BG2 with Bill Gurley & Brad Gerstner | NVDA, Chips, AI Compute Build Out, Impact of AI on Big Tech, Earnings & Macro Set-up | E03

26:03I don't think anyone has a Propexity came out. It was a lot faster. That was pretty cool But when I just look at the quality of the answers I don't, I mean, on different searches, one might be better than the other, but I don't see anything that so holistically notable like I remember when Google search came out like where I was I remember trying it versus All -Tovist and yeah, I mean it was like you could tell like oh, this is better, right? And I don't see that right. I did have that feeling of chat GPT versus a traditional Google search. Oh, yeah I mean, you're saying among the answer Now you have you have four or five highly highly back, you know, companies competing in this consumer AI space and there I don't see anyone yet now I as I said on our last podcast.

26:52I think if you get this memory right it could change and since we did that Open AI published a release that says we're working on it. We're working on it. Here comes memory But the promises where I think pretty thin and relative to what you want. Well, the bill explained to people again, because I think this is so important. You and I are in agreement that this could be the next 10x moment with GPT -like experiences for consumers. So, just double click again on what your understanding is of memory and why you think your sensibilities are that it's so important. So, I think there's two elements to this.

27:25One of them is a user expectation thing, and then the second one is a technical observation on the ladder. Let me get to that in a second. But on the user expectation thing, you know, it's funny. I always go back to the movie Herb, which I thought was just incredible. But like, you, I think want to be able to talk to this thing and have it remember everything that you ever told it. And if you had one that knew all of your emails, all of your contacts, could remember your to do this, when you're about to meet someone that could bring up the last four times you met with them and the reminders you left yourself at that point in time.

28:06You're talking, we talk about the programmers, 30 % productivity increase. This could be a human 30 or 40%. If you have this thing in your head that just remembers everything. So that's that and by the way and I said this last time I think most people have just extrapolated AGI into infinity and think it's going to do all these things but it's not doing it right now and you know this because you and I have been talking about this for a couple of months now if you talk to the people that are at the tops of these firms and you say hey why can't this thing remember everything I want and they go oh that's a hard problem.

28:45Exactly. And it turns out that just because of the way this thing works, it would literally have to retrain every night on each, every individual user and training costs are super expensive. Right now it's trained on the internet, it's regurgitating the internet. It's not training on everything in your database. Right. And so there are people, including OpenAI and all, I think every want this is another one of those things. I think they're all aware of this. But they don't know how to do it technically. And, and, and, you know, I invite anyone to come on the show that thinks they know how to do it technically if they'd like to correct this or whatever I'd love or, or if there's a startup that thinks they know how please come see us.

29:28And you and I've talked about this. I mean, and listen, I think open AI, open the communal a little bit. I think they're further out in front than they revealed. But you know, again, I come back to this idea that I'm really lucky out here, my assistant Brit. She's been with me for 15 years. She knows me longitudinally. My likes, my dislikes, my family, everything about my kids, everything about hotels I've stayed at, rooms I want to stay in, et cetera. So my expectation of her is that she can take offload a lot of that because she has all that prior history, right? right? And I'm very human had that exactly.

30:09You think about the productivity unlock for humans. If you give that for free to every human in their pocket, and I'm convinced it's going to happen. But one of the things I was just, I had a really interesting conversation with some friends about Apple. Okay. Because this is the giant, right? This is the thing in everybody's pocket. Nobody's talking about them. But they, they have so much information about me, okay? They have my contact list, they have my emails, they have my text, they have all these applications. And so one of the things I'll just drop out there, I think a little provocative about what they may be doing.

30:43Because I've read a bunch of stuff on Twitter about how they're building their large language model of their own. My sense is they're not doing that, okay? My sense is in fact that if you think about it in the context of my assistant, right? So here's the metaphor I give to Would you rather have an assistant with the intelligence of like Einstein, but they have no access to the internet and they don't know anything about your history? Or would you like to have an assistant that's just above average intelligence knows everything about you and they can use the internet? Okay, you would choose that right?

31:19You would choose the and so think about what Apple is going to do maybe more like a small language model. Like really understand all the language, really understand everything about me, really understand how I interact with all these applications. And then when I have a deep problem I need to solve, they can sub agent it out, right? They can send me down the path of chat GPT or send me down the path of Gemini or send me to meta AI for an answer engine if I want to go down that path. But I think that there's this layer on the top that's just a different architecture, a different way of thinking about this, there's going to be more like my assistant Brit that's just steering everybody in the right directions.

31:58I think Apple is superbly positioned to do this. But of course, you don't want it to just, you also want to be able to tell it things that you just to remember this or mark this down or attach this to a note. And we've talked in the past about how an LLM could be a user interface disruption. And so you could imagine a small business starting with a CRM that is only voice. And you say this customer, this, you know, and you just talk to it and you want it to remember. Right. But that has to be architect. Well, think about this. You know, Brett Taylor's new business, Sierra, we're looking at a bunch of businesses in this space.

32:37Again, you and I are talking about it in the consumer landscape. Remember everything longitudinally about me, but what is a CRM? It's remembering everything longitudinally about your other, your customers. Well, one thing I want to do just to wrap this up because I think that you and I are analysts. And, you know, I think oftentimes in our business, people talk about it, is this company good or is this company bad? And I think one of the things you and I think about a lot is distribution of probabilities and is it reflected in valuation? So if you pull up this chart, we did, which is the main comparison.

33:10It just shows the growth rates and the multiples applied to Microsoft, Amazon, and Vidya and Google. And here all we did was take consensus number. So these are not altimeters numbers. Our numbers are higher for some and lower for others. But one of the things I just want to point out is at the top, this is the 23 through 25 expected growth rates, 14 % for Microsoft, 12 % for Amazon, 42 % for Nvidia, and 11 % for Google. So Google's already expected out of those four to be growing at the slowest rate. But then what's interesting, if you come down here to the price to earning ratio, right? You'll see that Google is trading at the lowest PE, right?

33:5021 times 24 and 20 or 18 times, 20, 25, expected. So all the things that you and I just talked about, Bill, about growth rate and durability, free cash flows into the future, I would argue a lot of these are already discounted in the stock, right? People are already placing those bets. And so one might take the other side of that and say, yeah, Brad, yeah, Bill, I know all those things to be true, but they can cut a lot of costs and do a lot of things. and that could cause the multiple to go up. But if you go to the line under that to the peg ratio, because this is one a lot of people want to ignore on a priced earnings multiple, for example, in videos a lot higher.

34:27But if you actually look at it on a peg ratio, this year it's a much less expensive company. If you look at it on 25 peg ratios, it's just a little bit different. So there are two ways in which to look at future priced earnings multiples, one's growth adjusted, right? that tries to take growth out of the equation and just look at it in terms of strict valuation. So my big takeaway from this bill is we're not here to pick on Google. We're just to say this is an important case study to watch about innovators dilemma. And it's clear to me that investors are already discounting it that they have some of these headwinds coming.

35:05And I think there may be an opportunity. So, you know, I said the other day, if they managed to thread this needle, trust me, I'll climb on board that bus because I think their tremendous costs they can cut out that business. There's a lot of fitness they can drive into that business. And the real question is, how are they going to drive down the costs to serving these inferences and how are they going to monetize this in a way that we said. And by the way, it's funny because I think they have other assets. Like when you talked about Apple, you said they have the handset. Well, you know, Google controls the entire Android market, which is a big market.

35:39Yeah. They have a competitor to Microsoft's Office suite. Now, they have historically not invested a lot in that. It's not a big driver of their revenue. They could. They could all the sudden triple down. You know, they were ahead in type ahead. If you remember, like, my kids use those products and I was always on the Microsoft and I can remember, like it was finishing sentences for my kids. I was like, what was that? That was inside of Gmail. Yeah. first. And so they have assets that they could bring to the bear. And I think, you know, everything you said about Apple is true. Like having the physical control of the physical device seems real to me.

36:22Like meta, like the notion that my AI would live in my WhatsApp as a person, I like that doesn't feel intellectually perfect to me. Like it being in the phone Yes, feels perfect to me. Right. Like this thing's with me all the time. Let's talk about two things in that regard. So we talked about, you talked about memory being a 10X chat GPT moment. So you said GPT was one of these 10X moments to you compared to blue links. If we got memory, that would probably feel 10X like. And by the way, while you're there, I have to say one thing that relates to, to valuation. Yeah. Um, one thing that drives durability going back to our competitor advantage period is switching costs.

37:05If I start relying on one of these things as my memory and, and I don't have a way to pull that out and jump to something else, I'm stuck. Yes. Like I am hooked lock stuck. Right. Which means high multiple very, very positive for the person that gets there. So I think I'm looking for memory as a 10x moment. The other 10x moment I'm looking for here bill is actions. right, going from answers to actions. And so let's talk about that for a second. Yeah. You know, this company rabbit's been making some waves. They have their version one out. Yeah. I saw, I think Tony Fidel tweeted the other day, can't wait to get his hands on one.

37:43There's a bunch of cool demos online. We've spent some time with the company in it. Now, the thing that they have or that they talk about is a huge differentiator is what they call a large action model, not a large language model, large action model. And basically think of it like cursor control bill. So if I say in a fact we did this demo upstairs when they were visiting I said book an Uber going from but it was able to do it it literally had trained on the behavior cursor behavior Of people using these apps and it was able to book that without any other intervention by me. So I Took to doing some research and said could Apple do this because Apple knows exactly what pixel I'm using on the screen to hit a book button on booking .com or on Uber or whatever the case may be.

38:32Now remember Kaparthi talked about this when he went to open AI the first time he worked on a project that he called World of Bits and World of Bits, the iconic thing he tried to do there and I think this was maybe five or six years ago was to book a hotel. Could he get an AI to book a hotel? And he said at the time it was damn near impossible. He had to write all these very specific algorithms had to try to figure out what every booking page looked like. And he said recently on Lex Freeman, maybe a year ago, he said, I think if I tried to do it now, using the general capabilities that exist today, it would be a lot easier.

39:07So I think that Apple's working on this clearly startups like Rabbit are working on it. I think that is another 10X moment that's in front of us, which is we go from answers where I'm just asking it for information to actions. And once it can start booking my hotel, booking, reserving my restaurant, and then I just say, same thing. Do it again, right? Because it has a little bit of memory about my prior action. Those are really powerful. There's an element of this, it's just a fancier version of screen scraping, right? There's a hackiness to this notion. And I have often said, you know, why in the world are we writing in the self -driving world are we riding millions of millions of lines of code to infer the state of a traffic light?

39:57Like, why don't we just broadcast the state of the traffic light and it would be three orders of magnitude less code? But guess what? I think we literally are going to bypass. I think if we had done that, that also would have been intensive, right? Because then we had to wire everything up to be coordinate. Well, here's where I think the world's going. We met with these robotics companies we meet with Tesla, et cetera, imitation learning. Okay. They're not even going to know what the stop sign is or the traffic light is or the dog in the street. They're not going to write C plus plus for every one of those specific incidents.

40:31They're literally going to watch the behavior of the five star human drivers for enough hours and they're going to imitate it. All right, we missed my point back on the on the internet side, which is is telling having the having the AI like move my cursor around and click and fill out things is not the most efficient way to do this. You would have APIs with these different services and a way to interact. That's going to be an interesting evolution. There's a number of startups working on this too on different ways to try and drive action. and to, you know, some of them will sit on top of browsers and do that or some of them might try and sit on top of your phone.

41:19Of course, Google and Apple will stop them from doing that. I totally agree with that. It's funny I was asking our analysts, right? 10 billion queries a day on Google today. I say, do the number of queries in the future go up or go down, right? And I had somebody, if I said, Arvin, at perplexity said to me, well, the number of queries probably goes down because you don't have to ask it so many times. It'll just give you an answer. And I said, what about the positive reflexivity? Once I get the answer, I've got more questions, right? Like as long as it's fast and it's producing that information, I actually think actions and memory will unlock more interactions because it's so much more valuable to me.

41:57I'll start using it more and more for these future things. And I don't know, it'll be interesting to see there, you know, for a while we've had the Alexa's of the world or whatever, you know, do integrations, right? And so the, maybe the possibility existed that if I'm an Uber customer and an open table customer, that eventually I will tell them my favorite LLM front end and they'll come to some agreement there so that they can pass my registration information through and that that all happens seamlessly. But there's a lot of work to do to make all that happen. Right. I think there's a lot of agent to agent interaction that will go on and AI agent representing both of these parties.

42:37But what's interesting about the action model, you know, the hackiness that you talk about, right? I imagine this will get solved by startups in some pretty hacky ways to begin, but then it will ultimately likely be solved at scale in more elegant ways, whether it's APIs or agent to agent interactions, et cetera. But we're starting to see real experimentation and I've had some of the early prototypes of actions actually coming to pass. And that feels to me like the next two big breakthroughs are gonna be this memory. By the way, I said this last time and it's a subtle point, but I don't think Google, you know, it's another issue in the disruption.

43:19I don't think Google has treated its partners well in the search ecosystem. And so there's a lot of of angst there and a lot of mistrust. And so if OpenAI or Proplexity came along and said, would you integrate and pass tokens, they might say yes. I think they're going to be more reluctant to do that with Google. I mean, at a minimum, we know they would probably like more competitors in the game of sending them leads, right? So I mean, I think just the fact that you're a smaller player that you can be another source of competition and they're not so dependent on Google for upstream traffic is probably an advantage to you.

43:59Well, I know we're going to want to move to the topic of chips here in a second, but before Before we get there, we touched on Microsoft, we touched on Google, we touched on Apple, just by way of comparison. And people have heard me talk about meta a little bit in this regard. So again, the way we approach the analysis for all large cap tech, we said, is there existing business get better or worse because of AI and then do their new business opportunities get bigger? Okay. So in the case of meta, unlike Google, Google has this massive super profitable business that's under assault by answers and actions.

44:39In the case of meta, we've seen their core business get better as a result of AI. Why? Because you're targeting videos now on Reels, you're targeting. It happened pre -eal alone. That was already happening. It was starting to happen, but the big difference really, I think, between bite dance and meta was that you mean at bite dance adopted an approach around AI and GPUs before Meta did. And I think Mark really made that transition about three years ago. You can see it in their CapEx spend. But the big question was, obviously he had to spend the money before he got the results. And so investors like us, we're kind of holding our breath and we're saying, would this lead to better engagement?

45:20Well, now we know it's lead led to massively better engagement. And I'm not talking just on Reels. This is on the core big blue Facebook product. This is on WhatsApp. This is on Instagram. So they have these big platforms that are benefiting from both more engagement and the targeting of ads. Remember this stock was at 90 bucks and everybody said Facebook's dead because Apple attacked Apple pushed through their changes that disabled their ability to really track people into it. And basically because of AI, they've been able to backfill that monetization completely. Nobody thought, no investors thought that was going to be the case 18 months ago.

45:57So their core business got a lot stronger. Now, as we look ahead, think about the new business opportunities that are in front of them. And I'm just talking about the things that Mark's talked about on the call. Number one, they've got tens of millions of business customers that now they're literally creating these customer service agents for AI agents for every single WhatsApp business. Now we don't see that as much here in the US, even though WhatsApp is the fastest is growing messaging platform in the US. But if you go to a place like India or Brazil, people are transacting. Some of the biggest AI engines, right?

46:31AI bot companies are being built on WhatsApp as a platform in Brazil and in India where they have tens of millions of customers already using them. So these have become platform companies that are enabling vertical and horizontal bots and they're gonna build their own. They're gonna build it for celebrities. They're gonna build shopping agents that assist me buying things on Instagram. You know, I always see all these things I like on Instagram, but it's a pain to actually buy the things on Instagram. The one click never got that easy. Now I think you're going to see shopping agents that assist in doing that.

47:02And then just think about this content creation bill, whether you're an advertiser, just think what we saw this week was Sora. Text to video. I mean, now think of this in the context of an advertiser trying to drive emotions or a creator. right? So my sons are creators on these platforms. This is going to unleash monster amounts of creation in the world at lower costs. And so all of that, I think, benefits their core business. You have these new businesses that they get to move into that I just mentioned. Then of course, I thought another interesting thing from Morning Brew. I think the pod that Mark was on last week.

47:43He talked about the meta AI glasses that all my analysts have, right? right? He said, you know, most people, they looked at Mark taking the video, reviewing the Vision Pro from his couch and they see, you know, that got a lot of clamor on Twitter. But the fact of the matter is Mark said the way you want to think of VR and AR really is as your desktop or your laptop. But the meta AI glasses, he said think of as your phone. Right? Because I'm going to be able to text, I'm going to be able to call, I'm going to be able to listen to music, I'm going to be able to order my Uber, I'm going to be able to do all these things from those glasses.

48:19And I don't have to pull out this rectangular thing or I keep it in my pocket or whatever. I think that's why you're seeing such incredible demand for those. And of course, the foreign factors will change and it'll evolve over time. But that's an entirely new line of business. So this is a company that's spending $20 billion a year on these other businesses that haven't been generating a lot of return. and I think now the market's starting to assign some value to those businesses. But we should be fair, right? Like, because YouTube benefits from those same dynamics that you talked about. And if you're talking about units of being the phone or the Apple and Google already have a huge install base, like what, forwarders of magnitude to the number of Ray Bangladesh.

49:06Yeah, no, no, no, for sure. But I think the question is from where you are today, right? And so like I'll stipulate YouTube will be a better business in the future. Content targeting will be better, ad targeting will be better, right? And as long as Google is able to backfill the core of search like we just discussed, then it's going to be worth more in the future. There's no doubt about it. And of course, in terms of just their basic research and development around AI, what they did with Gemini 1 .5, I mean, like these, they have incredible talent and resources. The only liability is they have an incumbent business that is a monopoly business with monopoly profit.

49:45So, so that we can move on, let's do a fast drive by, I'm going to do one on Apple and then you do Amazon. Okay. So as a reminder to everybody, just our opinions, not investment advice. So for me, you know, Apple, you, you could argue they have the best asset in the world in and a spoon. And if you look at the user base of this compared to the Android user base, you know, it's just perfect, right? And they've been doing Siri for a while. And so you connect those two things, you say, shit, like they put an L on them on top of this, they could get to all the data. Whether I assure you could give a permission to read your emails.

50:21So you could literally get to all the data. So that's a massive positive. Now the negative is they haven't been known for internet services. you look at Spotify, relative to Amazon music. Series kind of been... It's been terrible. Not evolving, right? Like, it's very much like it was the day it came out. Yeah. And so, it would almost require a pivot of like Mark did on cost. And you'd almost have to see Tim come out and say, we're making a massive pivot. like 180 degrees, we're going to be all in like, like, we're putting our best engineers on this. And until that happens, I think it's a doubters camp.

51:07Yeah. Well, I mean, clearly it's been an underperformer this year. I mostly related to China market. Yeah, you have China market, but you have all these concerns in the market going just as to what's the durability of revenue going to be in the future. You clearly have. But they have to ask that to imagine if you took like five of the top AI people, I mean, I mean, I mean, I mean, and they were there. I'm like, they were there the way Tony Fidel was there, you know, early on for the iPod. Like if you had that, you know, listen, for the first time they have real challengers, whether it's a humane, a rabbit, a met AI glass as these other things, right?

51:44Like, I'm just saying the door has been cracked on the ecosystem. Siri has not evolved, right? So they, I think they're the first to acknowledge that. I think they are going to try to disrupt themselves about that. I'm not sure whether we'll see a big breakthrough moment this year. I think we'll definitely see announcements this year about AI on the edge running on the phone and all these other things. We'll start to see, they'll start to crack the door on this. To me, the big breakthrough on Apple is if they can run a $5, 10 billion parameter model on the phone, on the edge, without consuming all of my battery, which there's a lot of talk that they're going to be able to do that, they can maintain some memory about me.

52:24And then they can show me the early part of actions on this device. It will unlock a huge new device cycle. Okay. And that's what drives this stuff. I was the one supposed to do. You do it. I'm going to ask you quick questions. So on the e -commerce side of the business does AI help her Amazon? Yeah. Okay. And how much? Yeah. I think I think on AI, when I think about retail e -commerce, I think about it from two directions. First is Apple has been in the business of AI from a merchandising perspective, just like Ali Baba has been for a long time. Think about the largest retailer in the world, right?

53:03Think about the way Macy's used to work. There was somebody at the store who would say, we're going to show black t shirts today at the front of the line. And Amazon today, nobody knows why they are targeting Brad Gersner with certain things. It's a black Okay, so they're using it. But here's the thing I think that is happening a bit to them on that front. And by the way, Andy Jassy is getting fit. They are tightening the screws on costs and all the other things. But look at a company like T -Moo. Okay, the Pim Do Do Own China owns that quickly became the largest advertiser on Facebook. Its e -commerce sales are through the roof.

53:37Now what they're doing, I mean, it's so incredibly clever. It's full stack AI. So they don't even have inventory or merchandise. they literally go out and they collect data from customers about what they think they will want. They can assess how many of those things debilt and they literally are building it for themselves. So they vertically integrated this AI e -commerce business and then they're pushing it out the other end. And so I think there have been a lot of people in the US who've been dismissive, but they've been shocked how big that business has become in such a short period of time.

54:09We're starting to see this out of TikTok as well, where they're turning into an e -commerce business. I think this opportunity sits in front of Meta as well. So I think there's some orthogonal challenges. But in terms of the core, I think their core continues to get better because of better targeting and AI reducing costs. Think about their customer care cost. We do have to move on. Hit AWS as quickly as possible. Yeah, so I would say AWS, at the end of the day, these companies are in the business of renting AI services to end enterprises. Right? And as much as we talk about Azure and Microsoft running the table today, here's the truth.

54:47We've seen almost no share shift from AWS to Azure as a result of open AI. And if you would have asked me 12 months ago, I would have said, Jury's out as to whether or not that's going to happen. It didn't happen. Amazon responded quickly enough. And here's the other thing, you know, and Sluitman's talked a lot about this term, data gravity, right? It turns out all my data's in AWS. So long as they have a decent AI solution, I'm going to stick with them because I don't have to move anything. And I think they delivered that solution. I listened to a podcast and Jesse was talking about the fact that they have proprietary chips for both training and inference.

55:21And obviously as the AWS stat grew up, they had moved into networking chips. They moved into a lot of technologies. these people wouldn't have thought about Amazon owning or designing. Do you give them, and this would be a good transition? Do you give them any chance of being competitive in it from AI, silicon perspective? So I think the right way to think about it is not will they build a better GPU than Nvidia? The right way to think about it is can they supplant part of the super computer, right? the entire system are there pieces that they can pull out and plug in or workloads specific workloads that they can serve with a lower cost infrastructure because they're doing hyper -targeting silicon all the way to model.

56:11And I think the answer to that is yes. But I still think they'll be one of the largest buyers of NVIDIA GPUs in the world because it doesn't replace that for a lot of really important work. Well, the thing you also said that I think would be good for the listeners to hear. And, you know, I don't want to overstate, you know, where Alexi is, and we talked about Syria earlier, but he said that as Alexa got bigger, that the training costs are tiny compared to the inference costs. And he suggested, and maybe this is me interpolating, that the inference market over time is going to be much more susceptible to, you know, things that are lower power, lower cost, all the things that aren't just performance from a silicon perspective.

57:01And I totally believe that to be true. And with that, we can move to the biggest headlines of the week. We finally got here, which is this debate over the future of the compute build out needed to support AI. And to your earlier point about valuation, how unique is this revenue? How long does it last? And so we have a couple of charts just to ban or tweets to bang through here to kind of contextualize this first Nvidia stocks up a lot But it's because the revenue of the profits have greatly exceeded expectations So this chart just shows what their data center market share has grown to in the year, right?

57:38The world is shifting toward AI as a compute infrastructure and and and they benefited one of the areas I think I tweeted about this that I think has been greatly underestimated this idea of sovereign demand And I tweeted this week, you know, I think Jensen was over in the Middle East meeting with several of the GCC countries over there. And I think what people still don't appreciate is there probably dozens of sovereigns who are trying to get into the Nvidia order book and that they view it as one of their top three national priorities to build out AI capabilities. And I think the size you're talking about for some of these GCC countries is going to be competitors, competitive with the hyperscalers itself.

58:20So in that context, right? When Sam Altman suggested and blew everybody's mind, he's going to raise $7 trillion to, you know, build chips. I don't know if he ever said it. It was in firm. He has it repeated over and over. Okay. So he threw out a big number. Um, but I do think that we're talking trillions of dollars over the course of the next four to five years as we rebuild the world's computer for structure. And then finally, Masha did not want to be left out. And so he came out and said that he's going to raise $100 billion to build fabs and chips to compete with Nvidia as well. You've watched the semi industry for a long time.

58:59Okay. And more importantly, just the dynamics of supply and demand. So just step back for a second, right? What do you make of all of this. Well, I have some cynicism, but that comes naturally to me. The first thing I would say is they're all talking about raising money from the exact same people. So if I were those people, I would just be a little careful because I think to a certain extent, there's a smell of loose money. That's how I would interpret it. Because they're not saying they want to raise this money absolutely. They're saying they want to raise it from a very specific group in the Middle East.

59:46So that's the one thing too. I was struck when I read about sovereign server stacks, there needs to be a reason, right? It would have to be about wanting to have control over certain amounts of information. it could be proprietary information to your country, could be wanting to control how all of them operate in that country. Servers typically depreciate a bit like fish, you know, and now it's been true of DRAM and storage and all of that. You said fish, right? Fish, okay. They lost a day. Yeah, well, I mean, it's I'm being provocative, I was saying, but people have talked about with with those, like you wouldn't want to hoard any, because what happens is, you know, the next generation comes along and it goes down quickly.

1:00:36So I would just, you know, there was a time at which Microsoft was trying to convince the world that we'd all need an NT server for every employee. And, you know, when I heard that the first time, I was like trying to get my head to the, so I don't know, I don't know if countries need server stacks, maybe like I said, it have to have those particular frames in mind. The second thing that just struck me, and this gets more to the Altman and the Massa quote, is the idea that we're just going to go compete with Nvidia is pretty radical. There are already people competing with Nvidia, AMD's competing with Nvidia.

1:01:19There are other people that have somewhat of a head start like decades. Yeah. So you're just gonna go do it? That's bold. Right. It's not like chip design bends. Oh, yeah, an Intel obviously, but it's not like chip design bends to disruption or like software does. Right. This is hard stuff. Yeah. And then some of them, and once again, I don't know that there was an exact quote, but the idea that you're gonna be able to fab and compete, you're gonna compete with teams at TSMC and Nvidia at the same time, like, no chance. Yeah, like no chance. Cause like, let's say, let's say you got it, I mean, we all know how the math works.

1:02:04Let's say you got a 20 % chance of competing with either of them, like, then you're down to four, like a pool in the soft. And by the way, the time scale that you're gonna need, like, I mean, just read, well, we'll get into it in a minute because we'll talk about like what it means to have a competitive fab around the world. But TSMC is far, far ahead of the competition. And one of the reasons AMD has a higher market cap than Intel today is specifically because they got out of the fab business in bet on TSMC. So I think I think it's really important to tease out those two things, right? There's chip design right now.

1:02:41Obviously Nvidia is already designing for two to three generations ahead. I mean, I mean, the series B is already taking orders in the order book likely to launch in Q3 of this year. And, you know, is order of magnitude better than the H100s that are out there today. And then you're, they're already designing what comes after blackwell. And so it's not as though they're standing still. And anybody who knows Jensen, he's an animal and that, that company is, is playing for the future. And then if you look at TSMC and, and I shared with you a, a video, maybe just pull up a, a little bit about kind of the findings from that, but this is from the founder of TSMC and the CEO for decades, Morse Chang, talking about the competitive advantages in Bill, because this really gets to the fab.

1:03:26Like why are all the world's fabs in Taiwan? Okay. And why aren't the fabs in Texas anymore with Texas instruments or why aren't the fabs in other parts of the world? And what does that mean for the future of this buildout? And I think the implication of these releases is that we're going to start building a bunch of fabs in the Middle East. I think we know we're trying to build fabs in Arizona. There's some talk about building fabs in Mexico, but maybe just let's deconstruct that one. What does it mean to build a fab outside of Taiwan to make next generation AI chips? You shared this link with me and I it's a talk that Morse chain gave at MIT.

1:04:06I think in November, right? Very recently. He's over 90. And the first, it's like an hour long talk in the first 60 % is a history lesson. But then the last 40 % I would encourage everyone to go watch. Like everyone, including every politician in the United States of America. Because Morris makes the point that But the reason Taiwan is competitive has to do with the labor model that exists there. And the type of work people are willing to do and your ability to retain them. And he walks through his history of hiring in Texas and other parts of the US. Yeah. Interesting. And enough he ran a fab plan for Texas instruments in Texas.

1:04:55And he explains why Taiwan, like why Texas could never possibly compete with a fab plan. and he admits that US was a manufacturing prowess in the 50s and 60s, but the social requirements that we put on labor at that time are different than they are today. And so whether you look at TSMC and it turns out it's the same thing as true of a Fox confactory in Mexico. You have people working longer hours, sometimes living in dormitories, and he mentions that and he's tough. And he says the country must more likely to disrupt Taiwan would be Vietnam or India, not in advanced culture. And to think you're going to re -enshore a fab and ignore Morris Chang is just kind of crazy to me.

1:05:53And I look at the requirements once again that we put on companies around labor and say to myself, it's not gonna happen here. And the people will quickly react to that and say, oh, you're in favor of forced labor or like super hard labor. But the people that are choosing that at that point in time are choosing a better life right and to deny them that Opportunity like the the the the the individual that lives in Whereas that's commuting to this fox comp plant is getting a better life that's right with even with his Four nights a week in the dorm right um and to deny him that and insist on our circuit 2023 social norms on that country is unfair from my point of view.

1:06:46Right. And I think and I deny them their chance at disruption. So when you unpack and we'll put the link to the video here, when you unpack that message, it's really that Taiwan thrived because these operators and technicians would spend their life working in the same fab on the. Also getting getting better and better at the same thing. And he talked about a 12 % turn rate. I think when he was at the fab in Texas and he said the problem is the second a better job would come along. They would leave for a better job. He said that was it was 12 during during a recession and applying that was much higher.

1:07:21It was went to 25 when when when times we're good. And he said you just can't run a fab plant with 25 % churn among the operators. You you produce really bad product. And I think the point is it's not just better life. It's also So kind of these cultural norms. And so that's why he said, you know, in Vietnam and India, they have cultural norms, he believes that are more consistent with loyalty and stain was something longer. And on top of that, that it would be an improvement in the quality of life for the people who would take each other. And therefore the incentive is there for them to stay in those positions.

1:07:55And neither right nor wrong. Like, you know, I'm not provocative. Like, yeah, it's provocative because it says is that if America is really worried about the concentration in Taiwan, they should probably be trying to help build semiconductor plants in Mexico or Vietnam or that kind of thing versus bringing them here because the odds of operating them here in a competitive way, globally competitive way or low. So I guess is that make you skeptical of the chips act? I mean, I see the Intel is back in Washington looking for another $10 billion to subsidize the work that they're doing. I mean, I get the US national security concern, particularly considering that 100 % or virtually 100 % of these advanced chips are being manufactured in a place that has risk, has political risk associated with it.

1:08:42Bill, let me ask you this. By the way, I am somewhat skeptical at Chip's second. And then the other thing I would say to you is like China is probably in a really good place. Yeah. They're really smart. They have all the intellectual capability of being competitive and they probably still have this opportunity for them, you know, over time in terms of the willingness of certain part of their population to be willing to work in those types of situations. Yeah, I mean, I'll take probably the under on that. I think that the opportunity. Like now there is a global imperative to diversify the source of manufacturing.

1:09:26And I think Morse Chang was having this conversation at MIT recently because he understands the global imperative. I think you are going to see plants that get built in places like Vietnam and Japan. I think you are going to see them get built in India. You're probably going to see some attempts in the Middle East. Obviously, we're trying to do some of this stuff here. I think from a United States interest, both in terms of wanting to maintain leadership in AI and wanting to diversify our political risk associated with Taiwan, it's not so important that everything is produced in the US, right?

1:09:59That shouldn't be our goal or objective for all the reasons that Morse Chang points out, but I do think it would be better if we had four or five places around the world that were load balancing the manufacturing of these chips. That's a fair point. And I think the whole re -enchoring argument conflates the national security interests with a interest in American jobs and that kind of thing. Going back quickly to these new chip companies that are going to miraculously compete within video, I would tell you, and this is maybe an older venture capitalist talking and one who's watched different partners sit on the boards of startup semiconductor companies.

1:10:47It's ain't easy, you know, and you know, the first silicon that comes back doesn't always work. Yeah, and you might be 50 million to first silicon. You might be a hundred million to first silicon. Right. You've got to get in line at TSMC. Right. How are you going to break SMC time. Right. Right. How are you going to get on that place? Yeah. And it's hard. Yeah. And by the way, once you do get working silicon, your yields are probably crappy. Yeah. That's what happens. Like this is, this is physical material science type stuff. It's not software. Right. And you're going up against again, two companies that are run in pretty exceptional ways by exceptional founders.

1:11:39In the case of Gents has been there for three decades. He's devoted his life to this TSMC seemingly has similar types of leadership. But one of the things I wanted to pivot to on this bill because it begs the question, why is Sam throwing out this really big number? Right? Why is Masha throwing out this really big number? And I think the answer, like one of the things I want to talk about is this, which is just what is the size of the market opportunity that we're talking about here? And so I have a slide we had Jensen when he was in the Middle East. He mentioned and the quote was, and this was just from Feb 24th, he said, there's about a trillion dollars worth of installed base of data centers around the world.

1:12:22And over the course of the next four to five years, we'll have two trillion of data centers powering software around the world and it will all be accelerated compute. Okay. And so I asked my team to break that down a little bit. Like what, you know, how does that look like per year, right, to get to this number? So of course, I'm doing this from outside in. We take a swag at it and it pulls up this next line bar chart. So this is the AI data center build out in blue is the new accelerated compute, right. In green is the replacement data center that we think will go to accelerated. And then in gray is the replacement that's non accelerated compute.

1:13:05So this would be more like, you know, X86. And again, I'm certain this is wrong specifically, but that's what we're in the business of doing, trying to build a forecast based upon folks who are providing this information. The line running through the center that starts at 55 % and goes down to 26%. that is in video's share of that global compute build out based on current consensus numbers for Nvidia. Okay, so the consensus forecast that has the stock at $700 a share assumes if you believe this tam to be accurate assumes that their share will go from 55 % today to 26 % in 2028. Now I think if you just step back and you say okay do we think we're going to go from a trillion to data centers to two trillion to data centers?

1:13:58Just ask that at a high level. Okay. Will you and I just spend an hour plus talking about how a 10 billion queries a day are likely to move from information retrieval right to inference as we as humans expect to get answers rather than a card catalog. We talked about enterprises whether they're doing their engineers in co -generation or whether customer service centers or whether Tesla and full self -driving or whether it's sovereigns who are taking on national security issues, you know, drone fleets or whether it's proteins and life sciences or material sciences, there isn't going to be a single industry on the planet that's not employing accelerated compute in order to solve the problems of their enterprise.

1:14:45So if you said to me, with that is the backdrop. A year ago, I think the big question Bill was, is there going to be enough productivity gains in the world to justify this compute buildout? Remember, people thought, oh, we're pulling forward all the demand for Nvidia. This is like dark fiber in 2000. We're going to be way over supplied. We're going to have a glut. I think the evidence on the field is that that was wrong. I think the evidence on the field is that in fact, just like we talked about on the last pod, we tend to underestimate the size of these super cycles because when you have these phase shifts, everything changes.

1:15:23You have positive reflexivity in the world, more but gets more because it's better, right? And so I think the bigger question when I look at this chart and what I push my team on, what I'm certain of, you know, the rumored pricing of H100s to B100s is that B100s are going to cost even more than the H100s. And so I say to my team, like these margins have to get competed down, right? But the feedback and something I think that's really important is that although the B100 is more expensive, it's so much more powerful, right? It's like this. If you had an employee bill and you were paying them $100 ,000 and I said, hey, you ought to hire this other guy.

1:16:08He costs 200 ,000. You said, well, why would I hire him if he costs 200 ,000? I would say he does 10x the work of the guy who you pay a hundred thousand to you would pay him a two hundred thousand dollars in a second, right? And I think that's why Nvidia today is getting those margins. In the future, I expect that there's going to be more competition, whether it comes from this custom chips that you're talking about, whether it's from other competitors like AMD, whether it's from, you know, new startups that from Mosa or Sam, et cetera. But you and I just talked about the challenge to build a fab and the challenge to design those chips is non -trivial.

1:16:41And, you know, the probabilities are somewhat low. And so it's going to take time to get. And time, I mean, if you're going to take time, like, when would you have an impact? But one question I would have for you on this is, you know, if you write about this, I wonder about TSMC's capacity, right? You know, and that is a limiter. Right. So you're looking at the chart and saying, how do we get to two trillion of replacement in new if TSMC is gated in their ability to produce these? Now, Jensen gave this number. He's intimately familiar with TSMC and their ability to produce. So I said, I think he has a sense in his head about what they can get done.

1:17:23I think that the other limit limiting factor we're going to run up against here pretty quick. It's not going to be capital, right? It may or may not be TSMC, but the power consumption. So even for the B100, the data center designs, like you're talking about liquid, cooling, custom design data centers, they're going to consume massive amounts of power. And I think part of the reason you're hearing about this sovereign demand from the Middle East bill is they understand that their chief competitive advantage is low cost energy, right? And I don't think they're talking about building all of this just to service the needs of their country.

1:18:01I think they're smart enough to understand. They're trying to equitize all their petrochemical well into technology well in the future. And if I was running one of those countries, I would look at this phase shift as an opportunity to become the supplier to the world of of of computer -aided intelligence, right? And if they can do that because they have lower cost energy and because they can recruit the likes of Sam Altman, they can recruit the likes of TSMC to their countries to set up shop there to design chips there. It's not all that different than digging wells, right? Think about digging a well.

1:18:39You have to spend a lot of money, a lot of time, a lot of research. You're hoping you get your payback five, 10, 15 years into the future. And so I think this is a rational decision by them to build out this capability. But to your point, it's a non zero and non trivial undertaking to try to get it done. Now, if they do that, it's going to put them in competition with some of the hyperscalers, right? From core weave, AWS to others who are in the business of renting AI capabilities out. But I think it's good for the world because what we want to see is a lot of competition. We want to see the price of AI compute fall over time.

1:19:18That's going to lead to a lot more consumption. And because of the productivity gains from the end applications, again, self -driving cars or coming up with vaccines or solving complex problems, or just allowing consumers to have answers instead of 10 blue links, we need the cost to come down on all this stuff. If you, and maybe it's a good way to wrap, but if you bring that attitude to the table, I mean, it sounds like, and I don't want to put words in your mouth, but there's no, there's no reasonable insight from where you sit, like on this way. We're very beginning, And it's going to go for a long time.

1:19:56I said last week, and I think we had a little video that went out and I said, we are going to hit a zone of disillusionment. I don't know whether it's this quarter, maybe tomorrow within video when it starts, right? We're going to hit a zone of disillusionment where you have a mismatch between supply and demand and then all the skeptics are going to say, I told you so, the internet's a fad. A -I -I's a fad, mobile's a fad. I get happens in every super cycle. We're going to use that as a buying opportunity because we are absolutely convinced that the runway is longer and wider and the impact on humanity is going to be greater because the end applications are so compelling that are using AI to assist them in everything that they do.

1:20:39But that's really where the tug of war is in the world today, Bill. And that's what makes a market, right? You're going to have those people who are skeptics about that demand. That's what creates a wall of worry. You know, why does Nvidia trade at 20 to 30 times earnings? Right? I would say because there's a lot of skeptics and there's a lot of worry about whether or not these free cash flow is durable into the future. Right? But I bet to the worries more on pricing than volume. Right. And by the way, and it's it's unknowable. Right. I don't know. Nobody who follows this knows. So you have to assign some probability to that future outcome.

1:21:11But I think you're right. It is a good place to leave it. I wish you could be here I know you're you're anchored down there in Texas speaking of Texas fabs But this is fun to have you here in person good to see like old times and And I think we're gonna be talking and debating this For as long as we're doing this pod for sure no doubt