Invest Like the Best with Patrick O'Shaughnessy
· Dan Sundheim, Patrick O'Shaughnessy
Why would people be incredibly upset if you said you had to not use Spotify anymore? It's because it's personalized. It's because they've tailored the service to take a product which is a commodity and personalize it to the point where you're willing to pay a premium for that commodity. [00:00:00] S0: If you were giving advice to the executives at these companies and telling them what to lean into and what to look out for over the next five years, I'm curious what you would say. Because the scaling laws are so interesting in the sense that like, the models keep getting unbelievably better and that probably means the revenue available is like – who knows how big it could be, it could be the whole world. But the cost keeps going up by orders of magnitude. The Colossus 2 data centers, this unfathomably big thing, it’s like two gigawatts of power. It's crazy. What advice would you give them based on everything you've learned about these big, massive businesses?
· Dan Sundheim
[00:00:00] S1: The really interesting thing and challenging aspect of these businesses, the LLMs, is that the models they are building now, and especially in the future, can be applied to almost any aspect of the economy. You can take these models and you can make consumers' lives more efficient by having them be personal assistants. You could solve physics problems, you could help with drug discovery. You could make enterprises more efficient. The TAM is certainly not the problem. Focus is going to be a question mark. On the one hand, the more end markets you go after with a fixed asset, the better. You're spreading that cost over more end markets and having more revenue, which then can be reinvested. I think the flip side of that is that I rarely have seen any company succeed trying to go after multiple end markets at the same time. Usually, you have an A team. An A team is focused on one thing. Your culture as a company is oriented towards either consumer or enterprise. Even Amazon, which you'd say is an example of a consumer company that got into enterprise. They got into it like, seven years later, even after they went public. So trying to do everything at once is tempting, because if you're successful, you're effectively just amortizing that fixed asset over more revenue streams. At the same time, you risk not being the best at any one thing. So that is the trade-off. And I think that I'm not sure we have the final answer right now. The market has gone through periods where they thought Anthropic was Lyft and OpenAI was Uber. And up until recently the sentiment on OpenAI was more negative. I think OpenAI is taking the strategy of, “Let's do everything.” We are gonna go after Apple hardware, we're gonna go after robotics, we're gonna go after enterprise, consumer, science. They've been very successful in a lot of ways, but that's hard. I'm sure there are companies I'm not thinking of, but I can't think of many examples where that's been as successful. I understand the temptation to do it, and obviously the difference versus history is that the smartest people in the world are all going to work at these companies. So if anyone's going to pull it off, they will.
· Dan Sundheim, Patrick O'Shaughnessy
Anthropic took a different approach and just said, “We are going to focus on enterprise.” They tried consumer early on, but it came clear they didn't have traction, so then they just went all in on enterprise. They've had a lot of success with coding and enterprise. Because they've now taken a market leading position, generally the sentiment is that Anthropic is winning and they are like now the Uber, if you want to use that analogy. I think this is going to go back and forth over time and people probably get carried away in both directions, but I think those are the biggest differences. I would probably err on the side of focus, but I do understand the economic rationale for trying to do as many things at once. The only thing I – early on, we invested in OpenAI, this is probably a year and a half ago. I said to them “You have to do ads.” I understand, I've seen it so many times. People in Silicon Valley, the idea of ads is like… [00:00:00] S0: Gross to them or something. [00:00:00] S1: They’re allergic, because it's like – I have this amazing pure technology product and you want me to like [00:00:00] S0: Taint it?
· Patrick O'Shaughnessy, Dan Sundheim
[00:00:00] S0: Taint it? [00:00:00] S1: Taint it with ads. And like, you see Anthropic’s Super Bowl commercial. That being said, even the companies that were the most adamant about never getting into ads, like Netflix. If you go back and just listen to what Netflix was saying even 15 years ago, it was like, “Getting into ads?” Even Reed would've been like, “You are out of your mind. We would never do that.” Ultimately, they did it. And to me it's like, if you're going to do it, ultimately – one, you can't really compete against companies that are using ads if you're not. Very hard. If you're ultimately going to do it, you might as well start earlier because you have to build a culture around it. It just takes time. I don't think it's a big deal that OpenAI waited, but I was probably – rightly or wrongly, I was pushing for ads sooner than they've chosen to do it. I think now they're probably going to get it right. [00:00:00] S0: I'm so curious what you think is going to happen to the hyperscalers now. I saw this news report the other day that Anthropic is considering securing 10 gigawatts now of their own power. Which just makes me think, “Okay, they're gonna have the power, the scale's gonna be so big, why don't they just create their own clouds effectively?” The hardware might be different, more focused on inference, et cetera. Does that jeopardize what these business models might be – which I think people have thought of as pretty damn good at the hyperscalers – do you think the future is different as a result of AI?
· Dan Sundheim
[00:00:00] S1: I do. I've kind of thought this for probably about a year now. I am more confident in the thesis that the hyperscalers are a worse business model going forward. Now it's interesting because usually when you say something's a worse business model, you're implying that growth is going to slow, margins are gonna contract. I actually think you're gonna see the opposite. I think that AWS and Azure – maybe Azure doesn't accelerate, but certainly GCP – I think these businesses are going to accelerate for a while just because they are – their customer bases, Anthropic, OpenAI are growing at enormous pace, and as they get to be a bigger part of the business, the growth accelerates. The problem is that you went from a dynamic where AWS, Azure, to some extent GCP – their customer base was like every corporation in the world. So therefore they had fragmentation and they had the benefits of massive economies of scale that no single company could get, and it was a good, very good business. The problem going forward is that I think that economically it's highly unlikely that LLMs are not very concentrated in the hands of four or five companies. Those companies right now, they are obviously, as we discussed, they're investing a ton and they're cashflow negative, and therefore they're looking for compute anywhere they can get it. But if we're correct, and if anyone who owns these companies is correct, at some point in the next five to ten years, they'll be generating enormous amounts of free cash flow. When that happens, I think that they're likely to insource the compute, and every year AI is going to be a bigger percentage of the workloads at any hyperscaler. And so if you roll out 10 years from now, I think that the majority of the workloads will probably be AI. The LLMs will probably be providing a lot of those workloads, and I think that it will make economic sense to take it in house. Right now, I think that they look at the hyperscalers as more of a financing mechanism. These are well capitalized companies with big balance sheets. But I don't think these companies are better than them at building data centers. Like, building CPU clusters is different than building GPU clusters. Running inference on GPUs is very different than workloads on CPUs. And I think these – the LLMs are actually better at inference than the hyperscalers.
· Dan Sundheim, Patrick O'Shaughnessy
Then you have this whole dynamic of neoclouds. I think that the initial view for most public investors was that this was pure overflow capacity. There weren't enough GPUs and these things would be dead as soon as Microsoft got their GPUs. [00:00:00] S0: Caught up. [00:00:00] S1: Yeah. I absolutely would not make the case that they're fantastic businesses, but I don't think they're going away like people thought, because I think they're better at running GPU clusters than the traditional hyperscalers are, and I think there's a lot of interest from Nvidia and other chip companies to make sure that their customer base is diversified. Nvidia has a very big balance sheet and they want to keep these players in business. So over the next 10 years I think that these hyperscalers – AWS, Azure – will grow fast. I think the margins, my guess is, will be challenged both because the businesses are getting a lot more capital intensive, because AI is capital-intensive, more capital-intensive than traditional workloads. And also the customer base is getting more concentrated. Meta is not a hyperscaler, but they insourced all their compute. Why would they pay? I mean, they're just too big to use somebody on the outside.