Qualcomm's Data Center Debut, OpenAI's Jalapeño, and the Memory-as-Strategic Infrastructure Debate | The Six Five Pod Ep. 310

Speaker A: Daniel, you're kind of the chip guy now.

Speaker B: Well, you're, uh, you're, uh, you're more. You're more handsets and PCs. I mean.

Speaker A: Oh, uh, that is so true. Oh my God.

Speaker B: You're the. You're the. You're the devices. Founder and CEO of More Insights and strategy.

Speaker A: Patrick Moorhead, CEO of the Futurum Group.

Speaker B: Daniel Newman. Do we know what we're talking about, Morehead?

Speaker A: We absolutely know what we're talking about. Welcome to the 65 pod. It is episode 310. Daniel and I are back from New York City to make sure we can shoot this. Daniel, how you doing, my friend?

Speaker B: Man, I'm doing good. It's Friday, uh, shout out to my youngest child, Matthew, 10 at 10 years old. So I've been kind of doing daddy duty this morning. Celebrating his big day. Double digits, you know. Yeah, he knew me back when he wasn't even here. You knew back when he wasn't here, he was. So it's time's flying. Time is flying. But, uh, excited, exciting day.

Speaker A: Yeah. Um, I'm gonna hit. Hit, uh, up my family. So my, My son Patrick is moving out officially today. Moving trucks are going to be there, taking his stuff. He's got a nice apartment not too far away in case I get lonely and need to. Need to, uh, you know, go see how he's doing. But I might shoot, uh, over there after work and, and see how it's going. Definitely has a, an amazing apartment. Uh, really proud of that kid.

Speaker B: So.

Speaker A: Busy day. Sorry. Busy week, man. We spent, uh, three, literally three full days in New, uh, York City. Uh, we're going to talk about the Qualcomm, uh, investor day. Uh, Dan pretty much was on every single possible show. You could be on, uh, out there as well and broadcast. Uh, I hit up two, but, you know, I am, I'm blessed to do that. We've got a great, got a great show for you though. I mean, we're essentially talking about the, uh, control layer. Like everybody wants control. Like, is it. Is, is it chips? Uh, is. Is it software? Uh, is it the, uh, the model makers? Is it, you know, the energy producers? It's just. Is it just the memory, uh, vendors at this, at this point. Right? Um, uh, you know, you know, are we lacking money, uh, for capital? But anyways, I'm, uh, looking forward to, uh, pulling. Pulling that, uh, pulling the thread on those two. Not a lot of earnings this week, but, uh, Micron and, and Cerebras were, were two, I think telling, uh, indicators and we're gonna talk a little bit about Nvidia's $25 billion bond anchor as well.

Speaker B: Dude, it was the super bowl of, of tech earnings this week. You don't need a lot of earnings when you have Micron. I mean Micron storage, memory, commodity. Right? I mean, I'm kidding.

Speaker A: I mean I'm, I'm waiting for the uh, the trifecta, the ah, triopoly of, of memory just to start uh, buying ah, buying compute, ah, companies now with all that swimming in cash, right?

Speaker B: I mean, think like two years ago when Micron was just dying, it was just a disaster. Um, had like.

Speaker A: We'll get to that.

Speaker B: No, I know, I just, just like, more like think if like one of these hyperscalers had bought them, like how different things could be.

Speaker A: Yeah, totally. I saw some uh, some discussion about what if Amazon had bought, uh, micromes. Good stuff. Now we have a great show for you. But hey, let's jump into the decode where we kind of, we separate the signal from the noise about the, the biggest tech topics either on the news or uh, in posting on social, uh, media. Let's dive in. All right. Daniel, we could probably spend half the show on the payload of news that came after this, but you and I both went to uh, the Qualcomm Investor day. It really was the data center day. They did talk about a lot of other important topics, but it really was all about the data center. Daniel, let's uh, let's break this down and hopefully you can leave me some oxygen.

Speaker B: No, I'm gonna take it all. I'm gonna hog it. You, you already did all the, the rounds and tv. I didn't get a chance to talk about it. So the, the first of all, this is the moment for a victory lap. For many, many months, maybe even a year now, uh, I have, and I know you have too been very outspoken and despite quite a bit of, you know, I would say, uh, contention or even negative sentiment across a lot of parts of the technology. Pundit, analyst, uh, and equity side, Qualcomm is going to make an entry into Data center this time and it's going to be different. I think the challenges, and nobody knows this better than you, Pat, but when a company has kind of tried to get into something in the past and failed, that particular history tends to weigh heavily on them when they try again. Um, there was skepticism around AMD for a long time because it had exited a business, came back and I mean, obviously now they're back with a vengeance. And I think that stock only goes up. I mean I'm joking but it's been just a parabolic run. But for a long time we basically said that, look, Qualcomm has the ip, you know, going back to Nuvia, deals with Ventana, deals with um, Alpha, Alpha Wave. They've been buying uh, pieces for some time to, to and they've been sort of showing their hand but they haven't been like super crystal about what they're doing. And so while the entire market ripped for like three years, Qualcomm was sort of left to be a smartphone licensing IP company with an interesting auto business. And so this had to be the coming out party. And so over the last few months there's been little bits of news. Uh, humane win, um, potential a couple of hyperscale wins. They announced an AI 200 to replace, um, a fairly low volume AI 100 part, um, and an AI 250. Is it rack scale? Is it not Rack scale? They showed racks. It looks like a Rack scale. But you know, how are they going to come out? Are they going to do merchant, uh, type parts for the hyperscalers or are they going to do custom? Are they going to do both? Um, what about IO? Because if they don't have IO you can't go rack scale. So all these things have been sort of just ruminating about Qualcomm. And so this was the day that we started to get some clarity. So what are they doing? Well actually looks like they're going to do a little bit of all the above. Except maybe not the Rack Scale. It doesn't. That didn't really seem to be a thing that came out. But the Alpha Wave acquisition clearly entered them into a number of custom hyperscaler agreements. Um, they talked about two, didn't tell us who they were. Um, I have some ideas. I don't know, maybe you and I can come back to that. I'll leave that for now. Um, they also announced these CPUs. So they're going to have custom data center CPU, but they're also going to have a standardized CPU that comes later at high scale volume version. Um, they're going to have I.O. and they're going to do network, uh, chips and then they're going to have an AI accelerator. Um, I think the two big sort of moments, uh, during the presentation from their new data center leader, Tony Pialis, was his getting uh, both Meta and Microsoft to present. But not only getting them to present, he didn't get tier twos and tier threes to show up and Endorse, he got, he got the A team, um, Satya standing in front of his usual bookshelf doing his usual thing, announced um, ah, a custom A CPU deal for Microsoft. And then Meta, um, announced multiple generations of uh, CPUs that they're going to build together. So this was probably the proof point. And what did that all amount to? Well they're talking about all this together. Um, they put a really, I think a really modest target on revenue. $5 billion of 27 FY which by the way FY27 for Qualcomm means later in 26, um, and then 15 billion in 29. So my feeling, my gut was they basically forecasted the revenue they already have in the bag. Um, just a guess. Obviously everyone knows custom and partnerships on these chips. We're seeing it play out with Broadcom, we're seeing it play out with Mediate, seeing it play with Marvell. No guarantees, first generation succeed, no guarantee, multi generation deals, all scale and ramp. So all those things are still to be proven. So that's kind of what, what the biggest things that caught my attention there. And I'll just call out they also because we could, why not announced um, a 4 billion, $3.9 billion deal to acquire Software Play Modular. And um, this one's interesting because their story is all about making hardware agnostic. Um, you know, an AI runtime compiler that works on any chip. Uh, is this a lock in against cuda? Is CUDA even still an interesting debate, you know with a lot of developers right now with AI, uh, code development. Um, but having said that among the developers, my reading, I'm not a full stack engineer despite the rumor, um, is that this was an interesting platform that had a lot of potential and now people are sort of watching closely how Qualcomm deploys this because this history Arduino other softwares, they haven't necessarily come in immediately and embraced open community. But some of the conversations we had there, Pat, seem uh, to indicate that the customer is open or sorry, Qualcomm uh, is more conscious of this and is going to play their hand. Well, you know I suggested a little bit like IBM played the red hat hand might be the right way to play this one. So I said a lot but I didn't say I didn't go into tons of detail. So I'll give you some air to maybe go into some of the details. And this was like you said how we spent like two and a half days so a lot here.

Speaker A: No, I did a good breakdown and essentially you Know Qualcomm didn't just show up with a roadmap right which a lot of companies do. It really showed. It showed up as Satya and Zuck, uh as named customers, uh a an important software uh acquisition uh to give proof points and add capabilities that uh they can do data center software to in addition to spreading that out across their entire edge to cloud um um portfolio and put up a beefy but I think very conservative $15 billion. 20 uh, 29 target and by the way for people uh to say oh my gosh, 5 billion. So low folks we're talking uh, you know zero to essentially zero to 5 billion at ah this course, you know with a little bit in there for uh alpha wave uh ip that is huge. And just to give you that proportion uh amd uh in their entire first year of their first GPU accelerator for AI did $5 billion of revenue. Okay, um, and I think that's a good uh sobering a point here. Now Most of that 5 billion is going to be the uh custom uh the two custom wins that that were named. I'm going to go out on a limb here. You had Satya out there, uh who, who talked. You had Meta uh and Zuck. Well, well who's left? Right, you got, you got Google uh with tpu um they already have two, two to four vendors helping them uh full scale and back end already. Um so I don't think it's going to be Google but it sure seems like it could be Amazon ah at this point and my big learning into talking to all the hyperscalers the last six months and you can also buttress this up against um Nvidia's newfound heterogeneous computer and disaggregated uh compute view is you're going to need more than one accelerator folks. Okay so you know let's say there's Trainium or let's say you're trying to do some interesting package of, of of pulling um um an Amazon CPU um uh Graviton with a Trainium. You know that that might make a really interesting kind of high bandwidth and who knows you know maybe the interconnects NV link um uh between the two. Anyways I'm just making up stuff at this point but just given by who was on stage and and them saying it was a U S there's a US Hyperscaler uh in there. It's not too hard to, to figure out who uh it might, it might be the only I m Want to call it oddball or the only. Could it be Oracle? I mean, I guess, uh, but Oracle, you know, Qualcomm's uh, accelerator isn't ready, uh, yet. Um, and from a CPU perspective, um, they're buying ampere ah, chips already.

Speaker B: What about a Chinese hyperscaler? Could it be.

Speaker A: Yeah, I mean Bloomberg had a ByteDance rumor on that and I could see high probability and probably they don't want to bring it on stage because it's China. Right. Even though most of Qualcomm's handset revenue comes from China.

Speaker B: Right, yeah. Qualcomm's very strategic in China. It just would not surprise me for one of the wins to be. Be. Be China.

Speaker A: Yeah. Hey, one of the key technologies, um, that uh, that Qualcomm brought out was, was this um, hot hbc.

Speaker B: Okay.

Speaker A: And, and that wasn't. That's basically an architecture. It's not, don't confuse this with hbm, but instead of having a separate package for the memory, it, it actually um, um, is layered on as a metal layer, uh, just like you would do a monolithic chip. So, so, um, this isn't what I would say. You know, I'm not Babe Ruthing this, uh, because, you know, and you know, Daniel, you and I kind of had the same well, hey, if this is so good, why isn't everybody else doing it? Um, type of thing? That doesn't mean it's low value. It could be absolutely breakthrough. And it's not only available on the cpu, but also on the accelerator, which, which is, is pretty cool and essentially puts the compute, uh, under the um, LPDDR ram, uh, stack. So, you know, is it packaging? No. Does it accomplish very similar things as what you might expect with HBM at a dramatically lower cost? Yes. Um, has Microsoft and Meta signed up? Uh, and do they have confidence enough? I, I did hear that the chip had been taped out. And that doesn't mean that taping uh, out something, uh, is an amazing thing. And it's good and you got to get the samples. Uh, but there's work to do on how to do um, high volume manufacturing on that. So very exciting. So Qualcomm bringing in an architectural differentiation, uh, when some people want to ask why do we need more, like why do we need more accelerators? Why do we need more, um, uh, data, uh, center CPUs. So there weren't a lot of details about the CPU family. And it's three.

Speaker B: Right.

Speaker A: Uh, uh, one for a head node, uh, one for kind of a general purpose cpu and one for an agentic, um, um, Version. I don't know any details around that. You know uh, with agents it's, it's more about the bandwidth than it is about the uh, you know the you know int performance but it would be nice uh, uh to know what's going on there. I, I didn't expect this. The other, the other thing I didn't expect was uh, them getting into networking and you know I think I said on X I shouldn't be surprised uh but I am surprised. It was just how many uh you know AlphaWave was a core IP developer for some of the biggest hyperscalers by the way including Amazon. Um, I don't know if uh, uh people uh know that or not but um, they happen to have a good, a good serdes, even though nobody talks about them having a good serdes. And for all of those out there, um, SERDES is essentially the core building block um for all um, all networking and you know they, they gave a pretty good case uh albeit this is going to be you know uh, part of the revenue that hits a little later. They did give a, a three year um roadmap. I mean they do have in production 800 gig UM copper, 800 gig optical and EECs in 26 and 27. It's 1.6 T and then in development for 2028 UM 3.3 uh point terabit optical uh modules and AC. This is serious business. And this, this it's funny was probably my biggest um surprise coming out of this. And Daniel, to your point, you know they're not bringing out a full scale up uh solution uh at this point but uh scale ah out and scale across seem uh, seem very possible. Um uh at this point is there anything we uh, we may have missed, Daniel?

Speaker B: I don't think, I mean no, I mean we could dive into robotics and stuff and I think we can just put a, a very quick point. Nugul uh de Gaulle who's led the, the automotive was able to put an exclamation point that's been a steady diversification and probably part of what I call their permission to win in a new category has been their AB category so successfully. And then the only thing I'd say is in robotics they are really trying to build a stack that kind of cuts across all their different portfolios. Um and you know very software centric. Um and I do think even more so than data center path that robotics is an area that they really do have the right sort of ip, the right sort of low power, um, the right supply chain. Like they just have what looks like, and this is just a little further in the future. But I don't think people are valuing that part of their business at all right now. I think it's kind of like the way IBM's quantum is viewed. It's just something in the future, but I think it's something to keep an eye on.

Speaker A: Yeah, that's good, Good point. I uh, will note that the company did pull in their automotive uh, revenue target by two years and they also said that they will shortly become the largest automotive electronics supplier on the planet. Right. So given all the dispersions that were cast years uh, before, it just shows uh, what, what the company can do. I'll also add that uh, uh, the company updated its uh, 29 revenue targets and they essentially doubled uh, their non. Handset of business from a 20, 20% CAGR to around a 40% CAGR 22 billion to 40 billion. By the way, the only reason the stock didn't, didn't, didn't rocket ship, uh, was because the balance of uh, of the handsets, you know, so by the

Speaker B: way, I put a rocket ship after one of the tweets. That might have helped.

Speaker A: No, it may have. Yeah, absolutely.

Speaker B: So you should try it works well.

Speaker A: Okay. I think we uh, we drained that uh, gave it its uh, gave it enough credit there.

Speaker B: Let's dive into 50 of the show.

Speaker A: Yeah, exactly. Hey, let's get into uh, some more uh, uh, AI and and chip news. Um, there was a picture out there on X by the way. If you're not on X, just you need to be on X. It's, it's where it's happening. It's where CEOs, it's where Daniel and Pat, uh, are uh, exactly. Constantly posting and. But you got to get in there. But a picture of uh, Sam Altman, uh, and Hawk Tan essentially, you know, not only showing a, A full wafer of their new jalapeno chip, uh, but also you know, a small little packaged uh, die shot. So you know, I, without trying to get in, there were people trying to dissect whether it was an HBM 4 or 3E. It's likely, likely 3E. There were people saying it, it, you know, it looks like a training chip and other people are saying it's an inference chip. Uh, you know, the companies themselves said that it was inference. But I also remind anybody, everybody that Trainium, um, Train, uh, is actually primarily can train but is primarily used as an inference chip. Uh, I want to do a victory lap uh here years and years and Years ago, uh, three years ago I was talking about the benefits of heterogeneous compute, uh, and the value it brings. And I was very clear that for an xpu, if you know what software you want to use, um, it is going to be more efficient. You may not be able to use it as long as a gpu, uh, but it is more efficient. And now everybody and their mother is, is doing a, an accelerator not only on the hyperscaler side but, but now on the uh, on, on the, on the model side. So that's really, that's really the uh, I think that's really the, the big picture here and everybody, you know, telling me that everybody's stupid for not, not, not using um, Nvidia. I mean listen, uh, there are cons to not using GP GPUs uh, out there, but what is clear is these companies are willing to look beyond that and, and make it happen.

Speaker B: Yeah, I, I think that we've probably beaten this philosophically to death. It's an, and not an or. Uh, there will continue to be utilization within all these hyperscalers of heterogeneous uh, compute architectures, both compute types, you know, CPUs and NPUs and DPUs and XPUs and GPUs and DA, DA, da, da. But also in, you know, within even certain categories, you know, they're going to need more compute, let's be clear, like um, they have capacity constraints, they have demand, they have to fill that demand. And a lot of deals right now are being done with both short term in mind. How much compute can I get immediate access to? And then they're being done with longer term in mind, like can I continue to pay 75, 80% margin, um, to have access to what we could argue is the best uh, end to end compute platform in Nvidia. Um, and I think what'll end up happening is you're just going to end up with an amalgamation of all of the above and it's going to be necessary because we just still don't have enough. And so I, you know, I, I think in the long term these companies have to be thinking about their margins and their business and their revenue streams. And you know, OpenAI is just an interesting one though because they want to do everything. Uh, you heard their IPO got delayed this week. Um, you know, they're building chips, they're going to build fabs, they're going to build social networks, they're going to build browsers. They just need to figure out what they are though. Like they're, they're like, you know, that's a little bit of a different thing here. But one thing I do know is if they're going to serve the most advanced models beyond the frontier and absolutely be able to command premium margins, premium valuations, and get the support of Wall street when they do go public, they, um, can't have constraint on compute, so they need to control their destiny. And I think this is part of that process. But, uh, we'll see when it gets deployed, how, how well it does. And like we said earlier, Pat, first round, how good will it be? Because so far, uh, TP1 wasn't great. Training, um, one wasn't great. Uh, you know, uh, like, where's. What is an example of a custom AI chip that on its first pass, absolutely crushed it? Is there, is there one?

Speaker A: No, there really isn't. Gosh, I'm embarrassed I didn't bring that up. But Daniel, you're kind of the chip guy now.

Speaker B: Well, you're, uh, you're, you're more, you're more handsets and PCs. I mean,

Speaker A: uh, that is so true. Oh my God.

Speaker B: You're the guy. You're the, you're the, you're the devices

Speaker A: on the SAS guy now.

Speaker B: Right? You did finally really make that leap. I'm really glad it worked because you, all these years you were sort of chasing, I think I want to be in software. And you're like, but I'm a chip guy. You know, I'm a chip I hardware guy. But you're like, I want to work with software now. Finally, all the software companies are like, pat, I really want to work with you now.

Speaker A: Yeah, Daniel. I mean, it's very similar to my investment strategy. Right? Which is, which is just double down on investment, uh, when stuff is going cold and then get, get out of the stuff as it's about, uh, rocket

Speaker B: ship by the top, baby. Sell the bottom. You and Kathy Wood, like, you're following Kathy, you know, she like, she went all in on cerebris at like 350.

Speaker A: Oh, I know. That's so bad. By the way, I, I am looking for a, a potential way into both SpaceX and Cerebras when they hit the pavement.

Speaker B: But, uh, yeah, it's like my 4020 foot spreads in 2028. The ones where I can buy it at like 36 bucks if it goes there.

Speaker A: Yeah.

Speaker B: Um, by the way, anybody might hear this and get mad about that idea. I don't really believe it'll get there. But if you know a little bit about options trading I'm collecting the premiums, um, on the idea that someone does believe it'll go there and if it does land there, I'm going to be happy to buy it there because I think it'll be a great deal. So.

Speaker A: All right, I love that Daniel. Okay, hey, let's go into our next topic. Uh, you know we talked about everybody's trying to get a one up on each other and now even the Frontier Labs are trying to leverage supply chains. I mean quite frankly the OpenAI jalapeno chip, uh, hopefully it's not too hot. That's not good. Um, that's essentially trying to take control of your supply chain. Now we've got Anthropic, uh, and Micron kind of locking in a multi year ah, deal. Right. Similar pattern as Samsung and Sk Onyx pre funding Anthropic in May. I mean Daniel, is just this, just, just a bunch of circular mumbo jumbo or, or are these Frontier Labs truly trying to disconnect?

Speaker B: Can it be both?

Speaker A: Yes.

Speaker B: I mean, can it be both? I mean it's, it's, in my opinion it's the same circular stuff that Nvidia was doing two and three years ago now. But the ball has moved. I mean look at you know, uh, the Jim Chanos complaints. Uh, he's a short, famous short seller that always talks about like the companies that control the mean should do better than the ones that are dependent on it. Well, guess who's dependent on Micron right now? Everybody, everybody is dependent on these memory companies. Nvidia is dependent on like so, like when the uh, Neo clouds are ripping and core weave and Nebbys are running and people are going, well, why isn't Nvidia? Well now you got Micron that basically gets to be daddy over everything. So uh, you know, I think this is when you have the kind of cash flow they're creating, you have the kind of leverage that they have. This is where you start placing your chips on the table strategically. So they're doing the same thing. They're basically starting to pick winners. They're, they're aligning strategically. We'll talk about it in the earnings with all the new strategic agreements that they've set out, much of those agreements which have you know, basically floors but no ceilings on much of the agreement, other pricing. I mean these are, people are so desperate right now for capacity and just access to supply in the future that they're willing to do these kinds of deals. Um, so, so this is interesting and I mean this isn't going to be unique to Micron. Samsung will do this, SK will do this. Um, because we, you know, the constraint and you know we'd heard 27, 28, our data says 20, 30 minimum. Um, and it's just a lot. So. But what does it include? I mean this is a multi year HBM dram, ssd. Uh, you know, supply across their port days, their portfolio co design of next gen memory. So that's interesting. Co designing um, you know for storage architecture for AI workload. So they're, they're more and more customized versus commoditized potentially. Here Micron, um, strategic investment. That's the part where everybody says it's circular and then the, you know, Micron's going to use cloth. I bet you they already were. That's really not a, that's really not a big thing. Um, so that's kind of the headline here. Um, and like I said, I think everyone that either makes a chip or needs chips right now wants to become friendly. Remember the trips to Taiwan we talked about why Jensen and Lisa, they're so good because they made those trips to Taiwan and they, they sat down with CC and they, you know, and that people didn't understand that you know if you don't make that trip and build that relationship you will not get the supply that you need. And, and people are just upsourcing this. Now they have to make a trip to Taiwan but now they got to stop in Korea on the way home. Ah and, and, and apparently they have to make a pass through Idaho as well.

Speaker A: I love that dude, I love that. Honestly, not, not too much to add here. Right. And we, we're going to debate the, the whole memory strategic uh, memory tactical and slop. But um, this is, this is going to be, this is going to be a fun one. No, listen pragmatically and you know there was a recent rumor, I don't know if it's correct, that uh, even anthropic, uh, has what was it, a 14 FCF margin. Um, which you know, very rarely do you do you hear about an FCF margin. But just the fact that they're cash flow positive is absolutely insane. And at a time like this, when your valuation is so high and you're actually making cash, why not, why not try to. And you're not disintermediating uh, people necessarily. You're, you're um, just trying to take uh, more control. All right, let's get into our next topic here. SpaceX signs a 6.3 billion dollar compute deal with reflection. Uh, AI 150 million dollars uh, uh, per month. It's funny, some people are saying that that SpaceX, SpaceX is now the largest commercial AI infrastructure provider with $80 billion in revenue. Oh, it's funny, I, you know if you combine uh, the 300 megawatt colossus, one lease at 1 billion per month, you got Google at 920 per month. Now reflection, you know, you know SpaceX sitting on $80 billion in contracted um, um, revenue and that alone is $150 million per month, is larger than most public AI startups entire revenue base. So this is absolutely crazy. This just came of left field. You know Danny, you can't think that the, that SpaceX, um, and or Xai's theory uh, was we're going to buy a boatload of GPUs and then become a hyperscaler. Like I, I think the fact that XAI has not hit, hasn't hit for developers, it hasn't hit for uh, consumers, um, and it hasn't hit for really uh, for business users is they were, they were just sitting on a, on a pot of gold. Maybe at the time when they bought it it was 10. But um, through lack of use it ended up um, it ended up being absolute uh, gold. So I mean SpaceX kind of figured out how to be AWS without even shipping a console. Uh, and the backlog is, is surprisingly bigger than, than most clouds.

Speaker B: Yeah, I mean look, I'm like so torn on SpaceX because I'm such a fan over the next 10 years but I'm such a, just I can't. And um, I'm the rocket ships guy but I can't make this make sense. You got a $5 billion rocket business that's 85% of the market, that's growing at single digit percentage. You've got a basically satellite communications business which is interesting and it could disintermediate a lot of like you know, mobile connectivity businesses potentially. I heard that, you know I read something today that maybe they're going to come out with something to compete directly with like T Mobile, Verizon, like a mobile service. Super interesting for phone but like that entire industry together is like half a trillion bucks, maybe like three quarters max. Um, and then they got an AI cloud business which again is starting to look more like Nebbys. Um, then it looks like uh, anthropic. Right? I mean it's basically become a massive and to its credit like great growth. Um, I did call this out. Google owns 7% of SpaceX, so doing that deal right ahead of the IPO probably wasn't accidental. Um, look, all that capacity that they have that they don't feel the need to reserve it for what they're doing building is interesting to me. It's just interesting to me like, well, you've almost planned.

Speaker A: They're almost planning um, Xai to be. To fail.

Speaker B: Well, at least as a real frontier lab. And maybe it just becomes a massive infrastructure play and that's okay. But that doesn't trade at. It's a different multiple, it's a different business, it's a different promise. So that's what's really interesting to me is like what is Xai going to be? Is it like I said, a new core weave? Is it just a massive compute, uh, deployment? And what kind of valuation does that deserve, even if it's the biggest versus a full stack, which is really what I think the sale is, is that they're more like Google, right? They are uh, models, frameworks, uh, development, uh, platforms and of course infrastructure all in one shop. And of course they don't even have their own silicon yet, so they're still dependent on, you know, others. And we do know Elon's building a Terrafab. So I mean that could change in the future. So anyways, all the things, all the things. But this one's interesting. But I mean look, they, at least they're not sitting there empty. I mean, God, you'd think in this current economy how shameful it would be to have that much AI compute and not have anybody using it. It'd be wild.

Speaker A: An interesting thing. You don't hear much about Japan, uh, and AI but uh, Japan Sakana AI ship Fugu plus plus Fugu Ultra. This is not a big huge base model. This is more of an agentic uh, orchestrator. But it is interesting if nothing else. You've got uh, a uh, Japanese sovereign play.

Speaker B: Yeah, I mean this one was interesting. I don't pay a ton of attention to what's going on in Japan, but the way it was sort of pitched was that this thing has um, some agentic capabilities that outperformed the kind of top available frontier models and even potentially were on par with ah, some of the now unavailable fable mythos. Uh, models. Um, it's more moe, right? This is more distillation moe than it is a true model. And I think it was your son that came over the top and called it lame in my tweet and said there's doesn't. He doesn't even have like a harness or something. He Was he was telling me like why it's not cool. Um, but anyways, I mean ignore him.

Speaker A: What does he know?

Speaker B: What does that kid know? You know. But um, probably one of the most interesting things is they try to. This is another deep seat thought thing too where they're trying to sell it as oh it's meaningfully cheaper. Um, you know it's one uh, there was another story about GLM 5.2 this week but they love these stories that it's cheaper. Input token, output token, it costs less. But look, I think what's interesting is increasingly models themselves are not moats. I think that's the interesting part. It's gotta be something more. It's the full stack. It's the harness, it's the tooling, it's the connectivity, it's the looping, it's the agentic, it's the total compute capacity availability. Um because again as a whole the open source community is doing a good job and less to speak specifically to Japan and more to just speak generally to even what's going on in China is they are building really good distillations of the top models and they probably are stealing a lot of the IP because that's how they do it. But like the M. The mode itself has got to be more. And so we're, I think we'll just kind of continue to hear this stuff um, new models, we'll probably hear from the Middle east, they're going to do this too where these new locations are putting out new models that have uh, attributes that are on par if not better in some cases than some of the more popular Frontier.

Speaker A: Yeah, I mean the way that I look at this is it's kind of like open router. I mean people can correct me on social media but it just. Because really what they're talking about doing is taking different parts of the workload and applying the right model ironically it's like Perplexity Computer which is hey, I'm going to offer a lot of different models and I'm going to route you uh, to the right, to the right thing. This is more of the developers version um, of that versus the Vibe coders uh, like me but uh, I agree with you Daniel. We are going to see stuff like this. The thing about um, uh, Sakana AI is it primarily uses um, you know a lot of the different, you know it has its own models but it also is doing a multi agent routing across all of this. And you know the fable, the fable claim was very similar uh, to the, to the Open router claim is, hey, I can take, I can take 4.8 plus 5.5. Break up the workload and get, and get the same response as Fable. I like the innovation. That's why I, I really um, I, I really like this one. So hey, let's move to the flip here, Daniel. Um, you know all this talk about memory and you know, Triopolis, not monopolies. I use the term paybacks, uh, on um, on X in my X posting.

Speaker B: But hey, don't cry to me, Tim Cook. Yeah, you abused us at your top and our bottom.

Speaker A: Exactly. So the question is, um, is the area, sorry, is the era of memory a commodity over? Or has the beginning of memory as a strategic AI requirement? Are we here? So let's flip the coin. I saw Big Dan there. He got this. So Daniel, you are saying the era of memory as a commodity is over?

Speaker B: It is over. This cycle is different. Period, period, period, period. Rocket ship memory is strategic infrastructure, Pat. Look, the commodity cycle didn't get extended, it got replaced. Memory moved from critical path of the AI build out. And now it is the rules that now the rules that have governed it for 30 years. The boom and bust cycles of memory, unfortunately my friend, no longer apply. And look, look no further than the 16 multi year agreements. 22 billion committed, volume allocated book through 2027. Commodities don't get bought on multi year take or pay contracts. Infrastructure does. The buyers are revealing what they believe and uh, supply is scarce, strategic, and it's worth locking in years in advance. And margins, Margins are the other thing they tell you. You have a moat. Margins aren't going down, they aren't even stabilizing. Margins are going up 84.9%. That's higher than Nvidia. A true commodity can't hold that. Um, you know, because what happens is capital floods in, they build more capacity and then it crushes it. But the problem is this is an oligopoly, Pat. It's a triopoly as you called it. Um, we have yield problems, stacking problems, packaging challenges, and that only three companies can solve this and no one else can. And that right now is a technology barrier, not a pricing accident. And then the biggest thing is the demand driver has changed under the asset. Prior cycles were PCs, phone servers, but a lot of it was consumer discretionary. It was sentiment driven. Pretty elastic. Uh, this cycle runs on AI compute. Uh, HBM is the binding constraint on every accelerator chipped. Memory is no longer riding the cycle, it's gating it. So supply can't physically respond to the old timeline leading edge HBM capacity takes years to stand up. Management puts supply tightness past 2027, we have it past 2030. And so when supply can't answer demand for years, which by the way demand has been underestimated every part of this cycle, guess what? The well known memory cycle has been disabled for the duration. Uh the anthropic agreements are just another exclamation point here of how strategic memory has become. Memory now controls everything. Um, supply constrained technology moded demand inelastic sold on long term contracts. Memory now checks every box of strategic. This print isn't memory getting expensive, it's memory getting repriced as infrastructure.

Speaker A: Daniel I've lived through nine memory cycles. I worked for manufacturers for 21 years. I have crawled through memory factories, PCA, PCB chip, chip flipping, I mean you name it and so first by definition uh, long term agreements, uh, and these new strategic customer agreements, lol, um, very much signal that, that you're a commodity just in a current uh, uh, era of strength. Okay and that's the way that uh, that I read those. Nothing more than I need more uh and I can raise prices on you and I'm going to force you into making long term purchase uh and, and at a certain price, uh agreements that is ah, commodity uh to uh, to the max. Um and the second thing is that architecturally let's look at what's a commodity and what's not. In other words what's industry standard and and what's not. So what defines uh either PINS or micro bumps is an organization called jedic and let me go through the data center uh and PC stuff that is JEDIC compliant meaning anybody can make it DDR5 Cudim, Mr. Dimm, HBM3, HBM3, HBM4, LPDDR5, 5X, LPDDR6 CAM2, LP, CAM2, GDDR6, GDDR7. So these are all things that uh, are essentially I'll call them a uh commodity uh at the pin. And the only reason I brought that up is there were some people talking about uh, that on X. Now where it's going is like we talked about with Qualcomm with hbc. Uh there's nothing standard about that LPDDR that sits underneath uh the hood. Uh and uh, custom HBM4 is obviously not going to be standard, it's going to be custom. Uh and that's the way that, that I would view is, is if it's strategic I think memory is headed that direction for hbm. But, but that's um, but that's, that's, that's pretty much that, that, that's pretty much, that's pretty much, uh, it uh, one final thing here is new supply is coming, right? And I brought this up on Yahoo Finance. China DDR5 production ramps in 2H26. Um, you have Lenovo already using it, uh, you have HP and Dell who are qualifying it and we will see. Daniel, if it's strategic, uh, once China uh, comes online, maybe the US Government uh, comes in to save it. That would be so, uh, so, so dumb. Uh, but you've got, you know, cmxt, uh, coming on and I don't think this is a deep seek moment at all. I, I think this is real. Um, I can't say if it's high quality memory yet, um, because it's stuff's being um, um, qualified. But I do believe that at the end of the day, uh, memory prices uh, will get back to um, more than normal. Um, final point, another final point. Uh, on these uh, SCAs, um, a little fact check on there. Out there, it's not cash flow. Unlike TSMC that takes uh, pre buys on Capex, these customer deposits are not cash flow. This is vendor financing embedded in these agreements. It flatters the optics and how it looks, but it's not the same as a durable, um, uh, free cash flow. And I'll, I'll rest my case.

Speaker B: There. Fell asleep. What happened?

Speaker A: Uh, so much action.

Speaker B: I'm gonna, I'm gonna have to tweet that you're, you're a big time China bull. That is a memory bull for China. It's like all in China. I love it, I love it. What do you really think? Is it a commodity?

Speaker A: Yes, I do.

Speaker B: Okay.

Speaker A: Because there's so many standards around it. If you look at all of the other, um, um, like you look at a GPU like between Nvidia and Nvidia and amd and it's like those are so freaking different. There are six uniquely different accelerator architectures out there and like that's not commodity because they're different. When you're, when you're building to a JEDEC standard, it is by default a, a commodity. You might be cheaper, you might be faster, you may have built more Capex, but um, yeah, it's uh, it's. Maybe we just differ on, on, on the way we look at strategic. All right, good conversation. Let's jump into uh, bulls and bears. So arguably one of the most successful companies in the history of technology, Nvidia, is actually going out to get financing. $25 billion in investment grade bonds. Daniel, why does Nvidia need it? And this is a good sign or a bad sign?

Speaker B: They don't need it. I'm gonna be really clear here. This is raising money when it's cheap, available and uh, when you're in the cash flow situation they're in never hurts. When you're, when your, your balance sheet is gold, when investors are 3,4x oversubscribing this thing, um, they're not raising it to do things like the hyperscalers are. They're literally just putting a monster pile of cash in the in reserve and giving them ultimate flexibility in the long run. So this is a totally different situation than, than Meta or Google or any of these other companies raising. Um, and like I said, best time to raise, best time to go get money is when you don't need it. That's all I got to say about that. That's. This is not, this is not, uh, there's no red flag here.

Speaker A: I can't argue with anything that you've said. I mean, you know, when a company with negative debt, um, is, is the one rushing to borrow, um, you know, and you don't need it. Yeah, you're right. It's a perfect. But then again, why would you do it? Do you do it? Do you buy it? Do you bring it to do more buybacks? Like why do you, what are you going to sit on the cash for? What are you going to do?

Speaker B: Maybe they're concerned about rate increases. Less, less uh, you know, accessible cash long term. Um, there's, like I said, you know, you're just piling on the balance sheet right now. I mean there is some concerns about rates going up, but really, I mean with the type of cash they're bringing in quarter to quarter, they certainly didn't need it. Yeah, maybe it's just a flex like haha, huh. We can raise and no one can challenge us. Of course, their stocks performing like absolute dog shit. So the market clearly does not love Nvidia's long term, which is still really hard for me to understand.

Speaker A: Yeah, I've seen um, even the most ardent, um, Nvidia, uh, stock supporters call it dead money. Right. Like it's just not going anywhere.

Speaker B: It's not. I think, I think the excitement around their earnings is going to quiet. I think we've seen the pivot from everybody caring about Nvidia to everybody caring about Micron.

Speaker A: All right, Daniel. I mean with that uh, 10x forward pe before the big blow up I mean my God, the only reason that it doesn't even go up even higher is because nobody believes that uh, its memory is strategic and uh, the company isn't investing enough in Capex to uh, to uh, make a difference in the future.

Speaker B: So.

Speaker A: But yeah, I mean listen, the shiny, shiny object has to change. Um, and that's especially true uh, with the retail stock market. Hey, let's go to the next topic. Daniel. Oh my gosh, the hundred year flood scenario. Um, memory prices going up, uh, Apple fell five points uh, based on MacBook price hikes. Tim uh, Cook uh, says it's unsustainable right uh, now and, and really, really set that off. And you even had a, the chief Revenue officer at Micron, uh, in a coded method essentially say what I've been saying, which is its paybacks and you know you're going from negative 80% gross margins and your buyers are asking you to lower prices even more. You know, you can't invest into capex. And here we are with what, what was it, 87% gross margin, Daniel?

Speaker B: Micron ended with yeah, uh, 85.9. It, it depends on constant currencies and everything but yeah, it's in the mid-80s now and it's up from 81.

Speaker A: This is you know, Micron telling Apple to suck it. And then you've got um, ah, Apple basically, you know, blaming memory and then you've got Bernie Sanders blaming Apple. Ah, or raising this on consumers. But I do uh, wonder if, if there could be a potential blowback to the memory vendors, uh, based on the consumer uh, prices going up just like we saw with energy prices going up.

Speaker B: Yeah, well listen this, the, the market fell fast when this announcement came out and it was expected but it happened and remember the whole rally got momentarily wiped out and then you know, only Micron was up by the end of the day. A lot of AI because um, you know the market for the first time saw who's going to pay for all this and it's going to be the consumers, you know, the consumer. When the consumer starts to pay, that's when you start to get strain in the economy. Yeah, um, you know, paying 150, $200 more for that phone. That's a lot. Sanders point, which he's an idiot but his point is that Apple still makes a ton of money even at these higher costs. The problem is what he's not thinking about is which shareholders are going to be like yeah sure, take, eat the cost and you know, like go ahead and make less money next year. Apple like we'll still want to be in the stock. Like, they can't do that. They absolutely have to continue to show growth. The market is ruthless. It is, it is absolutely, unapologetically ruthless. So Apple has no choice but to raise the prices. And then you got to see what kind of inelasticity the pricing on apples and iPhones and MacBooks really, really have. Um, so anyways, that's kind of my read on this one. But like, um, the backstory is Apple absolutely took it to Micron when Micron was weak. They made a terrible deal that nearly almost crushed Micron. So what I'm saying is like, I'm sure Sanjay and Submit and the team are kind of laughing into the bank right now because Apple can certainly afford it.

Speaker A: All right, Daniel, let's go into your favorite company and your favorite topic. Micron, uh, blew the doors already.

Speaker B: Like I don't even know that we need to do it, do we? I mean, I guess we can just, just hit the headlines, right? Largest beat in company history. 41.46 billion 346%. Doesn't this remind you of the early Nvidia prints? Doesn't this really remind you the early Nvidia prints? Uh, crushing a 35.69 billion consensus? $25.11 per share earnings, 1,215% year over year growth, um, beating 2049, uh, expectation 84.9% record record margins. 10% higher than Nvidia, by the way. Um, a $50 billion midpoint guide versus a 43 billion, uh, consensus. This thing absolutely ripped. They had the 16 multi year agreements, 22 billion customer contracts. Um, the agreements were all from what they explained, like 100 plus billion dollar agreements over their lifetimes. Um, I think 14 of 16 at least were for sure. Uh, and again only somewhere around a third of the contract was had any sort of ceiling on price. Uh, I think it was maybe up to 40%. So meaning most of these contracts, it's all about just securing supply they need. They don't even have a cost. And that's some of the things that makes it feel uniquely asymmetrical long term for Micron is that historically entering long term agreement should have had some additional benefit beyond just getting capacity. But at this point it doesn't seem to. Um, that's it. I mean we've talked so much about it, I don't really have anything else to say. It was unfreaking believable. But yeah, actually at the same time it was what we expected.

Speaker A: Yeah, it's interesting Daniel. Um, it was an amazing beat. But what's amazing to me is is how they did it. And it was 95 price increases. They had no unit increases uh across the line. So essentially they um, they gouged their customers uh you know and it's paybacks like, like I've said before and you know it goes both ways and, and the buyers of memory, the OEMs and the ODMs are, are now seeing the um, the. The asend um of that. So that's not sustainable uh without these again you know, these, the these SCAs. But this is how you absolutely uh, mentally turn this into a um into a, into a, into a commodity. Let's move to ah, an interesting one. You've got Cerebris that uh had their first earnings since they, since they went ah public, doubled ah revenue year over year. Um you know their, their guide um ended up, excuse me, they beat on revenue, they missed on eps, they beat on, on revenue but really uh, had a huge sell off on the gross margin, uh deterioration. And listen, I just want to say that I'd love for Cerebras to do well as a company because they have a novel approach and the world needs uh, as much competition as the market will allow. But listen, growth stalled sequentially, right? Uh core revenue was 191 million of 12%. The Q2 guide was 194 million. Essentially flat. Um, margins are headed the wrong way. Uh, core gross margins were 47% going to 36 to 38% and 38 to 41% for the year. Um, it's, it's that. That's tough if you're a public company and you're not dominating uh somewhere profitability is, is deteriorating as well. Operating margins went to around 2% uh to a guided um minus. Sorry minus 2 to minus 30 to minus 30. Um, that's, that's kind of big. Uh, uh concentration, uh moved right. It went from uh Core 42 and G42. There was 86 of 25 revenue. Uh, uh to OpenAI who lent Cerebrus a billion dollars. It gets paid quarterly in, in warrants. They're kind of trading concentration uh risk between uh that and then finally on performance fastest. Uh AI is. Is no longer uncontested. You've got Grok bringing up some amazing numbers. You've got TPU8i bringing that up. Uh, you've got tense uh Torrent uh putting up some m. Pretty pretty beefy numbers. So again I'd love to see the company uh Do, Uh, well, it's just, it was, I read it as, as tough earnings.

Speaker B: Yeah, I mean, look, this was one that just ripped absolutely on the idea that um, AI is just has unfettered demand for anything that can be built. Um, there's still a lot of questions about this company, about their product, about utilization. It's got some interesting, another near memory sort of idea, uh, for strong inference. Very little in the wild right now. So, uh, there's just a lot to be seen here. But the initial rip was typical IPO nonsense. And now again, anytime Cathie Wood buys at the top, that's become the Jim Cramer sell signal for me at this point. But, uh, look, I mean they've got a bull case in OpenAI. If they can produce and succeed, that'll be a good case for them. But the margins compressing is uh, is something to be worried about at this point when it feels like, when it feels like margins are going up. But I imagine some of that sits in memory for them too.

Speaker A: Yeah, Cathie Wood is down, uh, 52% on her.

Speaker B: I don't even know how she gets. I don't know how she keeps raising money.

Speaker A: She made a boatload of money and in a prior, prior boom cycle and you know, she made a lot.

Speaker B: Like you've underperformed for how long before people are like, okay, I'm not going to give you any more money to invest.

Speaker A: I don't know if you watch CNBC this morning, but I'm, uh, trying to remember the name of the gentleman who came on and uh, he was interviewed for an hour on, uh, on why he should sell. And he made his name, um, basically during the dot bond. Right. And that was the, the only call he had made, uh, kind of, kind of since then. Um, and who's the guy that made the big short call?

Speaker B: Michael Berry.

Speaker A: Yeah, he's getting buried on, um, on a lot of his, his calls right now. But yeah, it's interesting. All right, guys, great show. Thanks for sticking in there. I uh, appreciate you joining. I hope you guys had a great weekend. This will be coming out, uh, on Monday. Hit that subscribe button. Be part of our community. We'd love to have you, uh, and you know, get on X and tell Daniel and I, uh, where we're off or, or where we did victory laps. Appreciate you. And take.