So Nvidia grew revenue 55%. You've got the hyperscalers spending 350 billion, 400 billion. Who can even keep track after a while? Where are we in terms of this story in your view? What ending?
Now I've been doing this job for almost 18 years. I've never seen anything quite like we're seeing right now.
AMD was thought of as the alternative. They said they were going to have a GPU chip. What's happened? What kind of inroads have they made? And I will say I've made my career being negative on Intel. What the hell happened to Intel? Can you just explain to me what the hell happened to this company? At this point, I don't see anything out there that's so wonderful coming out of AI that means the returns are going to be so high.
I think the doomsday scenario would be they're spending all this money and there's no return, right? And if that's the case, then I think the whole thing would come crumbling down frankly for everything else.
And that's what we're here to talk about.
Hi, this is Steve Eisman and welcome to another episode of the Real Eisman Playbook. And today I'm interviewing Stacy Rasgon, who is the chip analyst and semiconductor equipment analyst at Bernstein. Welcome, Stacy.
So Stacy, not much going on in your group.
Boring. Nobody really cares, but let's see if we can mine a couple of nuggets out of what you do. So I do this with a lot of guests. Let's imagine that we're at a cocktail party. I'm someone who watches CNBC, kind of knows what's going on, but I'm no expert. And we get introduced and I find out what you do, and I think, 'Oh, this is a great opportunity to learn.' So Stacy, give me a five-minute dissertation summary about the boring aspects of the stocks that you cover.
You bet. You bet. So again, I am Stacy Rasgon. I'm a managing director and senior analyst at Bernstein Research where I do cover the US semiconductor and semiconductor capital equipment space. Now, I've been doing this job for almost 18 years. I started this in April of 2008, about three weeks after Bear Stearns failed.
I've never seen anything quite like we're seeing right now. And it really is a tale of two cities in some sense. We've got the AI trade and the AI stocks, the AI numbers that are just ripping, right? And we've got everything else that, to be honest, is kind of lackluster. And it would be very interesting to see what would be happening to this space as well as maybe the broader situation if we were not in the middle of what is a very clear AI boom, because it really is AI that is driving everything right now.
Yeah, it really is. And just to give you, you know, just to step back a few years, remember it's only been a couple of years since this really started. Clearly started with Nvidia, right? ChatGPT was sort of the lightning rod for all this. ChatGPT was released in November of 2022. It wasn't that long ago. And Nvidia really started their run in around May of 2023. They had an earnings call.
Do you remember that one?
Actually, absolutely. But this was the reaction of the analyst community.
Yeah. I mean, the title of my earnings recap note was 'The Big Bang' and it really was the first quarter 2023.
I can't remember which. Yeah, they have a weird fiscal year. They always report late.
Yeah, they have a weird fiscal year. But whatever quarter it was, I just remember the consensus for the following quarter was something like 7 billion in revenue and they got 11. And today, by the way, 11 would seem quaint, right? But at the time, I had to look at the release like twice to make sure I wasn't looking at the wrong number. Like it was a different company. It was absolutely insane. And I mean, they're doing, you know, 50 billion a quarter now, like not 11. So even...
But by the way, before you continue, the one statistic that I always like to say is if you knew nothing and you looked at Nvidia's most recent quarter and you just looked at the numbers and you said to yourself the following three sentences: this is the largest market cap company in the United States, it's a market cap of over 4 trillion, and they grew revenue 55%. And then you just stop to think about that just for a second, like, wait a minute, the largest company in the United States, this is not a billion dollar market cap company. This is a four trillion market cap company and it grew revenue 55%. You have to take a step back and say, and you can't even grasp it, which is what has to happen for that to happen. And that's what we're going to talk about.
It's really amazing. Again, I've never seen anything quite like this. And that's the thing. It hasn't slowed, right? These were numbers that when it started seemed unbelievable, right? And it's just been going from there. But it would have been even bigger. You have to remember they got cut off from China, right? It would have been even more. And we're seeing other names catching similar, maybe not to the same magnitude, but similar types. Broadcom, for example, we can talk about the whole GPU versus ASIC, but they're starting to see a similar type of acceleration. And there's a lot of other names on the periphery, you know, cooling and power and memory, and even in the semi space that are doing well. So that's all doing very, very well. On the other end, you can take the more traditional semiconductor analog space. For example, you have your Texas Instruments and your NXP's and your Microchips and your ADIs, your Analog Devices, that actually were super, super strong during COVID. Like we haven't even talked about that. There was like a massive...
Everybody bought their laptop that needed all those chips.
So there was a massive overbuild during COVID that some parts of the market are still to this day trying to work off. And we're several years past the peak of COVID. The analog names are sort of in that category. They're well off the peak. There's some hope of sort of a cyclical bottom and recovering. We're still waiting to see that.
That's not a growth story.
It has not been. We can talk about, I've been a little more lukewarm on my analogs. We can talk about that. But that part of the semi industry, at least from a revenue standpoint, has been doing, I'd say, less well, right? And then you've always got your pockets here and there. But without the AI story, things would be in a very different place right now. But you can look at the overall semi group. I mean, the SOX, which is sort of the broader semiconductor index, at least as of Friday, had been outperforming the S&P year to date by almost 1400 basis points. This group collectively has been doing very well, but it really is on the back of these AI names that's sort of lifting the...
So that's the rising tide.
Where are we? Think of it this way. Okay, so Nvidia grew revenue 55%. There would have been more if there was China, but you've got the hyperscalers spending 350 billion now, Oracle 400 billion. Who can even keep track after a while? Where are we in terms of the story in your view? What ending?
Yeah, that is the question. Because if someone was short, believe me, I would never short Nvidia. You have to be out of your mind. But if you had a thesis that you were going to short Nvidia, you would say it's inning seven or eight and pretty soon the 55% is going to become 10.
Yeah. You know, there's a couple different flavors of that as well. And the bear case is just these numbers have gotten so big so quickly, how could they possibly be sustainable, right? You know, you've got Jensen out there though talking about, you know, in 2030 we'll be doing three to four trillion dollars a year of infrastructure spend. These are just like unbelievable numbers. So I think there's two ways that sustainability question can go wrong. One is the purely cyclical. These guys have spent a lot of money. They're going to keep spending money, but they'll take a pause for a bit. There's a digestion. And even before AI, you could sort of look at the hyperscale capex numbers and they would tend to build and digest and build and digest. It wasn't unusual.
And I always say, I've covered these stocks a long time. The chances of that at some point happening, it has to be 100%. Only a digestion period.
Digestion. It's not now. And I've been saying this ever since it started. It's not this year. Doesn't look like it's next year. If Jensen and Hock Tan over at Broadcom are correct, it doesn't look like it's 27, but at some point I guess so. I don't know when. Not now.
How would you... Let me ask you a question. What would be the sign that would tell you it's happening?
Yeah. I mean, you're looking, I mean, if the instant one of the hyperscalers cuts capex, like it's all over, you'd be too late by then. But the point I want to make, if you thought it was purely cyclical, it would be painful. But you could probably feel comfortable buying that dip because you say it's just cyclical and the outlook is still strong and these things, you know, they go up but it's never a straight line, right? Fine. I think the doomsday scenario would be they're spending all this money and there's no return, right? And if that's the case, then I think the whole thing would come crumbling down for all of these names and frankly for everything else. I think...
By the way, you could make an analogy to the 1999-2000s.
By the way, so I was on the sell side back then and sitting across the hall from me was a young Henry Blodget before he went to Merrill Lynch. And Henry was going out and he was making statements like dynastic levels of wealth are going to be created. The internet's going to conquer the world. And he was 1,000% correct. But so much money got spent so rapidly that the returns at first were not there. We had this enormous tech recession and then eventually we came out of it and Henry's predictions became true. So at this point, I don't see anything out there that's so wonderful coming out of AI that means the returns are going to be so high. I mean, the search is better, but it ain't crazy better.
That also is an argument that I actually do think we're still early if people worried about bubbles. And I get the fears, right? The numbers are very big. You know, people start to look at companies like Nvidia that are now starting to invest more aggressively in the broader ecosystem as well. People start to worry about...
You're pulling a little bit of a GE vendor financing.
We can talk about that, but people, look, I get it. It raises eyebrows. I understand. At the same time though, it's not like during the tech bubble, like they were laying fiber, for example. They laid dark fiber, stayed dark for 20 years, it didn't get used. Like we're not, nobody's buying GPUs and sticking them in a warehouse and stockpiling. They're all getting used. The demand is off the charts.
Just interrupt for one sec. The one difference though, important difference between then and now is then you had these rinky-dink companies that had just gone public that had no revenue, had a business plan that were spending money. These are real companies and these are the biggest companies.
My next point I was going to make, that's a big difference. It is. And people worried about bubbles and we're not anywhere near bubble territory yet. The valuations are actually fairly reasonable. They're elevated but they're actually fairly reasonable. Nvidia's, I mean, these guys, I mean, if you think the numbers are anywhere close to correct, it's not even expensive. It's way... And by the way, Nvidia is way cheaper today than it was before the whole thing started. Like the stock's up a ton. The earnings are up.
The earnings have gone faster than the...
Yes, they have. So it's not like these things are not trading at 100 times earnings like they were during the bubble, right? I always joke OpenAI hasn't even gone public yet. Like, can we be in a bubble? Like if that hasn't happened, so I don't worry too much. I lean a little more toward the idea that we're likely more on the earlier side than the later side of this.
So maybe three or four, like whatever you want to... I don't think we're in eight yet.
So let's just talk about the whole... I mean, Jensen talks and tells a story that would like insane. So just talk to me... pretty good and he's been a pretty good predictor and he's been right, but you still sometimes can't even believe it. So just talk about how big the opportunity is and then let's talk about from your perspective, what do you think this whole ecosystem will be capable eventually of doing? Because that's the real important question. What are the use cases for this?
Yeah, absolutely. So in terms of the opportunity, so these guys are already spending hundreds of billions. He thinks that number goes to two trillions a year. We'll see. Well, you're seeing countries like Saudi Arabia getting sovereign. And then again, I've got Altman out there who is very aggressive. And we were talking about this before we started. You know, so you got Oracle with their 450 billion in RPO of which 300 billion looks like it's OpenAI.
Let me just explain to the viewers what that meant. So Oracle came out and said they use this term, it's basically a fancy term for backlog, and they said that the backlog, quote unquote, type backlog had grown to 455 or something, 350% in like three months or a year, whatever it was. And then they sort of shut up and then people started to dig, people like you, and the word was of the 455 billion, 300 billion was just from OpenAI. And then the skeptics said, well, hold on there for a second, buddy Brown or Chad, they don't have, OpenAI doesn't have 300 billion to spend. They've raised 60 billion so far, but they just got another hundred billion in funding from Nvidia.
What do you mean sort of?
It's not 100 billion right away. They'll do it incrementally as they build it out. So...
But they... So let's say they got a hundred billion in the wallet, but they're losing money. So they're spending that too. So there's no way that OpenAI could spend a $300 billion yet.
They've got other mechanisms though. Like I said, they're still private. I think what's the current valuation? I can't even remember. Like 300, 500, but who cares? I mean...
The point I was making though was that Altman is very aggressive, right? So he's got that with OpenAI. He's got this opportunity now with Nvidia where they will be helping OpenAI to build out that infrastructure over time. I think Nvidia is going to deploy 10 billion, they're going to buy 10 billion of equity basically in OpenAI next year. Okay. Start with that first tranche and that'll be a gigawatt of power. And to your question on the opportunity, it's funny, the limiting factor it looks like it actually may be power. So how much, we'll come to that, but in terms of size of the opportunity, you can sort of think about one gigawatt. The numbers that have been tossed around, one gigawatt of power to power a data center, it's roughly call it like 50 to 60 billion probably of spend to build out that gigawatt. For Nvidia, that is probably 30 to 40 billion of revenue opportunity for the stuff that they sell into that infrastructure. Right. You know, I mean, look, Altman was out the other day saying he wanted to get to the point where they're building a gigawatt a week by the end of the decade. You can't build a gigawatt... I don't know what he can do. We'll see. But I mean, you can start to get big numbers, right? You can't build from nothing to things have to change. Certainly things have to change. So the opportunity is still huge.
The opportunity is still big.
Again, let's come back to the question of clearly these chips are revolutionary. They do things that couldn't be done before. How do you see like use cases? Because like for example, just from my perspective, when I go on Google and I do a search, I now get two responses. I get the traditional Google search and then I get the Gemini AI search. And the Gemini AI search is better and it's more comp... It's kind of like the college student versus the high school student. It's certainly better, but does it make my life that much better? Not that much better.
So in terms of the use cases, you've got those kinds of things. There's, you know, I have my OpenAI subscription for 20 bucks a month or whatever it is, and or I'm googling and I get my sort. So those are one set of use cases. I wouldn't call those revolutionary. They're interesting and convenient. I wouldn't call them revolutionary. I actually, by the way, I wonder if the real returns on this will not necessarily be, you know, we have the best model and we're renting it out for 20 bucks a month. I wonder if it's really productivity savings. So we're seeing massive improvements, for example, in productivity on coding.
Coding, right. We are seeing...
How much, let's nail that down a little bit. How much productivity savings are we seeing?
I mean, some of the numbers, I can't remember the Google, some of the others are thrown like half their code now. Half their lines of code are getting written by AI. So I don't know how that translates to productivity, but we're seeing a lot more of that, right? We're seeing companies that are actually starting to reduce headcount. So a lot of the big SaaS companies have been reducing headcount.
Software companies. We saw, I think it was IBM reduce headcount. This was last year, at least in their agent was thousands of employees in their HR departments.
You can imagine call centers, like call up American Airlines or whatever to change your ticket and you're now, you always were able to talk to a recording which was never that satisfying, but that's actually getting better. Okay. You could imagine, you know, you're going to McDonald's or whatever, you're going through the drive-thru and there's experimentation going on there. Um, and but I don't know what that means in terms of like, you know, employment and everything else at the end of the day, but I really do wonder if it's productivity savings that will help this. And then in terms of other broader use cases, you know, we're moving from these one-shot models to what are called reasoning models where the model itself uses a lot more compute, but it's a lot more productive. And even to what are called agents or agentic AI where you could imagine, you know, I want to book a trip to France and look for, get me plane ticket options that are leaving on these dates and I want to do X, Y and Z while I'm there. And, you know, the AI model will just go out and do all of that for you. And you can imagine it uses a tremendous amount of compute. We're not there yet, but that is clearly where things are going. So I don't really sign up for the idea necessarily that there were, that is the bear case, there's no use cases, this whole thing comes crumbling down. I don't think so.
It's not that there are no use cases. It's just that the use cases aren't good enough to justify the enormous amount of money that's being spent.
But again, we've only been doing this for like two years.
That's fair. It's still very early.
Still very early. In two more years, let's say by, you know, when we're into this five years and we're still having the discussion, I start to worry. Okay.
I don't think we're there yet.
And I also think that to your earlier point, the companies that are spending this money are not idiots, right? And they can see things that we cannot see. A lot of this used to be was all very open back in the early days, like when ChatGPT, the model structures and everything were very, everything's very closed now. Nobody's really, it's a competitive threat to put what you're doing out there. So these guys can actually see things that we can't see. And I don't think they would be spending money willy-nilly just to spend money. They have a purpose behind it. So okay, I don't understand. So when Broadcom comes out, Broadcom had an amazing quarter. They did. And this is not a company that I'll be the first to admit that I know that well. And they start talking about custom-made chips that they make for Google. Talk to me about the world of GPU versus custom-made chips. What's that all about? I really don't understand it.
So custom-made chip, you may hear the term ASIC. That stands for application-specific integrated circuit. It basically means custom chip. But I'm going to use that term just, okay, so it's not... So there are a couple different mechanisms to get the compute that is needed to do AI. So one is GPUs, and it's a whole other discussion how that came to pass, but one is GPUs, but the other is custom silicon. You don't have to use a GPU. You can design a custom piece of silicon to do this. And the idea would be, you know, you're not paying the Nvidia tax. The margins are pretty high. And ideally you can customize that chip to be more efficient for your specific workloads, for the workloads that it is designed for versus like a GPU which is general purpose and may do everything well but may not do individual things perfectly. So this is why we've been seeing this. And all of the hyperscalers are working on their own custom silicon. I will say that the only hyperscaler that has really deployed this in any great volume is Google and they have a product that they call a TPU, stands for tensor processing units, and they've been working on this for 13 years. They're on their seventh generation. Amazon has a version as well. They call it Trainium and they're on their, I can't remember, third or fourth generation. It's been...
So what do those chips do that the GPU doesn't do or do better?
So they can do anything that the GPU can do. So I would say you tend to design a custom chip for large stable internal workloads. Like it costs a lot of money and time to design a chip. You don't want to do it for a workload volume that's very small. You want a lot of compute and you want that workload to be stable because the issue with the ASIC is it's not flexible. Like there's no free...
The GPU is programmable and flexible. So there's no free lunch, right? So I'm willing to stipulate that an ASIC in theory should be more efficient for the workloads that you are designing it for. Otherwise, why are you bothering? But again, there's no free lunch. There never is. It's not flexible. If your workload needs change, if your model structures change, you may need to spin a new chip, whereas the GPU can be more flexible to handle that. So my own view is, if you do the math, by the way, in 2024, just looking at the value of the silicon itself, the ASICs were probably low double digits, 10, 11, 12% of the total...
Of total spend on AI silicon, processing silicon. So not that you have to remember, Nvidia doesn't just sell, they sell racks and all kinds of stuff. If you just looked at the silicon spend in 2024, the ASICs were probably low double digits. I bet this year they're probably mid-teens.
If I was to look forward, could they be 20 or 25% of a much bigger pie?
