Title:   The AI Semiconductor Boom and What Could End It with Stacy Rasgon | The Real Eisman Playbook Ep 63
Show:    The Real Eisman Playbook (Steve Eisman)
Guest:   Stacy Rasgon (Bernstein — senior US semiconductor & semicap analyst)
Date:    2026-JUN-08
URL:     https://www.youtube.com/watch?v=imrBSvXepYI
Length:  ~59 min
Note:    Auto-transcript, timestamps mm:ss. Saved for personal study. Rasgon's tour of the AI semi boom:
         AI now drags every sub-sector (memory, semicap, optical, power, CPUs) as investors rotate through
         bottlenecks; the YTD rally is all earnings (multiples compressed); Nvidia/Broadcom lag the
         constraint names and that divergence "has to normalize." Agentic AI is driving a CPU demand wave
         (NVDA ~$20B CPU revenue this year; AMD upgraded). Burry's depreciation bear case rebutted via
         rising old-GPU rental prices. The real long-run bear case is AI returns disappointing; the real
         physical constraint is power ("US has chips but no power, China has power but no chips").

================================================================

(00:05) Hi, this is Steve Eisman. You know, in the great financial crisis, people sometimes ask me, why did I foresee what was going to happen? And the answer in some ways is that the center of the univer universe back then was the entire financial sector and that was my area of expertise. So, I knew what was going on.
(00:23) The financial sector has not been the center of the universe for a very very long time. The center of the universe right now is semiconductor and semiconductor equipment companies because that is where the entire AI infrastructure is being built out. And so today we're going to talk to a recurring guest, Stacy Rasgen of Bernstein, who covers semiconductors and semiconductor equipment companies.
(00:49) And we're going to discuss what's going on, why it's going on, and what could possibly derail it. And afterwards, I'll be back with some closing thoughts. >> [music] >> Hey, this is Steve Eisman and welcome to another episode of The Real Eyesman Playbook. And today we're going to discuss a group that I would say is the center of the universe right now, which is semiconductors and semiconductor equipment companies.
(01:17) We're going to talk about it with a recurring guest now, Stacy Rasgen of Bernstein, who we spoke to >> maybe seven months ago. >> Seven Seven months ago. Stacey, welcome. >> Good to be here. >> So, Stacey, you are the center of the universe. How does that feel? >> Like, it's semis are always interesting. There's always something going on.
(01:37) Um, >> maybe. So, but you've never been the center of the universe. >> It is nice to be popular. Like, I'll I'll I'll I'll leave it at that. I always joke, >> you know, I I do I do well at at Bernstein. I'm, you know, but it's it's not just my sparkling personality, right? I mean, it is a fact that the group is of great interest both to specialists and generals alike now.
(01:56) And just given the rise of AI, I mean it's it's not just semis, right? AI is sort of dragging everything in in I mean in semis and out of semis along with it right now. Yes. >> It kind of feels like it's the only thing that's supporting everything else that's going on. Um so yeah, it's nice to be popular. >> I'm glad you're happy to be popular.
(02:13) We'll see how long it lasts. But [laughter] so let's take a step back. Give me a summary of of all the stocks that the group is a group like what's been since the last time we met. Yeah. when and when seven months ago, it's not like things weren't good seven months ago. They were pretty hot then, too. >> Yeah.
(02:29) >> What's what's happened in the last seven months? >> Yeah. You bet. I mean, it's gone into overdrive and and and what really what's happened is is like I said, AI has gotten so big, it is now dragging everything along with it. And so, probably 7 months ago, we were looking primarily at at the compute and accelerator names, the NVIDIA and the Broadcoms of the world.
(02:47) Now, it's everything in semiconductors. um whe whether it's you know was the accelerators and now it's it's memory and it's semicap and it's optical and it's power semis and it's CPUs now AI is dragging everything everything is is it it you know what it really is is one at a time all of these different parts of the industry have sort of become the the constraint as AI has gotten bigger and bigger and bigger different things have now become like the bottleneck >> the stocks have actually ripped because investors as you know investors love to
(03:16) play bottlenecks >> right Right. Um and so one at a time we've seen these different kind of um different areas of of the broader semiconductor space take off. And frankly you could have owned anything in the space um and you would have been just fine. Um I think semi now year to date I don't know what the exact numbers now but at least up 60% year to date probably more.
(03:36) And the issue >> Micron's up over well over 100. Oh, and look, you companies like like say a SanDisk or do my I don't cover, my colleague covers, but they just guided to a quarterly EPS that is higher than the stock price was when it went public like 18 months ago. And so this is what I find. Yeah. >> Say it again slowly.
(03:52) >> SanDisk just guided to an EPS number for next quarter. >> Just the quarter >> one quarter. >> I I can't remember. It was 31 or something like that. 31 32 whatever the number was but it is higher than the stock price was when it went public like 18 18 months or two years ago >> that's unbelievable >> unbelievable and so if you look at the run that the space has had yearto date it's actually all earnings multiples broadly if you take like the socks index as a prox the socks index is a broad index of semiconductor companies that
(04:22) people use >> the multiples actually come down a little bit >> it's actually come down a little bit right and and so all of the growth we've seen here today has been earnings right and so if you are worried about sustainability and people people I get it. But the the thesis you would have to articulate is why are earnings unsustainable? Because it's not like and the sector is not cheap, but valuations are it's it's not egregious at all.
(04:41) Like we haven't gotten anywhere near nuts yet, right? >> In terms of valuation. >> In terms of valuation, not not at all. And and and so and >> so give me what's what's like the the PE of Micron right now. >> Oh. Well, so the memory names are probably trading at single-digit PE. Micros and that's because you know look memory is is known to be a very cyclical industry and typically in cyclical industries when earnings >> wait what's the difference between a memory chip and a CPU chip >> different um types of processing
(05:06) different types of applications um different economics in the industries uh CPUs are what what they're a subset of a broader space which is known as logic they tend to use very advanced um manufacturing technologies um and they're used for computation right memory chips um I mean they're to to to for memory for memory they used to store.
(05:26) >> So Micron's a memory chip company. >> Micron is a memory chip. >> So if Micron's a memory chip company, which are your companies are CPU companies? >> Sure. I cover uh well, a lot of them I've got some CPUs and some that that want to be CPU companies. So the traditional CPU companies I have are Intel and AMD. Intel and AMD.
(05:42) >> AMD. >> Um Nvidia, who makes accelerators and GPUs, they actually make a lot of CPUs as well. And they are articulating a CPU story that could be even bigger than the other two right now. I've got other names like Qualcomm for example which is known as to be uh they make um chips for mobile phones mostly.
(05:59) Um they are now trying to get into the data center space and they have CPUs as well. >> Okay. >> So yeah. >> So let's start with the granddaddy of them all Nvidia. >> Sure. So, you know, for my weekly rap, I went over the numbers with like a actually I actually went over numbers really carefully and I I would say if I mean you'll give me more details, but I would say if you could boil down the whole story to like a sentence, it would be the revenue growth was 85% and two quarters ago was 65%.
(06:30) It's actually accelerating >> and so it's accelerating and the gross margin a year ago was 60% and it's now 75. >> Be careful. The gross margin a year ago had an impairment in it. So it was too low. It was 75ish or 70 I can't remember 72 or 73 without the impairment. >> Okay. >> So my first question is Nvidia gave some different disclosure this time.
(06:51) >> What was it and what does it reveal? And let's start with that question. >> Okay. You bet. So they did two things. They took their data center segment which they used to split up into compute and networking. Now they're splitting it up into what could you could take as hypers scale versus sort of non-hypers scale.
(07:09) So hypers scale would be the large >> the large >> the Googles and the metas of the world >> the ones that are building the massive data centers >> and then non-hypers scale would be everything else. So the enterprise customers and the neoclouds and the software >> would that be anthropic anthropy there or would they be in hyperscalers? Well, it depends on I think where Anthropic is getting the compute from, right? Because they're Anthropic may be building some and they may be also um going to the Neo clouds, but >> okay.
(07:32) >> Anyways, but um but yeah, so that was one thing they did and they g they kind of because there's this big concern with Nvidia about customer concentration. >> So what they actually said and they showed it is is the hyperskll and non-hyperskller are about equal sized >> in terms of revenue. >> In terms of revenue for data center revenue, they're about 50/50.
(07:47) >> Okay. And and data sentence is what percentage of total revenue? Uh oh d it's 90% right used to be used to be gaming. >> Oh yeah yeah yeah 10 years ago was almost >> get to my second change in the disclosure but but 10 years ago was gaming and there was crypto and all that nonsense >> right.
(08:05) Yeah but um but they're giving us some visibility and saying you they still have concentrated customers but they also do have they're not quite as concentrated as maybe you might think. >> They do have a long tale of of other customers and and and those are also growing very rapidly. just as rap almost just as rap >> the data center business is basically divided equally between hyperscalers and and everybody else >> last quarter it was equal they they bounce around a little bit quarter but but it's pretty big >> okay >> second is they took all of their other
(08:30) segments which is gaming professional visualization like workstation stuff automotive and and this other bucket um and they lumped those all into a single segment which they're now calling I think it was edge computing or edge AI >> and that's a small percentage of the total revenue >> pretty small yeah so they're telling you two thing what one well the biggest thing is they're telling None of that stuff really matters anymore. Right.
(08:49) Right. But the other I I think it gets to the the longer term. They they've talked about >> longerterm drivers, physical AI and robotics and and and even the autonomous driving. And so I think they're expecting hopefully if we're looking out, you know, 5 years, 10 years, maybe that other bucket will get bigger as some of that other stuff takes up.
(09:08) But I don't think it'll be gaming. It'll be things like robotics and and and automotive hopefully. But but they're taking all the other segments and lumping them together. They did not move stuff in between the segments or anything like that. Sometimes when companies resegment, they play games and they move stuff from one segment to the other and it's hard to They didn't do any of that.
(09:24) They just took the big one and split it up and they took the other smaller ones and shoved them together. >> I see. Before we move on, just define for us the difference between GPU and CPU. >> Sure. So, they're both logic. They both use advanced transistors. Um they and they both do processing, but they process differently.
(09:40) Um I'm going to grossly simplify, but CPUs, >> please do. You can think about um CPUs as doing computations sort of serially like one after the other, right? Um you can think about a GPU as doing computations in parallel. So for example, a lot of these um compute chips have compute different a certain number of compute cores on them like they've got different pockets of transistors on the chip that can handle logical operations.
(10:06) And a CPU might have anywhere from, you know, a few cores to a few hundred cores on it, depending on what that thing is is is being used for, a laptop chip or a server chip or whatever. GPUs would typically have thousands of cores on them. Um, smaller each core would be less performant than a CPU core, but there's a lot more of them.
(10:24) And the GPUs tend to do certain types of math exceedingly well, but not as useful for like the general purpose math that a CPU would would would do. Um GPUs are tend to be used I this is why they were used in in gaming and other things. They tend to be used for um uh uh it's called matrix manipulation matrix multiply and addition and I don't want to go into what those are but it's the type of of compute operations that were very useful for graphics. Okay.
(10:50) >> And as it turns out are actually exceedingly useful for artificial intelligence and machine learning. It's the same kind of math. That's why GPUs which are developed for graphics turned out to be very useful for for artificial intelligence applications. >> Okay. So quick question on Nvidia's stock since the the entire story basic I mean if you look at you know CPUs, memory chips, semicap equipment, the entire business basically hinges on Nvidia.
(11:19) In other words, if Nvidia is growing revenue 85%. >> Everybody else is going to do great. >> Yeah. And then we could have a discussion about who's who who's doing better here or there, you know, what's pricing. You know, you you get get into the weeds. If tomorrow Nvidia announces that revenue goes from 85% to 120%, everything goes up.
(11:41) If it announces that revenue growth going from 85% to 40, everything's going down. So my question is, why is Nvidia only up 14% this year and sells on a multiple that's much lower than a lot of these other companies? whose entire businesses hinge on them. >> It's a great question. It gets to what I said earlier about >> and I'm very upset about it because I own Nvidia and I don't get it.
(12:02) >> Yeah. I mean, try not to be upset, but you're right. >> I I'm not that upset. >> It It's lagged. Um now, to be fair, it's up I I don't know what it is. >> It's up a crazy percentage. Whatever. You know, that's talking about this year. >> So, there's a few reasons. So, that that is one. It had a big run already. Fine.
(12:18) Um, secondly though, and it it gets back to what I said earlier where I said AI was sort of dragging all of these other segments along. Um, and investors have been rather than playing like the GPU or the compute names, they've been playing the constraints, right? And again, remember what I said for just to pick on memory for example, some of the earnings revisions we've seen in in some of the memory names, um, >> where you've gone up an order of magnitude or or or even more in terms of the earnings power in the stocks, you're
(12:43) just not >> because pricing has gone crazy. >> Exactly. you're not seeing that kind of a of a thing from from Nvidia. And so the investors who love to play constraints have have been playing the constraint names and and it's they've been going from one to the other like like I said memory to semicap to optical to power right >> to to CPUs.
(12:59) So that is part and and it interesting because it brings up this very interesting divergence between the two because one of them has to be wrong to your point. The other stuff cannot work if >> doesn't work impossible. So >> I that's what I I think that there is an opportunity because I I do think that that has to normalize one way or the other.
(13:17) Either the constraints are going to go down or Nvidia I think has to come up. The valuations I think have to normalize >> but there's no constraint on GPUs right now. >> Well there there is and there isn't. So there there's leading edge logic and and what's called coas the which is the packaging technology and we can talk about that if you want to put the chips together.
(13:33) That's always tight, but Nvidia and memory is is tight, but Nvidia has been very good at securing supply across the value chain. They they they saw this coming, right? Um so if there's anybody out there that has enough >> J is an excellent CEO. >> Oh, yeah. Oh. Oh, yes. Oh, yes, he is. Um so they've been very good at at at securing the supply that they need to to meet the growth so they can accelerate.
(13:52) >> Mean supply from like Taiwan semiconductor. Well, Ty and also the memory guys and also the the packaging and now we know he's he's getting into the optical and and scing like lasers and all kinds of other stuff, right? Um but he's been very good at at at doing that. So they have supply. Um but this divergence has been very interesting.
(14:09) That's that's another reason it hasn't worked as well. And I think the third is just I mean look I can't remember what the market cap is. Is it 5 trillion 6 trillion now? >> It's over 5 trillion. >> It's it's big, right? And it's you know it's 8% of the S&P or something. And so for some of the large like especially the large longies it's hard for them to own more of it because they're already there.
(14:27) It's even hard to be a market weight because it's so big >> and so that's you know it's like okay it's 5 trillion is it going to go to you know >> posing just so our viewers understand if Nvidia is at 8% of the S&P there are plenty of institutions out there >> who have rules that say you can't do anything >> more than more than x%.
(14:48) So they can't they have to tech so by their own rules they have to underweight Nvidia. >> Yeah. Okay. >> So that there so there's some technical reasons I think as well. >> At the same time look I mean the thesis on it has been pretty simple like number go up right. [laughter] It hasn't you haven't needed any more than that.
(15:03) Um and I think it started to work a little bit better the last few weeks. Part of the reason there is you know people get very excited about the CPU names and people realize oh wait a minute they sell CPUs too. And then on the last earnings call they were trying to articulate they could have been a little clearer about I think we'll hear more at competit.
(15:21) They were articulating a CPU thesis and I >> what's their CPU thesis? >> Oh so so he was articulating a a CPU opportunity. They're going to do they think $20 billion worth of CPU revenues this year. So that's a basically all in the second half.$20 billion is about as big as Intel and AMD's CPU businesses. >> And why are they going to be selling so many CPUs? >> So two reasons and this is this is one reason.
(15:42) Now, why are they going to sell them, not somebody else? >> Well, other others will as well. So, there's two things that they're selling. One one is in in the large GPU racks that they sell. >> So, their their current mainstream product is something called a Grace Black. Well, it's like a GB300 NVL72. They have all these, but it's a big rack and it's got 72 GPUs and >> 72 >> GPUs and it has [clears throat] 36 CPUs.
(16:04) Now, these are Nvidia's own CPU design. It's called a Grace CPU. It's based on the ARM architecture. We could talk about that if you want, but every GPU rack they sell has these CPUs. Okay, so that's one thing. However, >> there is a much bigger demand for CPUs now just in general. And the reason is the rise of something called agentic AI.
(16:22) So this is over the last couple years, the the big push in AI has been what's known as generative AI. So >> LLM >> LLMs and you know I'm I AI slop like I ping the thing and it makes a photo or a video or like what it's generating content >> which is interesting and nice but ultimately it's not >> aentic AI would be I want to go do something I want to I want to go I want to plan a trip to Paris.
(16:46) >> Exactly. Or >> go book a do give me the whole itinerary. >> Yeah. Or or you know what's where you use now is for coding like I want to write an app to go do something like that. And [clears throat] the thing is like when when I'm generating AI that's that's running on the on the GPU, right? When I'm doing something agentic, I've got an actual an agent that is the model is actually like like creating an agent to go out and do a task and most likely that task is a real world task and it's running on a CPU. And so just to give
(17:14) you to use your >> why would a Gentic AI run more on a CPU? >> So to do your your travel, >> right, my travel to pass. >> Yeah. So I I ask my my model, I want plan me a trip to Paris. I want to go on these dates. I want you to look for pricing of tickets between these range and I want to stay in four-star hotels in these cities.
(17:32) So what is that? And and and it's going to have an agent that's going to orchestrate this. So that agent is going to go spin up a bunch of sub aents and each of those sub agents, they're going to go ping, you know, American Airlines and and Delta and United and whatever and all the different hotels. >> There's a lot of computation going on.
(17:48) >> A lot of it. All of those things are going to be running on on CPUs, right? >> Why not GPUs? >> Well, there's a GPU that that's running all this as well, right? But American Airlines server is not running on a on a GPU. Like when it's actually going to do the physical task to make the traveler agent to look at that's running on CPU.
(18:05) Another example, I'm using Claude to to make an app and I'm I'm coding something and I tell I want to do this. So that agent is going to go spin up 20 sub agents. >> The agents can spin up agents. Each of those sub engines is going to open up a a virtual machine like basically a representative um uh uh computing environment that's running on a CPU or more likely on on a on a compute core on a CPU someplace and each of those sub aents is going to be writing a block of code and there'll be other sub aents going to put all the
(18:32) stuff together and and orchestrate it and review it and everything and so that one GPU with that one test can be spinning up a ton of CPU um uh compute while it's running. The way people tend to think about this, they tend to think about it in terms of attach rate. So, for example, in in the old style, if you go back a couple of years, the CPU attached was like one to date.
(18:53) I had like, you know, I had like eight GPUs and one CPU in in this box in a server in these and then it went to a 4:1. And then if you look in these big GPU racks in Nvidia cells, it's 2:1. A lot of people talk about, oh, it's going to with with agents, it's going to go to one or even >> so you'll need more CPUs per GPU.
(19:11) By the way, I think that the the attach rate model is wrong, but it's useful, right? It's not I don't think that these CPUs are necessarily directly attached to the GPUs, >> but the GPU when when it's orchestrating these kinds of agentic tasks, it's it's going to require the usage of a lot of CPUs.
(19:29) So, the attach rate model like like all models are wrong, some are useful. I think that's a wrong model, but it's maybe a useful model. We need more CPU content as you as you do real world tasks with these models. So that's one reason why the demand for CPUs is going up. >> And and Nvidia is actually not only are they selling CPUs in their racks, they're actually going to sell standalone CPU racks now.
(19:46) >> Okay, that's interesting. Okay, let's switch gears just for a sec. Hi, Steve Eisman here. A few decades ago, private citizens used to be largely that, private. What's changed? The internet. Think about everything you've browsed, searched for, watched, or tweeted. Now imagine all of that data being crawled, collected, and aggregated by data brokers into a permanent public record.
(20:10) Your record, having your private life exposed for others to see, was once something only celebrities worried about. But in an era where everyone is online, everyone is a public figure. To keep my data private when I go online, I turn to ExpressVPN. One of the easiest ways for data brokers to track you is through your devices unique IP address, which also reveals information about your location.
(20:38) With ExpressVPN, your IP address is hidden. That makes it much more difficult for data brokers to monitor, track, and monetize your private online activity. And now, lowest price ever. Plans start at $3.49 a month. That's only 12 cents a day. ExpressVPN also encrypts 100% of your network traffic to keep your data safe from hackers when you're on public Wi-Fi.
(21:02) ExpressVPN works on all your devices, phone, laptop, tablet, you name it. Just tap one button to turn it on and you're protected. It's that easy. Select plans include Identity Defender, a new suite of tools to get your data removed from data brokers, alert you when your data appears on the dark web, and even ensure you against data theft for up to $1 million.
(21:24) A few weeks ago, I was traveling on a business trip, and I used ExpressVPN throughout the trip. I knew that my data was protected. Secure your online data today by visiting ExpressVPN.com/isman. That's e x psvpn.com/isman to find out how you can get up to four extra months. ExpressVPN.com/isman. If you've been a real Eisman playbook follower for a while, you know that I recently finished treatment for cancer and I feel very lucky to be in recovery.
(21:57) A cancer diagnosis is a lightning bolt I did not see coming. Breast cancer as a man was definitely not in my life's bingo card. As a podcaster and a cancer patient, I have learned that too many people affected by cancer don't have access to the care they need. This is why I feel very strongly about the most recent sponsor for our show, Color Health.
(22:24) Color has created the first ASCO certified virtual cancer center of excellence that is available all over the nation. Certification by the highly esteemed American Society for Clinical Oncology matters a lot. Color's goal? Reinvent how we all manage cancer. From screening to diagnosis through treatment and all the way back to the life you thought you lost.
(22:46) You no longer have to live near the world's top cancer centers. Color has created the first cancer center of excellence that is available all over the nation. The best possible care for everyone everywhere. Color works with employers, unions, and health plans to lead the way through the human, clinical, and financial impact of cancer.
(23:06) You've all seen some of my recent episodes about the state of American healthcare. Getting the care you need is too often unaffordable. Almost half of all cancer patients lose their entire life savings within 2 years of being diagnosed. Catching cancer early saves lives and money. This is why Color's work is so important. The facts are that Color delivers the speed needed to get from suspecting cancer to treating a cancer 66% faster than the average in-person clinic visit.
(23:36) As a virtual cancer clinic, it's accessible across all 50 states, 247, working with local providers. If your rising cancer costs for your company feel out of control, you have to check out Color because there is a better way to get strong member use, better health outcomes, and real control of cancer costs.
(23:59) Learn more about Color as a partner to overhaul your cancer strategy at color.com/isman. And while you're at it, get screened. So, Michael Bur many many months ago um put out a thesis about not per se Nvidia but but about all the hyperscalers arguing that all the hyperscalers had lengthened their depreciation from like three or four years to five to six years and >> mathematically speaking >> that increases obviously earnings that you can't argue with that that's a fact >> and he thought that that was illegitimate because Because the speed
(24:39) because because the GPU because you keep hearing from from Jensen, I got a new I got a new GP every six months I got a new GPU. So what does that how what could that possibly say about the old GPUs? They're obsolete. That was the argument. What do you make of that? >> Um that was the argument.
(24:53) I do not think it's it's correct. Um I although I understand the logic and I'm not going to knock on on Bur, you know, fine, but I don't think it's it's accurate in in the current environment. In fact, you know, you can look at the this is data that exists like you can look at the rental prices of GPUs by generation, >> okay? >> And take take the older generations, the Hoppers, which which Hopper came out in in 2022.
(25:15) You can even look at amp ampers, which is the generation, the Nvidia generation before Hopper. And by the way, just to level set you, the current generation is the current generation that's about to come out this year is called Reuben, the Ruben. >> The current one that's shipping is called Blackwell, >> right? >> The one that came before that was Hopper 2022.
(25:30) The one that came before that was Ampear 2020. >> Okay. >> Okay. And we can look at the rental prices because these GPUs are out there, >> right? >> They're available if you want to rent them. Rental prices for even for the older stuff are going up, not down, >> right? And and even for the stuff that's fully depreciated, they're they're very profitable to use.
(25:46) So these GPUs do not disintegrate after 3 years, right? They are still usable. So the question is, is it economically viable to continue using them when you have better GPUs that are out there? And at least at the moment it is absolutely economically viable to use and you can see it in the data like it it's I think it there's there's no question um the rental prices are going up not down why is that because demand is off the charts and demand is so strong we need every bit of compute that we can possibly get right now and >> even old comput
(26:13) >> even whatever is there and and that's been the interesting thing is more and more computers with with every generation you get more compute out of it right so you've had exponentially more compute coming online and yet the demand has been accelerating even more than than that. And so we need every bit of compute that we can have and the stuff is getting monetized no problem.
(26:30) If we run into a scenario where demand falls off and I don't need that old comput then I worry about lifetime and depreciation but we're screwed anyway so like who cares, right? I do not think that that is a there if we want to articulate viable bare cases on AI we can. I do not think that that the depreciation point is a is a viable bear case.
(26:47) >> Okay, we're going to come to the potential bare cases later. Let's keep going down some of the companies. Um, let's talk about AMD. >> Yeah, look, that stock has done great. So, so I can't argue with it, >> but AMD is a smaller company >> than than Nvidia. >> Almost every company is small. >> Almost every company is small, but quite a bit smaller.
(27:09) I I can't remember what AMD's revenue growth was in the first quarter. >> Oh, it was pretty strong. >> How strong? I can't say. I don't think it was >> I got I I'm not sure. I'm not sure because they've got other businesses. So just articulate the AMD. You're recommending it. >> We just upgraded it. >> I saw you.
(27:26) >> To be fair, I will say that my track record with AMD bull cases is horrific. So we However, >> we all have our crosses. >> We do. I've been much more wrong by not recommending it. I mean, clearly like we we've been we upgraded on on earnings. Um, and I'm actually kind of kicking myself because when we had previewed the quarter a week and a half before they were, I came this close to upgrading it into the print and I I chickenened out.
(27:48) I lost my nerves. So like, no guts, no glory. Okay, >> I'll tell you the reasons we were looking to upgrade and then the reasons why we did though. >> Okay, >> into the print >> again. People were getting very excited about this this agentic CPU >> story and they have very good CPUs and not not only is there is their data center their CPU business growing a lot.
(28:05) They actually have really good products. um they're taking a ton of share from Intel and and just just to give you an example for how strong CPU demand is right now. Intel had about 200 basis points of margin upside last quarter because they were selling previously written off garbage that was like lying around by by their own admission they their their server products are not competitive.
(28:24) >> Customers right now do not care. The stuff was written off. It was lying around. They just they could sell customers like we'll take it. They'd already written off to zero. So they were sold it at 0% cost basis. >> AMD actually has good products that customers want. They've been taking they took out oodles of share.
(28:39) AMD 10 years ago, AMD's market share in x86 servers was.1%. >> Say that again. >> AMD's market share in x86 servers chips 10 years ago in 2015, 2014, 2015.1%. Until 99.9% market share, >> 99% >> this on a revenue basis. Um, >> so basically 100% versus zero. >> Zero. Today AMD's revenue share in x86 servers is like I can't remember low to mid 40s >> and Intel >> uh it's the opposite the 60s but AMD's taking care of course they actually have products that that customers want to buy.
(29:13) Um that's and then the other thing that we were looking at is you know they are selling they also have GPUs >> and they've signed in a couple of big deals OpenAI and Meta um and this buy side I think was there but the sell side for whatever reason had not put Meta in the numbers yet. So I started to look into 27 and and for the first time in a while my numbers were quite a bit above.
(29:30) We we we didn't upgrade. We didn't pull the trigger on it at that point because the street was already modeling this the CPUs up 50% this year and I figured well people must know Meta is coming. Um it looks like now this the server CPUs for them are not going to be up 50% this year. They'll be up 70% like maybe more.
(29:45) >> Okay, >> we'll see. >> And how are they doing in the GPU part of the business? >> They're doing okay. Um a couple of things. So >> because that was the story. It still is part of the story, but you have to remember so the the the GPU thesis on AMD goes something like they're going to go from being a marginal player to a slightly less marginal player.
(30:03) [laughter] They go from 4% market share to 10 or 11 in a market that's growing to a trillion dollars plus >> right >> now to to be fair, you know, to to the bears on AMD to sign these big deals and they signed two of them right now. Open AAI and Meta multi-gawatt multi-billion dollar deals.
(30:20) They they basically had to give away chunks of the company to get there. They they gave warrants. >> Yes. >> Each one was about 10%. If if they're all fully exercised, there are purchase commitments required for the warrants. So, there's also stock price thresholds. Um I will be honest, I would rather see them sign big deals without having to sign the warrants, but I understand the warrants.
(30:39) The reason they they did it is I mean like they had to be on the rocket ship, so they bought their ticket. It was an admission on their part that they the the products as they stand today are not good enough yet to get the kind of share that they need. But they have to build scale rapidly. It's not just enough to have the parts.
(30:54) You have to have the developers and the ecosystems and everything behind you. And it's a chicken in the egg, right? Developers are not going to develop for your product if you don't have anything in the marketplace. So, they're doing everything they can to get the products into the marketplace to build that ecosystem.
(31:07) And I I don't like it, but I understand it. And frankly, if the stock goes to $600, which was the high end of the of the stock price threshold to exercise the warnings, I figured investors are not going to care anyways. And the stock price looks like it's maybe going there. So, they reported earnings. Everything looked good. Our server estimates, which I thought were high, looks like they needed to go higher.
(31:27) Um, and I started running my numbers and it was like, well, now I'm not the street had been at like 11 bucks for next year. I had been at 13 and change. Now I'm at like 14 and change. And if Lisa, the CEO, if Lisa Sue is correct, and and the and the server CPU thing does not level off if it keeps going because she also doubled her estimate for where she thought the CPUs could go.
(31:44) >> Okay. >> She thought by 2030, they had analyst day a couple of months ago. They said, "We think in 2030 it'll be$60 billion TM. Now she thinks 120." If that's true and things don't moderate, they continue to grow, all of a sudden I'm looking at 2028 and I'm pretty close to 20 bucks in earnings, which was their 2030 target.
(32:01) And at 20 bucks in earnings in 28, you can start to underwrite quite a bit of upside. So I admitted it. It's late. Like I get it, but better late than never. Um, and so we we pulled the trigger on it after earnings and upgraded it. >> Let's talk Intel. >> Yes. >> 7 months ago, Intel was here and now it's up here. >> What the hell happened? >> Couple of things.
(32:22) Number one, the CPU story as well is working for them, >> right? >> And I think they're getting bailed out a little bit, but look, you take lucky over good. It's fine. Like I said, their products by their own admission are not competitive, but it doesn't matter right now. So, they're selling stuff that ordinarily they probably would not be selling, but that is helping.
(32:39) It's clearly helping number one. Number two, a lot of the deals that they've signed like with the government and and you know, Nvidia and others took a stake. It got the balance sheet in in place. They actually they had had this um >> deal with private equity in they sold half of their Ireland fab to Apollo for 11 billion.
(32:57) >> And remember, the private equity guys don't work for free. So, it was actually a very earnings dilutive deal for them which they needed to sign at the time. >> Um, however, >> private equity yourself a sweetheart deal. Really? >> Yeah. Yeah. Again, if you're sitting down with private equity, you're probably not walking out with the better side of it.
(33:15) They needed it, you know, few years ago when they signed it. They don't really need it now. And so, their balance sheet is good enough that they were able to buy their way out of it. Again, at a cost. It didn't come for free, right? um they gave Apollo a pretty good return over the three years or whatever that they had it but but they bought their so you got rid of that potential earnings dilution from from that um and then I would just say in general the narrative is going their way so beyond just the CPU thing there's been some incremental uh narrative
(33:39) around the foundry deal the manufacturing right um and they said two things they said um uh their I'm going to talk a little about their process and some of the nomclature they they have different process technologies The current version that they're starting to ramp now is something called 18A and they make a a notebook product on it called Panther Lake.
(33:59) >> Okay. And they've got a next generation process that is in the works which is called 14A. Okay. Um and they said a few things. They said and and but they've had issues ramping the what's called the manufacturing yields like how many good products are you getting out of the fab. They've had issues ramping it.
(34:14) They said the 18 yields which were not great. They said they're they're ramping better than they thought. and they said 14A yields which is in development are better than 18A was at this point. >> Okay, >> now these were very carefully worded statements because I would say 18A may be ramping better than they thought but you can look at their margin guidance and kind of make some guess the yields are still not good clearly.
(34:33) Okay, >> maybe they're >> it's still a company that's not great >> and and 14A you know better than 18 at this point but at this point in its term 18A yields were pretty close to zero. So I mean fine but there were a few other things like you had um uh you know there's been rumors that they may be getting some foundry customers.
(34:47) Apple was mentioned again it will be it will be small >> but people hope that that will lead to something else. And then one thing that I think is a positive for them is there's a lot of demand for this on the AI side for what's called packaging. Again, when you look at an AI chip, it's not a single chip.
(35:03) It's a bunch of chips that are all put together. That putting together part is called packaging and it's difficult and it's actually one of the constraints and Intel actually has decent packaging IP >> and so they may actually get some packaging revenue and so that's all helped. And then finally, like there there's been one overarching bullcase which is why we have not been short the stock even though I've made my career being negative on it, right? It it it was a gift he keeps on giving for like 15 years until like a few months ago,
(35:30) right? >> But I think the overarching that I've been afraid of from the shorts is look, Trump wants the stock to go up. I mean, let's be honest, he took a stake. He tweeted out pictures of himself literally watching a chart of the stock price going up. Right. >> And the right thing to do, frankly, in hindsight, was to buy it as soon as Trump took the stake.
(35:46) >> Right. >> Clearly. So interesting. Okay. >> I would say then they still have a lot of wood to chop. However, >> I do like the CEO. I like Lipu. >> You do? >> I do. Absolutely. Um, he's doing the right things. You know, look, and he's People would say sometimes he's following Pat Pater, the old CEO who got fired, >> right? >> Following Pat's strategy and oh my god, he's getting all the credit for it.
(36:04) I actually think Pat's strategy was not the wrong one. I think Pat's execution of that strategy was horrendous. >> Pat came in and he acted like Polyiana. He said everything's perfect. He started hiring. He blew out the cost structure and then he had to fire everybody, right? I mean, Libu at least came in. >> This is what Pat should have done.
(36:20) Who who comes into a turnaround and act like Polyiana? Like, I don't get it. So you come in, he's underpromise and overd deliver versus the other way around. He he actually had to do a layoff. He got the cost structure in in in place. Um and he's doing what you what you need to do in a in a turnaround.
(36:34) And so I think Pat's strategy was actually very good. I think the execution of it initially was was not great. Um Lipu is executing on on the groundwork that was laid to execute on that strategy better. So they still got a lot of wood to chop. I think a lot is getting priced in at these stock prices, but the narrative is going >> rising tide.
(36:52) the narrative like is the narrative and the things right now are trading very much on narrative. Got it. The narrative is going their way. They'll have an analyst day in the second half sometime and they'll hopefully have something to say. >> What is the difference between I can never keep track and Qualcomm? >> Totally different.
(37:06) I I know they're different, but I can I can never keep it straight in my head. >> They were almost the same. You know, Broadcom tried to buy Qualcomm a few years ago. They went hostile. Yes. >> Actually, he would have succeeded. Um just for idiots. What does Broadcom do? What does Qualcomm? Let me talk Qualcomm first because a little a little simpler.
(37:20) So Qualcomm primarily makes chips for smartphones. Um processors and radios, they're called modems, >> right? >> Um as well as connectivity, Wi-Fi and stuff and and RF and so that's the bulk of their uh 70 70 or 75% of their chip business. Qualcomm also has a an automotive business relatively small but growing and they have what they call IoT.
(37:42) So this is like networking and and industrial stuff and they're also trying to get in other markets. They have a very nent PC business. And then now the big story for Qualcomm is data center, right? So they >> everybody's trying to make >> but how are they getting what product do they have? >> We've got a few products. Um they do have a CPU. >> Okay.
(37:57) >> And they they talked about a win with humane which is the Saudi Arabia consort AI consortium like a year ago. We haven't heard anything yet but we have that. >> They do have um AI racks. They get they have a 200 megawatt deal again with Humane. Okay. >> We haven't seen anything yet. And then they just announced a hypers scale ASIC.
(38:14) ASIC stands for application specific integrated circuit custom chip. >> Okay, >> they got some type of of AI ASIC. We don't know what the part is. We don't know who it's selling to. We don't know how big it is. We don't know when it's coming. We don't know anything except they have a a win. But that >> but that single announcement actually sent the stock up 70%.
(38:32) Tells you how okay >> tells you how nuts things are. >> And they're going to have they have an analyst on on in the middle of June, end of June. We'll hear more about. But that's Qualcomm. Mostly smartphones today trying to diversify away. >> Okay. They also have a licensing business, but I should mention um Qualcomm owns a lot of the cellular IP and other stuff that's out there.
(38:49) So, they get a license on every 3G, 4G, and 5G like smartphone device or and other that are sold in theory whether or not their chips are in it. >> Okay. >> Okay. So, that's a pure profit. Now, let's go to Broadcom. So, Broadcom is a lot of things. Broadcom has a semiconductor business and a software business.
(39:06) And their semiconductor business is AI and non AI. So, let me take these these bits one at a time. And I would say Broadcom historic, let me if I put the AI piece aside for a minute, which that's actually the bulk of we'll be the bulk of the company pretty soon. But if I put that aside for a minute, um Broadcom historically was um uh grower through acquisitions, >> right? >> And historically they did two transformative acquisitions and a bunch of little ones.
(39:30) Broadcom originally, I think it spun out of where did it spun out of HP and Agyant, I think, way way back in the days when they went public in like09. And back then they did, you know, they did RF parts for smartphones and and and some other things. And they they they bought a lot of companies. They bought um LSI Logic which got them into storage and other things.
(39:48) And they bought um PLX and and SCOPtics which got them into optical. And the biggest one back it was called a Vago back then I should say. That's why the ticker is still AVGO today. >> I was wondering. >> Yes. Um but Avago bought what I would call classic Broadcom. [laughter] and classic Broadcom did um wireless connectivity which was Wi-Fi and Bluetooth and GPS and they did storage and they did broadband cable modems and DSL but the the crown jewel was a networking they made chips for switches and routers and that got them a bunch of
(40:18) scales. So if I just look at the nonAI piece of Broadcom's business and but it's a complicated company that's why I'm going to go through but the nonAI piece they've got four or five segments they have wireless which is um uh RF filters for smartphones as well as that classic Broadcom Bluetooth Wi-Fi GPS >> they have storage so they do um um hard drive controllers and SSD controllers and and storage adapters and things like that they have again that broadband business from classic broadcom the cable cable modems and DSL and pawn and
(40:49) everything else. Um, they have a networking business, the the merchant silicon switching and routing. They also do networking custom chips. >> And what's the AI part of the business? >> I get there in a minute. >> Okay. >> And then they they have a small industrial piece. That's the non AI. >> They also have a software business.
(41:04) And the reason here is they had remember I said they tried to buy Qualcomm. When that failed, they started buying software companies instead for a while. And they bought um >> uh CA Technologies, which does a mainframe. They bought um semantic. Semantic that was it. They bought their enterprise security business and then they bought VMware which was the big VMware >> which which got which got them virtualiz and then so before the AI took off they were roughly 60% semi40% software.
(41:30) >> Okay. >> Now we talk about AI um so they do two things. They do networking and they do custom chips >> right >> like Google for example makes what they call TPUs. This is a tensor processing unit. It's Google's own internal custom AI chips. Broadcom effectively works with them to make that chip.
(41:52) So, and they've they've been doing these chips, by the way, for 15 years. It just it wasn't that big until fairly recently. They've been working with Google for for 15 years, but now with AI, it's just taken off. And so, this overall AI business across the uh the custom chips and and the AI networking, they guided for next year for that to be a hundred billion dollars, which is way bigger than the entire company was, you know, a year or two ago.
(42:14) >> Wow. >> Right. Um and they'll by the way they'll probably do a lot better than 100 billion would would be my my guess. Um but that is >> how how well has this stock done lately? >> It's been like Nvidia um it's kind of lagged. >> Why? >> Same reasons as Nvidia. Um people have not wanted to buy the accelerators.
(42:32) Um I also think because they have the software business. Software has been in the toilet, right? >> I've got a colleague of mine that covers it that I mean it looks to me like he's ready to slit his wrists, right? >> [laughter] >> We we a couple we about a month ago I had I had the software analyst from Beard on Rob Oliver who was great and >> the joke that we had was that it's the only group I said this I said it's the only group I've ever seen that goes down on good news bad news and medioc just just news.
(42:58) >> Yeah. And and there's people worry about the rise of of AI and agentic AI. You'll you won't need all these these SAS companies, right? Because you're going to replace it. And and by my guess is I I think for some of them you can argue about what the terminal value of those businesses are.
(43:11) There's probably some babies getting thrown out with bath water as well. And I would put broadcom software business in that. Broadcom doesn't do anything on the it's on the application. It's all infrastructure. >> Okay. >> So this is nobody's replacing the virtualization. The AI runs on the virtualization layer. >> Okay.
(43:26) Let's turn but it got impacted by that as well. >> Got it. Let's turn quickly to ASML, Lamb and Sure. Semicatch. What do these guys do? Explain what they do. >> That so there's a whole separate sector in semis which is the guys that make the equipment that make the tools that make chips. So there's the big five. There's applied materials, >> right? >> Lamb research, KIAC here in the US and then there's ASML in the Netherlands and Tokyo Electron in Japan.
(43:50) Okay, >> those big five have 70% low 70% of the total what's called WF wafer fabrication equipment market. >> They have 70% plus of the WV market. That percentage has kind of been going up over the time and they tended they do different things. Um the AAT and the Lambs and the Tokyo Electrons of the world do I I I should step back.
(44:08) When you're making a chip, you're doing four broad kinds of processes. So these chips are made on on a on a silicon wafer. It's a slice of silicon, leading edge 12 in, 300 millimeters, about that big around. And and what you do is is I do four things. I put stuff on that wafer, >> right? >> I pattern the stuff because I want to define areas where I want stuff to be and areas where I do not want stuff to be.
(44:37) I take stuff away and I monitoring and I watch what I'm doing. I monitor and control the process what I'm doing. And you repeat those things over and over and over and over again. and you build up the different layers of the chip of of the circuitry that make the chips. And if you were to look at a cross-section of a chip, it looks like a layer cake, >> right? >> I've got the transistors at the bottom and I've got, you know, I can have 30, 40, 50 different layers of metal wiring separated by insulating materials to wire all those transistors together. And
(45:02) the features are very small at the bottom and they get bigger as you go up. Okay? But the the companies that make the tools make the processes to do those. And there's different flavors, different materials and different ways to put stuff on the wafer, different ways to take stuff away, right? But they're all versions of of those kinds of things.
(45:19) And so applied materials and lamb research and Tokyo Electron primarily do the put stuff on the wafer and take stuff off the wafer steps. ASML does that patterning step. It's known as lithography. It's the most critical step especially for the advanced semiconductors because >> mean for the GPUs. Well, the GPUs and even the CPUs and everything else because the the most advanced uh chips have the smallest features, right? And it's that patterning step, that lithography step that defines how small of a feature you can print on the wafer.
(45:46) So, ASML is does that almost they have 90% market share. They got almost 100% in the in the most advanced tooling. Companies like KAC, um CLA do that process control that monitor monitor the wafer. um they do in inspection. They look for problems and defects on the wafers. They monitor it while it's running and and there's other companies that do that.
(46:09) ASAT has a process control business. There's smaller companies like Onto and Nova and others. >> So, which of these is doing the best? >> Well, they're all doing good. And and so this is the thing with semicap the correlations are pretty high. And as I say, like if if if it's working, they will all work to greater or lesser degree.
(46:24) And and you could own all of them. You you'd be okay. You could own the basket. There's been divergences. Lamb has probably done I I definitely but year over year Lamb's probably done the best of at least of my three. I cover AAT Lamb and KLA. >> Okay. >> Um Lamb's probably done the best. KLA's probably done the quote unquote worst, but I mean they're all up.
(46:40) It it but I'm going to make up the numbers, but it's like it'll be like Lamb is up 200% year-over-year and KLA's up 100% or something like that. They're they've all done. >> Okay, let's switch gears again. Let's take a step back. >> Yeah. >> I mean, this story is insanely powerful. Yeah, >> let's talk about what could derail this.
(46:57) >> Sure. >> I mean, there are lots of theories out there. I've heard theories of, >> you know, people being basically forced to be token junkies >> and and and the the companies that that they're using are are dramatically underpricing the use of the tokens and eventually they'll have to charge for the tokens and when people get charged for the tokens, they're not going to want to use the tokens as much.
(47:21) That's one theory. You know, another theory is that Open AI is kind of a shell game and that the guy who runs it a liar and and it's going to go public and it's gonna and it's not going to be great. And then >> when they go public, you have to open up the kimono. >> They got to open the kimono. >> Um and and uh they're the kind of the weak sister of the whole story and but they're a big percentage of the whole industry.
(47:46) So if they fail, think things go bad. I mean >> from where you sit >> Yeah. >> I mean you've you've heard it all. I mean, this may go on for the next five, seven years. It's certainly possible. Um, but if it didn't >> Yeah. What would derail it? >> What would derail it? Sure. You You bet. So, I mean, you you have to be thinking about this all the time.
(48:03) >> And look, I've been thinking about it since the day it started. Like, you got to remember like, so Chetch, it's not been that long. Catch BT November of 22. Nvidia started its big run in May of 23. Like, that's And the way it started with >> I'll give you the date, May 25. I looked it up May 25th because it's when they reported >> the numbers that they actually reported were very very good but they but they guided to 11 billion 11 billion so the street >> which was 50% higher than where you were and where your colleagues were at
(48:32) >> the street had been at seven and they guided a little stock was up 24%. >> And I and and I remember and by the way you can't even see that move on the stock price. >> Not now. Yeah. Yeah. But I remember actually looking at that press release and and seeing the 11 and thinking that I must be reading like the wrong line on the now it seems very quaint because where did they just guide >> you started you started to go let me make sure I'm on the right line >> and and to be this for this next quarter they just guided 91 billion just just to
(48:58) give you 11 from from 11 in May of May 2023. >> So in three years they they've almost 10xed the revenue >> right >> and and that that that 11 billion was something that we' never seen before. The title of my note the next day was the big bang. >> Big bang. >> Right. Yeah. Okay. So I that's It hasn't been that long.
(49:15) >> And it hasn't been that long at all. >> And I've been thinking about it since that day, right? What could derail three years? Yeah. So let me talk about like the I mean the thing you would see is probably capex numbers getting cut, right? And and but but by the time you see it, it's too late >> clearly. Yes.
(49:30) You know you the day medic comes out and you know cuts capex or something like it's all over. But we are not seeing anything like that. If anything, the capex numbers keep getting revised higher and higher and higher. And in some sense, at least for the big hyperscalers, it it's not only are I I do actually do think that they're getting a return on this, especially companies like like the metas of the world that have a lot of internal uses for this, you know, fine, but in some sense it's it's also existential for them, right? Meaning, well, everybody's
(49:55) spending. I have to spend because if I don't, they may win and I be I may be out of business. The meta, by the way, the meta problem >> is that >> Google is spending 180 billion this year. And actually, I'm very proud of myself that I actually know these numbers now. And Amazon's spending 220. >> The table stakes in this business have exploded.
(50:14) >> And Meta, poor little Meta, which doesn't >> I'm I'm using Meta just an example. >> No, I'm just I'm not pick on them, but here's the problem is Meta, which doesn't have a data center business, is spending 135 billion. And people are upset because they keep increasing the capex because they can't afford it as much as some of the other guys.
(50:33) >> Yeah. I mean [clears throat] to be fair though this is not 2000201. I mean these are these are real companies. >> Real companies with with the most profitable companies in the history of of man of humankind right also. So um and they're not idiots either. And I may have said this last time I I was here.
(50:48) I you did. Um they're not idiots. Right. So they're not spend I don't I don't believe this. They're they're spending for no no return. They can see things that we cannot see. They're not fools. In some sense, it it sort of is existential, but they they can they can run it for a while. So, like I'm not really worried about capex numbers rolling over.
(51:06) If anything, we're seeing capex continue to go up. And my take has always been capex is too low. I mean, it's funny, you know, Jensen talked about in 2030 we might be doing three trillion plus in in infrastructure spending. And it seemed like a crazy number when he first gave us like a year, two years ago. People were like, is that a cumulative? No, no.
(51:23) We're we're doing pretty close to a trillion dollars this year, right? We're not that far off. So at that level three trillion isn't isn't as much of a stretch anymore as maybe it used to be. But that that would be the the sign like so what would drive that? I mean >> what would drive >> So clearly >> I mean obviously comes down to return >> right but like you said by the time we actually saw meta report I'm cutting my capex numbers it would be too late.
(51:46) It would be but but but the question I would question is why would that happen? >> Yeah. Yeah. So, I mean, ultimately, it's going to come down to return. Like, either they're spending all this money and getting something out of it at the end of the day or they're not. And if it turns out that there's no return, then the whole thing by the there's nowhere to hide, by the way, if this were to happen.
(52:04) If it turns out there's no return on AI, it's all it's all a shell game. It's all everything's coming crumbling down the in semis and out of semis. The only economy right now, >> the whole economy, >> the whole economy. [laughter] I mean last year I I calculated that it was something like if you the AI capex was something like 75% of the growth in in GDP.
(52:25) >> I think for these kind of you have to almost go back to like the build of the railroads like it's like a percentage of GDP to find something that's sort of comparable to what we're seeing. So it's a lot right. So that that would be that would be problematic. So look ultimately like if if it turns out there's no return and but there's a couple ways there's there's no return >> and or or and there's return that's not so great. Yeah.
(52:47) Yeah. So, and there's a couple ways that this could happen. So, one is just by the way just the air pocket. You know, we even before AI, you look at the hyperscalers, they would tend to build and digest and build and digest and build and digest. So, if there's a digestion cycle, which I guess could happen, that would be bad.
(52:59) But if if you thought it was just a digestion, you could probably own through it or look through it, right? The doomsday scenario would be there's there's no return or the return is much smaller than we than we think it is. Only thing you can do right now is monitor proxies. But I I mean, look, you are seeing token usage explode.
(53:14) And maybe you argue, well, they're using too much, but I think they are monetizing. I mean, just as one point example, you can look at Enthropic. Enthropic does periodically release their annual like annualized revenue run rate. >> They've gone vertical, right? So, I mean, the last number they gave, which was a few weeks ago, they were 40$44 billion annualized revenue.
(53:32) >> A month before that, it was 30. In January, it was 14. In December, it was nine. A year ago, it was like a billion or like whatever it was. So they've literally just in the last like few months done that. >> Okay. >> Right. So companies are clearly using them. We we're seeing layoffs and companies are are laying off and spending the money on tokens and in many cases they're now spending more money on the tokens than they're spending on the and to be fair I don't know how many of these layoffs are actually AI driven and
(53:58) how much of it is just some of these companies just stuff themselves full during co and so it's a convenient excuse. I don't know but absolutely we're we're starting to see adoption rates and but it's it's not like broad-based like like adoption. A lot of it is this agentic stuff and specifically for coding which I think is a real use case where there actually is demand and and there's an appetite to pay right so I like and I don't know what form you know AI workloads and usage will take but but agent coding is
(54:26) clearly one where we're seeing like we're we're it's it's it's reached takeoff velocity right >> um and so this whole idea that there's no return on AI I I I don't believe I don't believe it I I'll be fits and starts And you know, we'll we'll see how it what the trajectory looks like, but I think we're already seeing evidence, clear evidence of use cases.
(54:44) >> Yeah. So, let's talk about power for a second. I mean, I mean, some people say that that's the binding on the entire industry. What's your thought here? >> Yeah, you bet. So, if you were to ask me, you asked me earlier like what what could blow it up, right? I mean, if you ask me like let's say the demand is there, again, Jensen says we're going to do three trillion plus in what would stop us from getting there? Assuming the demand is there, it's it's probably power.
(55:07) I mean certainly the US electrical grid is not capable of of adding what would need to be added probably for the demand. And so actually what we're starting to see now at least here in the US is is local like on-site generation. So turbine oh even there like what's the lead time on a turbine is probably three years. And this is where the whole idea >> oh one of the big gas turbines.
(55:24) >> Yeah. It's at least three years. >> So you know we didn't even talk about China but like I I I did a piece of work um not that long ago and the the title of it was something like the US has chips but no power. China has power but no chips. Like who's bringing more capacity on them? And by the way, >> they have all the the >> they can just throw up a coal plant, right? They don't they don't care >> and and and their chips are are not competitive right now.
(55:48) And actually, it's partially because I think the semicap sanctions have been um successful like they can't make fact I know there was some announcements from Huawei over the weekend. They're they're they're you know in general we're forcing China to be creative by the way on on how they make semiconductors because they can't pursue the options.
(56:04) They can't buy ASL. >> Exactly. So they're doing Huawei by the announcement from Huawei, by the way, was I think is not the stuff they're doing is not unknown, but they're pursuing it now earlier than I think the rest of the world because they have no choice. Right. And that that's we're going off topic, but they they've got power. They don't have chips.
(56:22) Um but if they can get chips, they they can power them because they can throw up a coal plant like wherever they want. They've got they got plenty of power. the US and other places, it's much harder to bring the the the the centralized um power capacity online, the grid capacity online. So, they're doing a lot more things like like on-site generation, right, to power these >> small nuclear reactors.
(56:39) >> Yes. Mars and I mean, they even turned on they're turning three Mile Island back on. So, >> but yeah, but power is a huge is a huge controversy, a huge concern. >> Stacy, thank you. >> Yeah. Oh, you bet. >> Great. That was really great. >> And we're back. So, yes, Stacy does cover the center of the universe.
(56:55) There's no question about it. And some closing thoughts are think from seven months ago when I saw him last things are actually accelerating from back then. So for example, Nvidia just reported and they reported 85% revenue growth. But two quarters ago that was 65%. So whatever's going on is accelerating and this rising tide is lifting all boats.
(57:25) Nvidia is up only 14% this year now. It's been a great stock for many many years. So people can't complain. But the stocks that have done the best are the memory stocks and the CPU stocks. And the reason why those have done well is that investors are playing bottlenecks. And the bottlenecks are in CPUs. The bottlenecks are in memory.
(57:45) And there's actually an interesting CPU story in that because of um Agentic AI, the number of CPUs that you need per rack is actually increasing. So the CPU companies have all gone up enormously. But what's interesting is that yes, they've gone up enormously, but the earnings have actually gone up more. So the multiples have actually contracted a little bit.
(58:06) We talked about, you know, what could derail this story. And he, you know, really, I don't think he thinks at this point anything's going to derail it right now. But long term, what would derail it is that the returns that AI creates are disappointing. And if that becomes clear, people will pull back. But I think at least according to Stacy, still too early in the story for that really to be an issue.
(58:34) And so given that capex is increasing, he's still very bullish on the stocks. Thanks for listening. This podcast is forformational purposes only and does not constitute investment advice. A host and guests may hold positions [music] in stocks discussed. Opinions expressed are their own and not recommendations.
(58:55) Please do your own due diligence and consult a licensed financial adviser before making any investment decisions. >> [music]