my opinion, but I think it will be a more sudden inflection point that we probably won't see coming early enough and it's going to inflect pretty, pretty crazily.
Speaker 2Yep. I think, yeah, I think so. And I think that's an important part of what's taking place in an anthropic, but let's, let's, let's move on. Cause there's a lot of other good stuff in here that I think bears. All right, let's move on next. Oh, actually, actually, sorry. Let me, let me, let me just close on that one. One thing I did. There's a lot of dog fooding taking place. I mean, we kind of all knew this. They're using their own product to encourage their own product. Yes. But they said that 90% of, uh, of code contributions now are written by Claude. 90% of the work that anthropic is doing is being done by Claude. So when we get to a hundred percent, that's recursive self-improvement and, you know, big things happen anyway. Yep. Um, and again,
Speaker 1I think we should assume this is just the same. I would be shocked if there's not. Not that much a difference between the underlying meat of what we read between anything open AI puts out and what entropic I think these are very similar businesses in their shape, their function, their usage, et cetera. So I, I, I kind of feel like they're going to be similar. Maybe I'm wrong. I don't know, but I feel like they are.
Speaker 2Yeah. I mean, we don't, we don't have timing on the open AI IPO. It's, it's not this year. So it will be very, it'll be, it was a sleepless night. For me this week, sleepless week for me this week, reading through that S one, I'm going to have a, another sleepless week when that S one drops. But I think returning to the fact that we're a podcast about semiconductors, I think there was a lot of interesting stuff in here about their approach to semis. I think the big, the big sort of question around anthropic now is one of the things that stood out was the fact that they are, have been incredibly capitalized. They haven't raised that much money given how fast they've grown because they've been using cloud, right? All, all of their training basically was done in the cloud. She's starting with Amazon. They're changing that now. They've realized that, that they can no longer rely on third parties. They can't be cloud customers. They have to now start directly controlling their data centers. And they have this project in place called Anthropic Direct, where they will begin. Signing long-term lease agreements and various forms of controlling, not quite. owning or controlling their own data centers and this is a big deal because they're going to go from a software capital light business model to now a very capital intensive business model and again for semiconductors that's good news because they talk a lot about what their multi chip strategy is right they use nvidia they use amazon tranium they use tpu uh they use a lot of amd and they use uh they're now designing their own chip with broadcom so they're what's that five yeah five five chips they're using and i thought that was that was pretty cool but i noticed in the language this this is the level of pedantry i was reduced to this week they talk about this multi-chip strategy and they describe their relationship with each of their chip vendors and it's all slightly different right there's a lot of different things that they do and they're all slightly different so there's an awful lot on broadcom in there because broadcom is funding the development of their chip the the build out of that chip when it's that's a whole other that's dozens of pages but looking at like nvidia tranium tpu and amd they describe those relationships very very differently and it caught my eye right they say that in nvidia and amd both offer the opportunity to collaborate but for nvidia it's the opportunity to design and engineering to optimize future nvidia hardware for our workloads and for amd it says collaborate directly with amd through a multi-year engineering collaboration so on one side they're they're getting they have a relationship with nvidia where they can sort of like optimize future nvidia chips and what i read that is optimizing it probably means extra you know sort of custom software tools maybe some custom nvidia cuda kernels some something like that as opposed to amd where they have this multi-year engineering collaboration and then for tranium they have the opportunity to influence future generations of tranium and google is just there's no collaboration going on and so this is something i'm going to be watching is like who who is really doing what for anthropic it's all a little bit different what level of customization are they getting from these from these chip vendors it's clear it's different like it's just it's a little bit different but it's a little bit different and i think it's a little bit the the language is super vague but it's clear that they have different they're getting different amount of love from the different chip vendors yeah yeah
Speaker 1which you know again this is not uncommon right just back to server cpu days both intel and amd would do custom skews ish of server cpus for the hyperscalers if they wanted it throttled up one way or another or lower energy etc they would do that um same design right little different performance profile uh what what i think was that that is a hundred percent this this broad point exactly the point with the article i did on on um everybody wants the workload to be fungible right so the future of ai is fungibility that you have your workload and you can move it across diverse bits of silicon all with different perhaps performance profiles economics profiles and your master orchestrator decides where that goes for your bottom line they're probably the best example of this i'm not sure open ai or anyone else has that level of diversity across compute infrastructure with the same workload and but that again that shows you the fungibility now to your data center point this this is why meta has always done what they did right they've believed that under under that owning data centers while cost expensive gives them economic benefits that's the point of their whole blog they put out with nvidia on prometheus is that owning that owning that compute and that tranium lets them tune and find you know fine-tune things for different economic profiles based on train models faster release products faster all those things so it makes sense both open ai and anthropic are going to do this they just they'll never satisfy their own compute with their own data centers they have to outsource but they're going to get better margin profiles on that probably fill it with some of their compute right hardened silicon to their models that makes sense um i again i'm going to keep beating this horn as we have last week and many players cyber security is going a big deal they're going to need to build their own cyber security models they might run that on their own cyber security models they're going to need to build their own cyber security models asics i don't know like this is all feasible so sure being vertical there makes a ton of sense
Speaker 2i i agree but if this is this is a pretty radical shift for the industry because in in you know back in the stone age of cpu days when all we cared about was cpus you you didn't bother doing this you didn't bother i mean yeah you get a little tweak here and there from intel but you're right
Speaker 1you're right you were happy using your hyperscaler partner or
Speaker 2whatever you yeah you just use whatever and here here you have the big computer that's going to be using your hyperscaler partner and here you have the big computer that's going to be using your hyperscaler partner players have gotten so deep into it that they're doing their own and they're also working with almost literally every vendor out there right i think open ai actually has one i have a list and open it has one more partner than anthropic does but they both have six or seven each and right they wouldn't have they wouldn't have done this in cpu times this is this would have been just one one vendor and one backup and now they're just getting really
Speaker 1though doesn't this just make the case that not necessarily that gpus i i hate to say this broadly but not that gpus are necessarily appreciating assets but that owning that owning that owning your accelerator the thing that computes the most computationally complex workload in software the industry's ever seen is a strategic advantage like that that has to be the end the end point here
Speaker 2i i think it's hard for us to tell because the demand has come on so strong and so suddenly that we're everyone's playing catch up and i don't know that what we see today is necessarily the pattern going to be in the future um it's just like everybody needs everything they can right now because growth is you know a thousand percent a year two thousand percent a year when that slows down when everybody in the world has has ai in their pockets or wherever in our eyeballs how does that compute actually sort out and i i i think it will i i don't think we know i think it'll depend a lot certainly certainly if you are like like anthropic you have a business that's premised on operating at at at or beyond the frontier of ai capabilities that sort of lends itself very very well to having your own silicon um sure sure but
Speaker 1but am i wrong in thinking it and i can certain this is certainly i think undeniably true for the next four or five years but like is it perhaps out of the realm to think that like nobody will be able to just fully have their own capacity like for all their compute needs like i don't see how you can't it's inevitable that you spread this out i i don't think we get to a spot where anybody's big enough but i think it's inevitable that you spread this out that they've got 100 of their capacity anytime soon including including so i was like so so you have to pick my point here is if that's true you the hyperscaler or you the lab have to pick the most important workloads that you want to run on your data centers and your silicon you can't satisfy all of it so you have to pick and make targeted decisions at which that is
Speaker 2absolutely five years from now i don't know maybe but right now it's like you know what's that old saying in a in a in a hurricane even turkeys can fly people are going to try everything right now because they just they need anything they can get where does that settle out steady state is less clear to me i think i think i think i think it's very likely that it remains very fragmented and it gets more fragmented
Speaker 1agree i think that's a fair but but again my point is you know i think it's very likely that it remains very fragmented and it gets more fragmented even it's an it's an it's a it's a affirmation that you need to have land power shell because economics necessitate it great but you can't satisfy all of it and then be that you want to have your own compute which will include networking like all that stuff right you're going to probably still be as vertical as possible hearken back to props to ourselves very early on maybe the fourth podcast we ever did was on why do companies want to go vertical and we talked about apple and we made the case that like vertical integration is the best strategic advantage that anybody has and it's why people should go and do it and and again apple apple's approach to this as we talked about was they only do custom silicon where it's a valued workload a workload that to them is a differentiator everybody here i think applies that same lens and you just have to decide land power shell what your infrastructure is what workloads you're going to have to do to have that strategic value in in
Speaker 2something there i yes but i sort of i see it flipped around which is you only should do your own silicon if you have if you can generate some form of strategic advantage from that and that's not true in all cases i think it is true for the for the frontier labs it's absolutely true but it's but you have to you have to be able to generate you don't just do it to save money you don't just do custom silicon to save money you do it because it gives you some strategic advantage that gives you an advantage over the competition agreed agreed now
Speaker 1again i'm not going to say we know what that is but this to me is the lens they're using for that analysis like we should we should control top soup to nuts top to bottom some part of this but but i mean that's the
Speaker 2interesting thing is that anthropic has gotten as far as it has now without any of that right they use nvidia they use tranium because that was what available to them oh and and they didn't even own it now they're going to own it because the market is so tight but they've done an awfully they've done they've come an awfully long way without
Speaker 1any of that yeah yeah
Speaker 2i'll tell you where this leads me is um like if you think about anthropic like what is it is it it's not an internet company like it's available on the internet but it's not an internet company it's kind of a software company but we typically now we you know for years we've learned to think of software companies as as sas companies that are asset light this is a company that is very r&d dependent and has now gotten to a place where it is very capital intensive too and so you know we're always searching for comparisons or analogies for ai to me this looks a lot like a leading edge foundry right where you have to you have to do intensive r&d and then you have to deploy an immense amount of capital in order to monetize that i mean it's it's that's tsmc right we're going to spend 50 billion dollars a year on capex and we're going to have 3 000 phds working for us and that's how we stay ahead
Speaker 1you're right no you're absolutely right that is all true um it's just a super interesting dynamic and you're right it's a big change and it is this thing that we're talking about that is driving the concerns in the industry that we see like people are just like they get it but they've not had to like wrestle with what does it mean to kind of own that level of infrastructure for your differentiation so that it's not you know you're not you are not a commodity you don't become commoditized you have sustained strategic advantage these are the things that now enter the equation and then whoever you think has the best of that is perhaps better set for returns right so anyway yeah i mean and i think that's
Speaker 2that's the thing that gives people pause here is this is going to cost a lot of money right the one of the key things that they they left blank in in the preliminary s1 was the amount of purchase commitments they've made for for silicon and the reason they left it blank is because the ultimate value of those depends on their stock price right so when you file an s1 you leave the stock price blank until the very last minute because because it's being negotiated right and so if you have anything in the s1 where it the the number depends on the stock price you don't fill it in until you know what the stock price the offering price is going to be and so that includes all their purchase commitments with companies like broadcom and whoever because some of that is is valued and they're paying for some of that with stock and whatever that number is it's massive right i sort of laid it up i came up with like 800 billion it's probably the high end of the range it's somewhere between five and eight hundred billion dollars in purchase commitments over the next it it it's basically the next five years and those are big numbers and that kind of has people freaked out and it's going to change a lot of the things like a lot of their numbers because just to be clear they were actually profitable in q2 they're cash flow generative and profitable and now that's all going to change because they have to spend a trillion dollars on compute
Speaker 1but did you not think i saw this come up and you read it so i i did but you didn't think that some of that was because they were getting some subsidized infrastructure from their partners like they weren't are they paying traditional market rates are they profitable paying traditional market tokens i guess is my question
Speaker 2i think they are getting lower rates because they're buying in volume right they're they're like i mean they're they're they are probably aws is not probably they are aws's largest ai customer by yeah i would say that's true by a very large amount um and so yeah they they get a they they merit a volume discount on that alone and did amazon give them special terms because the equity probably that too so you know like i've done a lot of math on trying to figure out compute costs and every every calculation i get i know it's a little bit too high because open ai and anthropic are paying less than market rate right often by you know 50 less so yes it's subsidized to some degree
Speaker 1so okay but but but to me am i wrong in this view that that needs to continue in some degree the deals that they get with their infrastructure partners and i get this is win-win this is win-win for amazon this is win-win for because as we'll talk about the hyperscalers for a bit they they sort of alluded to this that these big customers uh despite having some discounts their margins are margins meaning amazon margins are still good but part of that is why we enjoy some subsidized token prices in our 200 prices that's all you can eat yeah that's the
Speaker 2that's the uber argument right like in the early days of uber everything was very cheap it was all subsidized venture dollars i don't know how how true that is i mean like okay i mean yes yes they're i mean their margins are very high right they're probably 70 80 gross margins again it's not totally clear from the hundred no they're
Speaker 1definitely that high if not higher but yes so i'm with you so that
Speaker 2that to me doesn't sound like a company that needs subsidies um will like once these guys get to steady state and are profitable you know whatever 20 29 or 20 30 whenever they get to like sustainable profitability will they raise rates okay yeah maybe yeah but all right
Speaker 1no that was that was what i wanted that was a good answer um so
Speaker 2use use your cheap tokens while you can people
Speaker 1okay let's we could talk about this all day let's move let's move to hyperscaler earnings all three happen this
Speaker 2before before we do i have a few quotes that i think are just precious right okay
Speaker 1let's see wait
Speaker 2number one is they say model architecture decisions are made with inference economics as a priority which goes to the point we were just talking about right it's just like they admit it yep yep in discussing how they go to market um they they distribute to everybody and what they they say is our relationships with these csps the hyperscalers are multifaceted and complex which i did read that
Speaker 1one that was
Speaker 2fantastic that's like the understatement of the year your your biggest supplier is also your potential biggest competitor okay hold
Speaker 1on hold on to that line because i want everyone to think about that line when we talk about some folks earnings in a bit yeah that line is the exact same ir should say multifaceted supply chain is multifaceted and complex
Speaker 2yep that's a good one um another one i i think i just have to touch on this is their governance structure okay right all right their governance structure how they're voting different classes of stock it's like five pages in the s1 and it took me two hours to read it and absolutely i had to throw it into claude repeatedly to explain it to me because it is so complicated and so absurd and so cringy they have five classes of stock right one of them is going to go away so they can have four they have four classes of stocks with different voting rights i i could talk about this at length but like the simple summary is the founders are in control that's that's pretty much all of it but oh it is so so much more complicated than that and it's very clear that they wrote this with the experience that they all went through leaving open ai and that whole turmoil around the coup they clearly wrote this put the structure in place as a reaction to like the experience they had at open ai and it it gives these the seven co-founders effective control and i think okay that's that's fine like mark has that control the founders at google have something like that level of control what i don't think they've thought through is like all the unintended consequences that are going to ensue from that you have seven people it's not one person it's seven people and they kind of have to vote in a block they vote in a block and like how they figure that out i just think it's it is a recipe for drama for years and we're we will come back to that right i'm so excited about this a recipe
Speaker 1for drama it's it's
Speaker 2yeah and and like part of part of what's going on is that they want to have like an ethics board they call the i get it i get it right the long-term benefit beneficial trust which is this yes sort of a super board that is comprised of experts from other fields who are supposed to be sort of the moral compass that make sure the company adheres to its principles right and and so repeatedly through the s1 warns people that that the pecuniary interest of shareholders is not the priority and i i love the fact that they use the word pecuniary repeatedly because like what it's basically saying is the the company purpose its goal to benefit humanity whatever that means may take shareholder interest which is sounds noble in spirit but in practice i don't think it's I don't think it may play out quite as neatly as that. Last one, last one. This is my conspiracy theory. I'm kind of very curious about this. Think back to the history of Anthropic. We all know its origin story as it came out of this team tried to remove Sam Altman as CEO of OpenAI. They failed at that and then they left and they founded Anthropic as a competitor. That's the story we all know, right? But if you think about that, that took place in 2023, that whole drama. They actually founded Anthropic two years before that in 2021, right? And I was like, wait, that's before all that happened. They had been, so I think that means they were planning it ahead of time. And so I guess there's already a movie being made about this story, but still, if there's ever a movie made about it.
Speaker 1Of course there is.
Speaker 2The real story is these guys were planning this two years in advance and they knew that they might lose and so they had a backup plan. Oh my gosh. Okay. I could go on for days about all this stuff, but that's the main thing.
Speaker 1Hey, well, hey, look, they're going to IPO. We're going to get to talk about that. Stuff will get updated, so we'll do that. All right, let's move on to the hyperscalers. Next bag, next bag. Despite, everybody continuing to say that they've got an absurd amount of compute, the AI demand is there, there's other things going on under the scenes in either CapEx or revenues or whatnot. So, you know, sort of like Microsoft, great. Google, Amazon, questionable for other things. Amazon's got more business than AI infrastructure. And so they had to talk a lot about other things than just AI demand. Microsoft, right? No secrets in PC industry being terrible, what's going on in Windows. Everybody wants to talk about Azure. All the cloud growth was there, but I guess people wanted more and wanted more clarity on kind of the AI revenue side. There was some questions on CapEx guides, where again, they were like, well, we're going to raise them. We're just not telling you all of it now. So if we could just maybe go one by one, I don't know, or do you want to bulk them all together? Let's talk hyperscalers.
Speaker 2So I think there's two things we have to talk about the hyperscalers. I think we should bulk them all together because they've all gotten blurred in my head at this point. All right. Number one is they started the week with everybody raising money. Both Meta had an equity offering. Amazon had a debt offering, both in 70, 80 billion dollars. I don't know how much, but big, big capital raises from both companies to fund data center expansion. That goes along with lots of companies doing that earlier in the year. We also had Oracle, who's kind of maybe sort of a hyperscaler in this context, had their analyst day and they actually said, so they said that they were not going to do any capital raises for the foreseeable future. Like not never.
Speaker 1They don't need to.
Speaker 2They don't need to. Well, and, and so, and it was interesting because all of that was pretty well received. And that was on Monday. Just remember Monday morning where like, you know, before 7 a.m. We'd already had five major news stories, including S1's Anthropics S1. Yeah. So all of that, all that was pretty well received, right? Yes. These companies are doing big, big capital raises, but the markets was like, okay, that's fine. They kind of looked through it. Then we get through the week and we have Meta earnings, Amazon earnings, Google earnings. I think Meta's earnings were fine. They were able to talk a lot about the engagement that's being generated by their AI agent Muse. And I, I, I think their stock was off afterwards, but it wasn't off by a lot. People were just like whatever, whatever they had talked about was good enough. It wasn't a big concern. Then we had Google earnings, which were not okay. Um, it like just there, there just weren't seeing a lot of growth. Yeah. And then the stock was off pretty heavily. Yeah. And the Amazon, we're forgetting Amazon. Amazon was in there too. Same, very similar. Story. People weren't so excited with their numbers, but I forget all the details.
Speaker 1Yeah. I mean, I think again, I, it came down to a lot of other pressure. There wasn't as much guides on, on the go forward. You know, everybody had growth in the cloud business backlog. Um, you know, to some degree it was, I think lower than some people had thought. If you're just tracking backlog growth, obviously they wanted to see continued CapEx growth higher, which is sort of what I heard. We thought they were going to go. They're going to willing to spend more on infrastructure, you know, but, but their calls like just strip away the other parts of the business. They had to talk to you. Like Google was like, Oh, you know, search is okay. At a time when I think people are really freaked out about what happens to Google search. Right. Um, it's, people forget Google search is the largest part of their business. So if that starts to hurt earnings, you're going to start to hurt despite what happens in Google cloud. Um, so I think there's, that's, that's what, again, why I say there's, there was more in, in their businesses that they had to talk to than just, if you just isolate it, and this is going to be true of a number of people we talked about on this particular podcast today, if you just isolate it to the data center or the computer, the AI side, it looks okay. When you look at the other side of their business, which might very well be impacted by the success of other people's AI business, AKA Google search, and not showing up when you use agents to search. Um, that's a fair question. All right. On their revenues, on their growth, on their growth trajectory. And so it's, it's, it's, it's positive, but also then to some degree negative. If some people's businesses are or parts of their business are up for disruption and the other one's not going to be big enough to offset it for a really long time.
Speaker 2Yeah. And I think the hyperscalers got themselves in this bind because earlier in the year they started to get questions about ROI, their return on their AI investments. And they all came back and said, well, not all, but a lot of them came back and said, Oh look, our AI business in the cloud is growing really, really well. And now we're at the point where, people are at the investors are asking for more than that. They're like, no, no, that's not enough. And to your point, like what you're saying is what about the rest of the business? Like, aren't you seeing returns from this and, and part and parcel with that is this sort of growing realization that like a lot of AI companies have compute and they're like, okay, we don't have a business model, but we're just going to resell this compute to somebody else. Elon has talked about this. Elon has done a lot of this, but there are lots of startups coming out this week saying, Hey, we we're pivoting. We're going to be, we're going to just resell the compute that we locked down. And, and ultimately that leads to the question, well, if you don't need it for your AI magic model, who is going to, who does need that? Like, right. Who's going to buy it? Ultimately who is buying all this AI compute? Is it just open AI and anthropic are the only ones that are going to be buying AI compute forever? That doesn't seem sustainable or healthy. And I think that that's kind of what's, what's going on now is like, what is your actually, what is your core business? Is it benefiting from it? Yep. Right. Yep.
Speaker 1Yep. Yeah. So, okay. So this is, this is kind of then what I wanted to talk about because while Google and Microsoft talked about this, like, despite AWS, his stock, like they, Andy, Jesse hammered this point home right off the bat, he talked about the flywheel and you know, and really like we have analyzed this, right. That all three of them, it has to be a flywheel. Like so, so meaning that parts of your, but met is the same in the same boat, right? That parts of your business in this infrastructure need to diffuse and benefit and lift all the parts of your business, right? So for Google, it's search, it's everything related to that. It's gotta be YouTube, like all these things, the flywheel of the investment in AI has to a cover itself or make money from your general purpose, your public cloud partners, but it also has to be a flywheel for your products in growing revenue. And so I guess the question is, and you might just tell me, all three is the answer, but this is what I've been noodling on is who, who, for which of those three names for Google, Amazon or Microsoft depends on the flywheel working the most.
Speaker 2So the way I think about it is Google is one. I think some of the founders there are true believers in AGI and are willing to spend to get there. And so they'll, they'll keep going regardless because they have founder voting rights. And does it help their business? I think there is just from the outside, it would seem to me that AI could greatly help their business somehow matching ads. It's, it's a, it's a possible threat if people sort of use AI instead of search. But that seems like something like they have a good, they have pretty good models. And they have a pretty good understanding of how the, the under workings of ads work, that they should be okay. And they like, they should be able to navigate that it's, it's Google, it's complicated, but I think they can get that right for Amazon and Microsoft. I think it's much trickier, right? It's much, it's much trickier. I don't know that they can get the, what they, what they need both is to see the enterprise adoption of AI grow. sharply fast but i don't know that they can capture that because they're so dependent on anthropic and open ai for the models yep
Speaker 1agreed so yeah so so what you're getting at is sort of what i was getting at is you know who who needs more help from the outside and yes you're right i think amazon is a big part of that i would agree with that and then microsoft the same way right it's got to work for office 365 and all of their services products has to diffuse from azure in the back end and convert enterprise into those two things like they're required whereas you're right google has this big business that can convert to kind of like meta my my only concern for them is it's the customer for what they're dependent on is the consumer whereas the customer that microsoft and amazon are more dependent on is largely the enterprise and i don't know how i feel about you know because search searches mostly revenue from that is is consumers globally searching whatever it is 50 or 60 queries a day um you know the back end is office right or infrastructure to some degree amazon prime commerce yes consumer but that feels pretty insulated um i'm just not sure how i feel about google and the consumer side of this because i'm very confident enterprise monetization i'm not so sure how that looks in consumer so yeah
Speaker 2i mean you and i have different views on this i think the the thing they remember for google is that consumers are not their customers right we're the we're the raw meat we're the inputs right
Speaker 1we're the ones seeing the ads
Speaker 2right but we still
Speaker 1have to see ads though
Speaker 2so yeah so so yeah we still have to see ads i i am i i know there's a lot of people saying that you know open as an ad product and people are just going to use claude i get you yes search correct yeah i i think if google i think google could navigate that right i think that they can just need they they need to sort of buckle down and really tighten it up what their offering is um because if they they could if they use a to have better search that should protect them i i think the problem right now is that open ai and anthropic have a search product just their ai which feels better than google because they don't they're not under pressure to monetize it and you get a google you do a google search and there's there's a lot of spam in there there's a lot of ai there's a lot of ads and so i can google can can navigate that and start they just it's a question of just improving the quality of their search product it's not an existential thing they just need to do a better job of something they've been doing for a long time but it is absolutely a risk it is absolutely a risk so you know right and and i think that's one of the big questions around whether or not and open ai can actually do ads anthropic's not doing ads open ai is trying to and you know it will open ai search will will chat gpt search be as useful if it's as heavily ad loaded as google isn't sure right well
Speaker 1that's that's why i like this conversation because their their businesses i mean amazon and microsoft every their businesses are all very different right they they each have a little bit sorry amazon and google each have a mix of enterprise and consumer microsoft's pretty much in that enterprise camp largely um and but this is an interesting tension because like i said they all have to have a flywheel and so it's which parts of their flywheel have the dependencies that we believe can convert on the back end of ai and i i think that's an interesting tension yeah i i think
Speaker 2they're all exposed is where i come down to right on the consumer side you're right google is very exposed on the enterprise side i i think both amazon and microsoft have something to worry about in that how much of enterprise workloads are going to go to are going to be dropped out driven by their ai decision first and then they'll say oh well i can you know i can get a lot of my ai directly from from anthropic and i'll just figure out a way to integrate that into all the back end i run on aws and then suddenly the aws side of it isn't as important that i think that's the risk there all right
Speaker 1let's move to i guess what has dominated the conversation really since yesterday and today as we're recording this again on friday october 30th um so you will all hear it then but you'll all know that it's going to be a lot of work and a lot of work and a lot of work and a lot of what i'm talking about which was intel so okay i want to come at this a couple ways first wait wait just give let me
Speaker 2give the headline intel reported yesterday they had a very strong q3 and then a very weak q4 guidance
Speaker 1go okay i have deja vu jay uh-huh this is the second time pretty much this exact thing has happened in the same quarter two years ago it was the end of pat's i don't even want to say tenure five year four year um everybody was you included jay has been our permable on intel everybody who remembers going back to the earliest of days jay was starting to turn stuff was looking good and they just they just pooped the bed and and while the whole thing was going to be horrible there was a whole slew of execution questions that dan came up on the back of that and then pat lost his job and i'm not saying libu is going to lose his job i will say however every question i've got you and i have talked about this i'm sure you're getting him has now come back to what happened with the execution it was looking good and now you can only explain this largely and we're going to get into it from some lack of planning or execution so okay okay you tell me where you want to start let's just dig through
Speaker 2it so so i i look at this as the basic premise is they have a limited amount of capacity for reasons that aren't current management's fault but they have a limited amount of capacity how do they allocate that capacity do they well what they've done is they've produced more units for pcs which is going through the worst downturn in 15 years you tell me or
Speaker 1hold on hold on i'll tell you one second it uh the worst it's ever been off an annual is negative 16 which is 2022 this is ben's pc model uh we're tracking to negative 14 for this year so the worst since 2022 okay i think
Speaker 2it's i think it's worse than that it's but i think it's gonna be because they've they talked about negative 20
Speaker 1that was for the quarter not the year annualized that was it was negative 20 in the quarter the first first half of the year was actually okay for pcs so it's it's gonna be somewhere around 12 to 14 i'm pretty sure by year end so they're
Speaker 2making more chips for for pcs even though pcs is going through an awful market and they're not making enough chips for ai servers which is going through the greatest market of all time of all time right literally and quite literally and i think what what has thrown people was they warned about capacity constraints two quarters ago and last quarter it sounded good like oh they've they're and two quarters ago they said they would work through it during the year last quarter they seem to make some progress on it and by the way this is the exact same setup as two years ago warning good quarter i think i actually wrote a professionally i was wrote a note saying hey intel's turned the corner right bad warning a good quarter and then a really bad quarter exact same setup and and then listening to them talk about it on the call and afterwards it just kept getting further and further away like on the call they said next year next year will be a growth year but it'll be back half loaded meaning it'll come in the second half and you're like uh that's a long way away two quarters ago you said this year is going to be you're like oh my god and so they are currently suffering what i would describe as a crisis of confidence among investors who are just like yeah
Speaker 1that's fair and i'll tell you
Speaker 2almost every conversation i've had with people in the last 12 hours since that since the earnings was oh is this the old intel i mean everybody's saying the same thing is this are we doing are we really doing this again and um i would say uh you know lip lipu didn't not sound great on the call and my take is he was as shocked by this as we are and i think this is his very rude uh introduction to intel corporate culture and like how much you know when when did he realize how bad this was what are they telling him internally those are the questions i'm asking because i i kind of i'm willing to have a lot of faith in him that he he knows better than to sort of throw miss miss misstep people this way and i just so i'm wondering now like what's the intel culture like internally are they just feeding him good news and hiding the bad news from him um yeah okay all right
Speaker 1so let's uh let's unpack the pc problem this has been everybody knew this was coming i think the degree of how bad was perhaps not entirely obvious to everyone even though i've been telling everyone who listens this year is going to be worse than you expected every quarter of the like hour i think it's going to be like it literally went through this i can just backtrack because i can remember my brain because i had these arguments with and with those who track this in the supply chain versus our our data it was like it's going to be eight percent for the year this was first quarter i'm like nope it's gonna be worse then it goes on it's gonna be like nine second quarter it's gonna be nine percent for the year it's gonna be worse come third quarter it's like in ten percent it's gonna be worse it's like it just kept getting worse and and no one wanted to acknowledge how how bad that was and again to the degree that they needed to do what you and i talked about i wrote a whole note on how much the pc downturn can lead to in server wafers they needed to make that call six months ago it seems like it's just happened and then it's like all right well how long is that going to take if you're just now starting to think about allocation and then clearly on the call they were like well but you know we got another pc product next year we think it's going to be super competitive right called nova lake and so i'm like please don't double down on this because i i almost want to say stop freaking making pc ships and just go to but but i know that's unrealistic i know amd is also not going to do the same thing and and whatnot but but like just make more server cpus so now it's like when can these decisions that happen late enter the market and then again to show to what degree is this uh not bad enough or not did they not are they too late right for and but but then i want to go back to the other point my other frustration is exactly what we're talking about now what you just pointed out in the earnings which was like oh it's the the the turnaround is farther than we thought is exactly what happened two years ago it's the same thing it was like ah it's farther than we thought that was two years ago and now we're at 26 right 2026 fall october and now it's like dude now it's not even next year because you're constrained and so it's 2028 and i just that's not great
Speaker 2it's not great i i will say there is a big a big difference between now and the last time they did this two years ago which is i don't think we're up against existential questions about their future right two years ago it was like they needed to get to 14a and they needed to make it work and now i think we're at a point where we're all pretty comfortable that 18a works 14a you know it's looking good and 14a is is real and will come and their balance sheet is in much better shape than they were two years ago and so it's no longer this long-term question of oh will intel survive it's much more sort of near-term tactical like what what's it going to take to really tap into this demand yeah
Speaker 1but i think i think what was also happening was there again nobody knows intel's capacity we don't know why we make guesses guesstimates i'm not sure they know i'm
Speaker 2not sure they know
Speaker 1okay fair fair but but but regardless like we we don't know right they've got mix and match of tile nodes so all their server products as well as panther lake go back to intel 3 as a base as a base die and sometimes even there's products on intel 10 so those products right didn't ramp right they they never ramped those so sure they're constrained so i get it but but the problem is i think people were hopeful that they had more in the supply chain to surprise to the upside to be able to capitalize on server cpus and maybe to some degree have an early customer and foundry than they want despite them talking about the triple digit growth for uh custom custom a6 which is a black box of a of a category right now we have no idea what that number is it's in dcai and so it's growing up triple digits from what a couple billion probably um so it's still just in the single digit billions at the end of next year um fine not bad but is that gonna surprise to the upside to offset any of this right so i i think i think the concern i walk away with is sure growth's gonna happen next year but growth is constrained and capped based on that you make your judgments accordingly as to where their value's at and it's not going to be as big or as high as you thought because now any inflection for them in either capacity that comes up for foundry or bring in manufacturing right or sorry construction isn't till 2028 right and probably second half of 2028 so
Speaker 2i'm not going to talk to valuation i will say i i think your timing timeline sounds about right in the sense of when are we going to have a better signal on their business and know how they're doing i you know it's it's not going to be solved quickly it's not it's not and and i'll tell you part part of my frustration was like we in in callbacks they and on the main call they kept like doing doing this weird thing where they're like oh we need more of this product so we're going to do this fab at this node and it's going to be a different like they kept moving around i like it was really hard for me to keep track of what's what's going to be an intel 10 what's going to be an intel 3 what's going to be an 18a and they're making big decisions there that like i if i understood it correctly they're talking about allocating fewer wafers to 18a and doing more on intel 3 which i guess makes sense but like i just like it i i i guess the idea is like the server parts are a node a node behind so you need more of that so you're going to go to intel 3 rather than 18a but then what is that impact on learning for 18a and getting 18a yields where they need to be because they're still right not quite there and and i just realized as i'm saying this is like i'm pretty sure this is exactly exactly the conversation we had two years
Speaker 1ago yeah like oh
Speaker 2yeah so oh man so then
Speaker 1so the so then the the challenge is so intel 3 is the base die for panther lake and for the server product so that's partially it right so you you need that base die to scale so you can and you're right right panther lake was more 18a and uh that that'll happen with their next generation server parts it's just a little farther out than than we want um but yeah it is the chiplet part of this that means it requires some parts that never really scaled for your chiplets and so your demand is again you might have some tiles that you've got capacity for and you might have some tiles that you don't and you need to continue to ramp that so that's this as much as i love chiplets and disaggregated design and how it helps your effort this is actually the downfall to not being monolithic planning for this across nodes is very a very interesting challenge when you're mixing matching anywhere from two to three process technologies on on one chip in different in different tiles so again take that for what it is that's a complex problem what did we say uh complexed and nuanced what was the term the anthrop yeah yeah yeah that's what we're talking about here it was it was it was complex and nuanced or whatever the case is and so that's one of those things where i get it but again right exactly what you hit on is now everything i heard which is i mean why why didn't they see this coming why didn't they plan accordingly and then my other part complaint about the capex side is it was very clear right when they talked about they're doing this capital race so they can go get all of the components that they need and everybody was like that's because they need substrates so that's why this happened this month or last month you know what's happening in substrates is so constrained which includes things they need for for emib um is so constrained that i'm worried that you go and buy this now and you're capped at how much you can get for next year how much can you really get 28 because you know all your competitors have been trying to secure this up and try to secure it up through 2030 i'm worried that they took too long to do that capital raise and go get the substrates they need even for scaling
Speaker 2in 2028 yeah i i um uh my my take was they because they blamed the miss on substrate constraints that was that was pc is being off and pc's off right and both both of those were things that were very well known in the industry six months ago and i i think it is less a question of oh you couldn't get enough substrates that's because you didn't plan it all well enough in advance and i think this is it's very clear to me that somewhere along the line in the last 10 years or the last five intel lost its planning capability um i assume it had it at one point but they've clearly lost that muscle and it is it is painful it's painful to watch this now so i i think they can sort it out i like i said i don't think that we're in survival mode if you a lot of the data points they gave out about about foundry were actually really encouraging like where 814a yields were about the demand for emib that side of it sounded really good it's just like it's gonna that was always a 2030 story it still is um and it's just the near term is gonna remain complicated it's multi-faceted and complex multi-faceted
Speaker 1and complex that that is exactly right um okay just to tie this up and then we'll move on so like my again my concern is well they're in unfortunately they need to rebuild confidence again so like we have to go through this hopefully next quarter even though they're guiding down and whatever it's fine but they present positivity but i agree with you the direction of travel is right this is this is this is not a question of can they survive they're absolutely gonna get customers i mean to be honest with you like just for example this came up like in so many different supply chain conversations i had on the back of um uh andrel making a making a big deal about their u.s-based manufactured pc using qualcomm manufactured in by tsmc i have already heard through the rumblings that like the government's going to say anything that's made for our military or our use has to be u.s companies and u.s made on u.s. So TSMC is great, but it's not that's not an American company. Right. I think that they're going to double down on these things. And so Intel is going to win customers for people who want United States based GPUs made by a United States company. Right. On this soil. That could be them. That could be NVIDIA. Like, it doesn't matter. They're going to want it manufactured by a U.S. company here. They're going to get those deals when they can scale them. How much capacity do they have? How much revenue is that going to be? Again, this is an end of the decade story. So fine. That's going to happen. It's just that all of this gets pushed out. So direction of travel. Good timing. Questionable. Management has to rebuild confidence.
Speaker 2There's our powers. Yeah, that's I think that's ultimately comes down to is going to be a while before anyone trusts them again.
Speaker 1Again, I know what Lipu comes off as like so nice, you know, like on these things like I want to know. I want to see angry Lipu behind the scenes like is he livid here with people like, you know, like where like where I would love to know. I mean, she's certainly not like, oh, you know, it's OK. And it's like, no, something happened. Like, I want to see angry Lipu behind the scenes. Like, are people scared rattling in their boots now to to make sure this doesn't happen again?
Speaker 2Yeah. Is he going to do it again? Like, I mean, he kind of did it when he. He joined. But you're right. I mean, I think he should do the publicly. I think next earnings call, he should get on and say, hey, we really let you down last quarter. We're going to fix it. I've taken steps. I've you know, we're going to cut these. We're going to cut a whole bunch of people. This has to improve. OK, I don't know if he's going to I don't know if he's going to do that, but like that would be that would be.
Speaker 1We know we have we have key people from Intel who listen to this. So there you go. There's our advice, folks. On on onward. How how assure us that this won't happen again? Yeah. OK. OK. Last on the docket of this very long episode, the longest circuit in history, which the week's the week's news demanded it. Really weird news across the wire reported by Bloomberg that Marvell is in advantage. I'm not going to go into too much detail on that, but I'm going to go into a little bit more detail about what this is about and what it's going to look like in the next couple of weeks, and then I'm going to go into a little bit more detail about what's going to happen in the next couple of weeks and I'm going to go into a little bit more detail about what's going to happen in the next couple of weeks and then I'm going to go into a little bit more detail. I have some opinions of what's going on here and why, and why this makes sense, I guess. Apparently a deal could be announced at any time next week, which might be about the time that you listen to this. But what was your high-level take?
Speaker 2I think Marvell has repositioned itself as a networking company. We talked about this at other Analyst Day a couple of weeks ago. Yep. Yep. It is. Something that they've done very well. They have deep history there. They have a lot of good assets there. They bought Cavium. They bought Infi. Those have been very, very good acquisitions for those companies, for Marvell. And if you look across the optical supply chain, it is certainly a big area of growth. And Lumentum arguably is, you know, one of the critical players in. All of that. Set that aside. Set that against the fact that it is, you know, what's Marvell's $200 billion market cap and Lumentum's $100 billion market cap. So these are both big companies. And I think the reason it's all stock is that's the only way that Marvell could afford it.
Speaker 1I mean, yes. But, okay, but why would Lumentum agree to that? I don't know. Like, what does Lumentum know? What do they know that's coming that would make them agree to this? All right. You tell me. I'm very curious now. You have a theory? I do. I do. I think you're right in that maybe this is the only way this happens with Marvell. Again, it comes back to, like, why. Okay, so let's say that Lumentum has recognized what I think, which is, A, unlikened copper. Where what I'm about to say probably couldn't have happened because there's a very diverse set of suppliers for that. I think in optical, it makes a lot of sense for networking providers to be more vertical. Basically, I think everybody wants to be Credo. I think Credo's approach, plus if you add networking silicon, is very, very interesting. Marvell has that networking silicon. Broadcom, similar. You could argue that Broadcom. In fact, our friend at Irrational Analysis. tweeted out something that I retweeted saying, you know, he thinks that Broadcom's lasers are as good as Lumentum's. If that's true, Broadcom's more vertical than people think, right? Suppose Broadcom walks in and probably really hurts Lumentum and Coherent, if that's true. Okay, so what do you do if you're Lumentum and Coherent? You don't have CERTIs. You don't have additional. You're a component provider of this, right? Your moat was that you have a part of a very complex supply chain growing these lasers that hasn't ever had to scale. And that's it. What happens if that's not the standard? What happens if micro-Vixels or micro-LEDs or other things, right, come out and play? And you're not the only component provider. Okay, so my theory is they're afraid of disruption. They see the signs on the wall that they're not going to have the moat or the leverage that they have much longer. This is a huge risk to their business, and this is true of Coherent. And they buy the idea that verticalization in networking. Networking silicon down to the optical level makes a lot of sense. Okay, so if that's the case, who else might be in this mix? NVIDIA, you know? Like, Marvell is not the only buyer, if I'm right, that networking to optical needs to verticalize. And you want to control more parts of that supply chain down to the physical lasers as a component. Why just Marvell? Like, if you're willing to go stock only, go. Hit up NVIDIA. Like, you know what I mean? Like, there's more buyers here for this if this is right. But I think they fear disruption and commoditization of this space. And that's the writing on the wall that I think they are wrestling with or have concluded to. But regardless, I think verticalization here, it makes a ton of sense.
Speaker 2I think that's a very bearish take on optical. Why? Because what you're saying is it's the optical. But how is it not going to be a commodity?
Speaker 1No, I'm serious, though. I'm the curmudgeon here.
Speaker 2How is it. No, no, no.
Speaker 1But in all reality, other than the fact that right now there's no standard, and two, they grow these lasers, and so that's a choke point, right? That's the moat. But if there was standards, and more people. Or eventually, when optical wafers for this, right, scale, and laser, you know, infi starts to scale. Because they are. Like, the industry is getting to that point where they are scaling these things. It's just going to take a little bit of time. Once that happens, how is it not a commodity? It's interchangeable. Who has the best price wins?
Speaker 2So, I think Lumentum has something differentiated in their product. Amen. I don't know how long they can keep that, but I think that they are not a commodity. I think in the class of laser we're talking about, it's them and Broadcom. So they have a very strong position. And so I think I agree with you. Long term, like five years from now, are we really going to be talking about optical companies anymore? I don't think so. No, I don't think so. I think we're going to have this big, big build out somewhere in the next few years. And then everything will get built out, and then the industry will settle on some new steady state. It will be higher than today, but the growth will fade. That's kind of implicit in your argument. I don't know.
Speaker 1It makes sense to be a part of other people's parts of the supply chain who, like I said, Credo is a great example. They win a lot. They win a lot. It's because of how vertical they are in that side. I think there's an argument a lot of people in networking need to be about vertical. And in this case, dude, I'm telling you, if Broadcom's lasers are as good, everybody's screwed. I'm sorry. Like Broadcom swipes the floor with you. Like I don't see any way that that shakes out if that's true. And so that's, again, that's why I'm saying like these dynamics change for someone who's that dominant in silicon to be able to tie it together. Tie it down for CPO and NPO have a laser strategy integrated. That's very bad for everybody who's just a component provider.
Speaker 2Yeah, I know. I get that. I think that's true. And I think it makes a lot of sense for why Marvell would want to own Lumentum. Agreed. To bulk them up against Broadcom. But why is Lumentum willing to sell is another question. Because if what you're saying is true, that implies that they're sort of top ticking. They're selling at the top. Yes. Right. Correct. And I don't have a view here. I don't know what the right answer is. But I'm just saying like if you're saying this is as good as it gets, let's sell now. Who knows? But I think that I do think, though, there is I'm not I'm not sure who else could could step in here and buy them. I think that's why not Nvidia. So I think. So first off, shout out to Reuters. They had an article out this week reminding everybody that the Department of Justice is still investigating Nvidia across a whole host of antitrust issues. I remember this. Yeah. Right. That investigation started a few years ago.
Speaker 1And Hugging Face is a part of it.
Speaker 2Yeah. Hugging Face, SkedgeMD, which is which is Slurm. These are some really, really important assets they're picking up. Oh, and let's not forget Grok, which was not an M&A, but kind of an M&A. Yeah. Right. And so the DOJ is investigating them across all of that. I'm not sure that Nvidia could get this cleared while they're being investigated for those things. Now, could Nvidia make it hard for Marvell to buy Lumentum? I think that's that's possible, too, because I don't think Marvell I don't think Nvidia wants to see Lumentum get acquired unless it's by them. And I don't think they can acquire.
Speaker 1Oh, I'm sure. A hundred percent agree. Yeah, absolutely. On the other hand,
Speaker 2on the other hand, like like Nvidia does own what? Five percent of Marvell.
Speaker 1That's right. You're right. So that's maybe by by association and partnership works out for them. I could see that. Right. Still still benefits Nvidia.
Speaker 2It's one of those enemy of my enemy is my friend kind of thing. That's exactly right.
Speaker 1Because this is it. It's it's against it's the world against Broadcom. Yeah. A hundred percent. Yeah. But anyway, like I said, this has just been speculating. But like I said, you think about all we've studied in optics. I do. I've always believed that this needs to be more vertical. I'm a big fan of Credo. The reason they win is because they're vertical in a lot of these segments. And I just felt like that makes a ton of sense. So who can verticalize here all as as deep as to the supply chain as possible for networking to networking? Silicon is going to have a differentiator. That's a big advantage because, again, it's it's you against Broadcom. Broadcom has all those things.
Speaker 2Yeah. All right.
Speaker 1All right, everybody. We pulled a long on you. We appreciate your time. Hopefully it was worth the hour and a half ish podcast. We look forward to what next the next podcast and the week will bring in the industry that never sleeps. And hopefully everybody's having fun and not losing their minds. And we will talk to you next time.
Speaker 2Have a good one. Thank you for listening, everybody. Tell your agents to subscribe to us. We'll see you next time. We'll see you next time.
PodcastTranscript AI