Guys, today we are talking to Doug O'Laughlin from SemiAnalysis. SemiAnalysis has some of the best research and reports on the semiconductor and AI industry. I always look forward to their reports. Doug, thanks so much for hopping on.
Yeah, thanks for having me. So yeah, let's chat.
Big week to come on, man. Big week to come on. You came on at the perfect time. How sick are you of talking about Nvidia earnings?
Surprisingly, this one can maybe — guilty confession — this one was kind of boring. So it wasn't exactly like there's a lot of... Yeah, it was very boring execution.
It's a perfect print. I have no problems with it. It's like exactly what you needed. It's better than buy-side expectations. It's better than the street. A little better than what we thought they would do too. Like it's just a really good guide, a really good result. Even with the H20 not even coming back, they're still selling a lot of compute. Gross margins, you might have a slight complaint, but like that was kind of expected. I think it's a really good print, but I think — and like look, the market took that to be it was a really good print the morning of, right? But and then we faded all day. So I mean, I thought that this was going to be the all-clear. Obviously, we even got the all-clear and it was still a selling event, which tells me it's not just about Nvidia in terms of the broader market environment. But I do think when it comes to Nvidia specifically, it was like such a good print, it's boring, you know, just imagine like, you know, they beat all the numbers for the most part and the guide was great, but still it's not good enough.
That's what was surprising to me was that how just like again across the board, they just crushed it. There was zero concerns moving forward. Even the guide was great. But yet the market was just kind of like they brushed it off, which was I think concerning to a lot of people like, okay, well if this isn't enough to get the AI trade back on, does this mean that the AI trade is fully cooked or what else is going on? What was your take on that?
So I think look, the AI trade's had a really good year. Let's be real, right? It's had an amazing year, even after a really good year last year, and honestly I would even argue part of the year before, right, like into the end. I think, you know, stocks do go down sometimes and I think there's been a little bit of... Yeah, I know, up-only doesn't work. It is not going to be up only forever, and I think the reality is we've had a really, really, really strong run since the tariff tantrum, right? I think we've been above the 50-day for like — this was a really surprising thing I learned — we are above the 50-day moving average for the second longest streak in history. Just recently we broke through it. The second longest streak in history. I was like, wait, that means we've been in a huge ripping bull market, which we have been. And I guess I was just really shocked. I was shocked by that. And so I think that this is just an example of like, hey, stocks do go down. They don't go up only. And yeah, most of what's actually driving I think the price and narrative is the Fed, right? The fact that this morning the reason why things have ripped at all actually is because some of the more dovish comments. We went from maybe no cut to definitely no cut to actually we're cutting again. So we went — it's so over, it's so back — or no, sorry, the other way. It's so back, it's so over, and now it's so back, right? And that seems to be moving stocks more.
So we're recording this on Friday morning, November 21st. And so this morning a Fed official came out saying that oh, a rate cut's back on the table when there seemed to be no hope of a rate cut. And so yeah, those comments are moving the markets back in the green. So it seems like the market is more focused on rate cuts right now. But like bigger picture though, if stocks, if this AI trade, let's just say if this AI correction continues to unwind and we still see a bleed, whether it's a fast bleed or a slow bleed in these AI names, do you think that's going to change the decision making from some of the hyperscalers? Because, you know, it's easy announcing a hundred billion dollar investment in a data center in Louisiana when stocks are at all-time highs. It's a totally different thing when you're doing it when you know you're off 30% from the peak. Do you think that's going to change the decision from some of these hyperscalers?
I think it just really depends on which hyperscaler. Look, Meta — okay, so specifically Google and specifically Mark Zuckerberg, they are post-economic, money doesn't matter anymore to them. It's about technology and it's about winning. You don't become a billionaire because you're not a crazy sicko who wants to keep winning, right? I think Zuck, part of the year of efficiency, right? It wasn't like Reality Labs was part of it, but like everyone else was doing it. I think that there was a belief and driving willingness to continue to figure out this new technology, and it was clear to me Reality Labs wasn't it, and it is clear to me that AI is right. I, you know, I messed around with an Oculus when Zuck was doing Reality Labs and I was like, this is pretty cool, but I don't think this is going to change my life. I messed around with ChatGPT when, you know, obviously the ChatGPT moment. I use Claude, all in various amounts, and it is life-changing. It is a transformative technology. Now, like every other transformative technology in the history of capitalism in our history, often there is a big capital cycle that builds capacity and supply to be ahead of demand. And we're starting to see that right now. In the beginning it's very valuable and it makes a lot of sense to invest. In the end, it's overbuilt. And that's kind of the cycle that every single one of these technology transitions go through. You know, the internet is the one that is probably the most recent one, so it's the most available. But there was a giant overbuild for like even semiconductors at one point. Even, you know, the automobile industry whenever that, you know, in the early 1900s, there was an overbuild. There was an overbuild for railroads. Railroads are probably the most famous.
Right. But so I want to stick to that point here. You mentioned obviously the buildout of the railroads. Everyone talks about that. There was a huge boom there and then a bunch of railroad companies went bankrupt. The internet, you had a bunch of fiber buildout, right? Fiber everywhere. But I think what people are concerned about with this buildout is that like fiber became useful all these years later. People still use fiber. The railroads are still used today 100 years later. There's concerns around all this money being spent on data centers and specifically these chips, right? What is the useful life of an Nvidia chip? Is it just three years, four years? You have Michael Burry tweeting every day talking about how these depreciations are underreported and how the useful life is only like three years compared to what Nvidia is saying that it might be six years or even longer. How do you see that? And how does that play into the bubble fears compared to like previous bubbles?
So I think that's a great question. I definitely think there's a five to six year useful life if it makes sense. But the way that we think about it is less about like accounting useful life and more like economic useful life. There is — I mean it's kind of complicated really, like we could go into this in a lot of different ways. So look, it's not going — you know, you're building all these assets and they are not, let's say, long-term. But the thing is what happens is if you end up spending for another amount for a new chip, you end up having meaningful improvement in parity. You're right that this does create a little bit of a treadmill aspect, but we're also seeing the inverse of it where there's a lot of older chips, for example H100s, that are still being used today. So that's like the counterpoint, right? Is that older assets from in this cycle still are being used, and CoreWeave for example booked H100s at 95% of the original price for a contract. That's huge. But all of this is like, you know, you are right that there is this kind of almost this crazy decline curve that's going to be really interesting to see. And I think that crazy decline curve, you know, you see this in all kinds of things like shale for example has a 12... Are you familiar with the shale well decline curve?
I live in Houston, so yeah, I mean I used to work in the oil and gas industry. Yeah.
Okay. Yeah. So right, this 12 to 18 month — it's the same kind of aspect here where like you have these crazy cash flow IRRs and so you get all the cash up front and you can actually ironically reinvest in it. It's just a function of supply — like what's going to happen is supply will keep — it's a function of demand rather, supply is going to keep reacting to demand, and as long as there's demand on the other side to purchase these things, people are going to reinvest their cash back into it. And so you can actually have even this more perverse thing, like same brain mindset of like, hey, well if you make so much money so fast on these shale wells, you could just take that into another shale well, right? It could compound infinitely and also crash like crazy. But... Yeah. And look, I think that's going to happen. This cycle will happen. Like I can't be like that's not going to happen. But the way I look at it today is like there's a supply and demand curve. We know the supply and demand of tokens. We know that there's a supply of — we at SemiAnalysis work really hard to estimate and understand the supply of total tokens. But we know there's a demand curve for tokens too. Now the biggest dream of all dreams is all of labor in the global economy. So, you know, a hundred trillion or whatever number you want to say, like a hundred trillion dollars, right? The reality is always going to fall short, that's just how these new technologies are. But we do think there's a meaningful real demand for most of these use cases, and we think that there's a real line of sight for multiple billion people using ChatGPT or some proxy day-to-day in their life and paying some amount of money for it. Not 100% of them are going to be paid, some of them will be free. There is a path for, you know, OpenAI and Anthropic to be making triple-digit billion dollar ARR. That's valuable. Now the problem is triple-digit ARR still can't even support some of the capex, and that's I think the real crux and concern that people are having in the market right now. It's real. I don't...
The $1.4 trillion spend that Sam was talking about, and then right now they're sitting at say 12 billion ARR or whatever, maybe it's 20 billion, whatever the number is... 20, but that's still not even close to — that doesn't support the $1.4 trillion buildout. Now, maybe that's just a marketing number, right? It's not really a real number, but I think when people start seeing these big ginormous numbers and they see like the revenues coming in, it's like, well, how do you justify everything? And then that's where the concerns start happening, and whenever the market starts dipping, then people start feeling worse about it, and then there's like this cycle.
Yeah, there's definitely a self-fulfilling cycle of fear, right? But let me specifically talk a little bit about that. It's a 20 billion ARR, right? So let's say it can support — we're going to do gross margin zero or whatever — that can support a hundred billion dollars of investment over five years, right, because it's a five-year useful life. I think one of the reasons why it's $1.4 trillion and some of that is double counting, some of that's like over-announced, and some of it is like definitely it's very in the back half of all this. What happens is like they announce a $500 billion deal, right? There'll be like a hundred billion dollars of it will be like already investments in flight, and then they will like — it's like okay, 100 billion, 100 billion, 200, 200, and so it scales and ramps. These are often very back-half-weighted deals, and these deals that are back-half weighted you pay with them in future revenue. So that's kind of the calculus I think is that, you know, they still believe it can grow triple-digit revenues, right? And I don't know, you know, 20 to 200, that would be pretty crazy, like 20 to 100 or something is going to be a huge, huge, huge lift. But we're talking about — there's never been, or maybe not never, but it's really rare that you get to 800 million users and you don't go straight to like two or three. And so internet scale we'll call is three billion, right? Like that's Google, that's Facebook, right? And three billion users, and the vast majority of them aren't even paying. There's ways for you to make money from that. And so I think OpenAI has to show that they can monetize quicker, and that's kind of going to be some of the push. We wrote about how we think they will. We think purchasing, right, which essentially is going to be taking a vig on every transaction that happens, is going to be one of the ways they'll be able to quickly monetize things.
