00:12:21 · Joe Weisenthal, Tracy Alloway, Alex Altmann
[00:12:21] S2: Reconcile this. If there is this perception that people are still bullish, there is a, the train is still moving at full speed. Why should capital be hesitant? Like, okay. So it's like, you can, you can see on the numbers, like the core weaves and the irons and the cipher and like a bunch of other ones that I had never heard of. Like, yes, clearly there's the capital markets are telling something, telling them something. But why? If the view is that AI is still generally going straight up, what is the fear about these companies that are at the tip of the sphere? Wouldn't the bull say, look, someone is going to pay for these GPUs. The demand is going up. If they buy or rent or whatever, if they have electricity, if they have GPUs, if they bought them, whatever, they're not going to zero. They're going to be worth a lot of money somewhere or nothing. [00:13:17] S0: Yeah. [00:13:17] S1: So that, that is the bull argument is there's nothing wrong with it in terms of build it and they will come. I would say that, that the challenge will be to the early two things to the earlier point, tip of the spear. So when you start to see cracks or canaries in coal mines or however you want to characterize it, you should probably pay attention. Right. [00:13:36] S2: Um, and, and.
00:13:36 · Joe Weisenthal, Alex Altmann, Tracy Alloway
[00:13:36] S2: Um, and, and. [00:13:38] S1: The second point is, is that, thank you. Uh, the second point is we, we need to think about revenues at the end of the day. And this is something which we addressed at the end of August in a note called placing your AI chips more carefully. And the idea was, is that we basically decomposed. It was around the time of the SpaceX TAM. If you remember, that was around about $ 30 trillion. [00:14:02] S0: It was a total addressable market of like the entire universe, not just earth, something like that. [00:14:08] S1: Yeah. $ 30 trillion, which is obviously around about the GDP of the United States. And so we went, again, to the point about people throwing around trillions. Fun fact for you. Since 1990, I can't actually take credit for this stat, by the way. This actually belongs to Cameron Crease at, who also works here? [00:14:25] S2: Our colleague, yeah. [00:14:26] S1: And so give him the necessary plug on this. Looking at the cumulative revenue of the two most successful tech companies in modern history, it was Microsoft and Apple, and it was the cumulative TAM of those two was, I think, inflation-adjusted, not even $ 10 trillion since 1990. But SpaceX's addressable is talking three times that number. So that kind of prompts us, where's all this money going to come from? So punchline again, $ 17 trillion of S & P top line revenue. And they've got around about $ 11 or $ 12 trillion of expenses. And so that means that the net profit approximately of the S & P is somewhere between five and five and a half trillion.
00:14:26 · Alex Altmann, Joe Weisenthal, Tracy Alloway
[00:14:26] S1: And so give him the necessary plug on this. Looking at the cumulative revenue of the two most successful tech companies in modern history, it was Microsoft and Apple, and it was the cumulative TAM of those two was, I think, inflation-adjusted, not even $ 10 trillion since 1990. But SpaceX's addressable is talking three times that number. So that kind of prompts us, where's all this money going to come from? So punchline again, $ 17 trillion of S & P top line revenue. And they've got around about $ 11 or $ 12 trillion of expenses. And so that means that the net profit approximately of the S & P is somewhere between five and five and a half trillion. [00:15:14] S2: Dollars, right? [00:15:15] S1: So it's like, all right, we can make some bold claims about the amount of money that's going to get allocated to AI expenditures and all the rest of it. But last time I checked, like we all look at the ramp data or the open router data, and we can clearly see that that token expenditure has kind of begun to level off a bit, right? At least near term. So it's like, all right, our company is going to spend this much money. And that's where the other point to answer your question, Joe, is. [00:15:41] S2: Well, you know, your company Barclays is, I know, managing the anthropic IPO. So, or as, you know, involved in this, we're not going to ask any anthropic questions to you, but I will say to Tracy. [00:15:54] S0: Okay.
00:15:54 · Tracy Alloway, Joe Weisenthal
[00:15:54] S0: Okay. [00:15:54] S2: No, I won't make you comment. You know, we understand the rules here, but I will say to Tracy that anthropic numbers on the top line look great. And I do not think that nobody, and I don't think we learned that much about open route, open router data. because that captures a specific slice of people who are usually looking for cheap Chinese models, which they're doing well, but the top line numbers for Anthropic look unbelievably good. That's the interesting. [00:16:22] S0: Thank you. There's earnings progress there in the S1. You can see the earnings progression. However, billions in earnings versus valuations at trillions, I think that is the gap that people are debating. [00:16:36] S2: I'm just pointing out that There's lots of charts. There's ramp data that's floating around. Lots of people look at the ramp data and lots of people look at the statistics on OpenRouter. And they start to say, oh, the American Frontier Labs, I guess, are they in trouble? And they might be. But I'm just saying, if you look at the data that's filed with the SEC, as of pretty recently, at least one of the companies, the one that is filed in S1, that number is going straight up still.
00:16:36 · Joe Weisenthal, Tracy Alloway, Alex Altmann
[00:16:36] S2: I'm just pointing out that There's lots of charts. There's ramp data that's floating around. Lots of people look at the ramp data and lots of people look at the statistics on OpenRouter. And they start to say, oh, the American Frontier Labs, I guess, are they in trouble? And they might be. But I'm just saying, if you look at the data that's filed with the SEC, as of pretty recently, at least one of the companies, the one that is filed in S1, that number is going straight up still. [00:17:03] S0: Why don't we talk about the frontier models more generally? So, you know, the argument has been that they can come up with a better model, a better agent, a better technology, and that people are ultimately going to pay for it. And the persistent, I guess, doubt in the market has been, well, you know, open source models are kind of catching up very quickly. They're worse on their benchmark test, but not by that much anymore. And they're a lot cheaper to operate. What are people saying about the open source versus frontier debate? [00:17:31] S1: I would say here, this is weirdly, is a bit of a contradiction versus that bullish narrative. Because the data is the data you can evidently see on a whole bunch of different sources that open weight has kind of gone from a 20% market share to anything up to about an 80% market share. And then you'll also see that the token cost is a fraction, one one hundredth. [00:17:58] S2: Versus- Incredible token deflation that's happened every year.
00:17:58 · Joe Weisenthal, Alex Altmann, Tracy Alloway
[00:17:58] S2: Versus- Incredible token deflation that's happened every year. [00:18:01] S1: It might be one of the fastest deflating assets in history. It makes those NFTs back in 2021 too. [00:18:06] S2: But GPU rental rates have been going up. They have. [00:18:11] S1: Yes, absolutely. And so there is clearly a two-way debate. And I think it again goes back to the evolution of this broader AI narrative. It's not just necessarily a one-way train, rising tide lifts all boats. There's going to be degrees of differentiation. But you asked about open weight models. And I think where one of the leading edge questions right now is not even just open weight, but combine that with locally run open weight. And this is something which I like to call being long or short human creativity. [00:18:42] S0: Right. [00:18:43] S1: In so far that if, when the iPhone came out in 2007, stay with me here, uh, that if I had, if someone from the future had said 20 years from now, you're going to have whatever the orders of magnitude compute differential And the battery is going to be the same size. That 2007 person would have said impossible. You'd need an iPhone the size of this table, right? So clearly we got there amongst many other things. And so the idea that we necessarily need to ping a bunch of hoppers or Vera Rubins in an Ohio data center just so that we can sort of get our. [00:19:21] S0: Agent to- Manage my email for me. [00:19:23] S1: Yeah, or to send an angry response letter to the hotel. [00:19:27] S2: You're going to be able to do that with the chip on your phone by next year.