It's the year of the fire horse, and we get to celebrate it twice because I think we—I remember we talked about it at the beginning of the year, but the Chinese New Year did not start until today.
It's Lunar New Year, right? Worshippers burned large incense sticks on Monday outside a temple in Hong Kong to mark the Lunar New Year, which falls on Tuesday. People around Asia celebrate the start of the year of the horse.
In the Ultradome, it's the year of the horse every year.
Every year, I think so. Everyone is talking about the Cournot equilibrium. At least Dario Amodei and Dario... and me and some folks on the timeline, we're going back and forth and basically trying to get to this question. I feel like we should give the context on this, on the titling strategy piece, because we call the run of sh—like, we'll title an essay, 'Why is no one talking about the Cournot equilibrium?' because this one guy had this super viral thread like a year or two ago, and the title was 'Why is no one talking about Marc Andreessen?' We were laughing about it so much because he's one of the most talked-about investors in venture, constantly. He's written essays, it's been viral a million times, so many—yeah. He's someone that everyone in the industry has an opinion on already. He's not like a Midas lister.
Yeah. Yeah. Not a lot of people. It's like everyone's talking about. But in this case—
The basic idea is that if there's only a few players in a given market, you can think about, you know, any specific market, lemonade stands or whatever, and they aren't competing on price, they will compete on supply, and they'll try to predict what their competitors are doing and then respond accordingly. And this is really, really relevant to the AI lab discussion because you can tell that even though all the leaders of the AI labs say, 'I don't think about the competition. I don't talk about the competition.' All to use general terms, they're all obsessed with what everyone else is doing and they think about it constantly, very clearly. And if someone's buying 10 billion of compute over here, they're going to counter with eight over there or try and jump to 12. And everyone's sort of keying off of each other. You know, Microsoft pauses, AWS goes all in. There's all these like horse races. It's why SemiAnalysis exists and provides great, you know, cross-functional data. Outside of tech, there's this discussion that I see that's always funny to me where people are like, 'The price-to-earnings ratio—'
Ryan says they don't want you to talk about the—
They don't want us to talk about it. They don't want us to talk about it for sure. There's this discussion—I saw one person say the price-to-earnings ratio for OpenAI and Anthropic is just simply too high. And I was like, earnings? Like, these companies are losing money. They don't have a price-to-earnings ratio. Divide by zero.
This is going to blow your mind.
Yeah. It's so much worse than you think, right? They're not making any money. The other side of things is the inference factory. So this is essentially a manufacturing business. You have variable costs—GPUs, power, engineering overhead—and then your revenue is subscriptions, API usage, and enterprise contracts. And so when you just look at inference, you see positive contribution margin. And we can see that because we can compare the cost to inference a model of the GPT-5 class size or the Opus 4.5 size. You can see what does it look like to run an open-source version of that model on commodity hardware. It's way, way cheaper than what you pay to Anthropic or OpenAI. So they must have good margins. And everyone sort of agrees at this point that inference margins are in fact healthy. The question is how do you balance those two pieces and when do you risk overinvesting, and that's sort of this Cournot game of chicken that everyone's playing. The Cournot equilibrium comes when a small number of labs, an oligopoly effectively, choose supply at the frontier level and then the market clears at a high price for frontier access. So choosing supply in this case means how many data centers get built, how many GPUs get ordered, but also how much low-latency capacity is allocated to the top tier. So, you know, right now, OpenAI just did the Cerebras deal, there's Groq fast, and there's a whole bunch of different modes that will deliver faster inference. And how many of those fast queries you get, how much of the best chips are allocated to a particular tier that you're paying for is an economic question for the labs. There's a ton of developers and knowledge workers who are happy to pay hundreds of dollars a month or more, but they always want the best available model. This is most people in executive roles in startups. Yeah, I got my $200 a month subscription. I'll pay 250 or a hundred or whatever, couple hundred bucks, and it just makes me better at my job. I just do whatever I need to do, but don't give me the old thing. I want the best. I want to know that the hallucination rate is as low as possible.
1% of the time it makes a career-ending mistake.
So, having a product, not just an API business, gives you leverage because at some point the models are smart enough where you don't need to train them. You don't need to train a model that is 4% better because people are still coming to your application and having a good product experience. Right. Yes. So historically, one of the critiques of Anthropic's business was that they have to just be on this constant, constant fly—you know, sort of hamster wheel of training the best model because the majority of their business is this API business. Swap it out for a smarter model. That said, they have Claude Code now, which gives them some more leverage over the market. And the really interesting thing is that Dario is now talking about being near the end of the exponential or maybe producing like the final models because we've talked to a few people about this, but it's very unclear if it's possible to create like a superintelligence that's like 5,000 IQ. It might just be they get good at all knowledge work and they can answer all tasks, but it's like the digital god. At that point, it does commoditize and you drop out of Cournot equilibrium and you become more—customers are more aggressive about switching to cheaper models to cut costs because the frontier is now commoditized in the entire backlog. Everyone is at the frontier basically. And so in that scenario, you switch over to Bertrand competition, which doesn't really mean that profits go to zero, but there is more competition. And it looks a lot more like the hyperscaler cloud market, which is I think what people have been sort of signaling towards, and also it sort of explains why a lot of the VC firms are getting in multiple companies because they don't think it's going to be winner-take-all anymore. They think it's going to be much more oligopolistic for the long term and there will be competition between the major three or four labs and it will be much more about how can you marshal enough supply, create a huge barrier to entry. Like you and I could start an AWS competitor tomorrow, but it's going to be extremely expensive to bring up data centers that just serve web apps everywhere, let alone AI stuff. Right, building all those data centers—you're thinking what I'm thinking, you're thinking, you're thinking.
Ad. When I was hanging with my buddy, my buddy Ben, on Sunday. We both live in Malibu. He was thinking of just getting some chips and setting up Malibu Inference.
There we go. Just the name alone sounds like you could get at least one Malibu Inference.
Would be a beautiful name for a neocloud.
Yeah. Let's play the clip of Dario and Amodei discussing the economics of AI labs. Like we have a—you know, let's just imagine we're in like an economics textbook. We have a small number of firms, each can invest a limited amount in, you know, or each can invest some fraction in R&D. They have some marginal cost to serve. The gross profit margins on that marginal cost are like very high because inference is efficient. There's some competition, but the models are also differentiated. There's some—companies will compete to push their research budgets up, but like because there's a small number of players, you know, we have the—what is it called—the Cournot equilibrium, I think is what the number-of-firm equilibrium is. The point is it doesn't equilibrate to perfect competition with zero margins. If there's like three firms, if there's three firms in the economy, all are kind of independently behaving, behaving rationally, it doesn't equilibrate to zero.
Help me understand that because right now we do have three leading firms and they're not making profit. And so what—what's a good question. Yeah. What is changing? Yeah. So again, the gross margins right now are very positive. What's happening is a combination of two things. One is we're still in the exponential scale-up phase of compute. So basically what that means is we're training—like a model gets trained, it costs, you know, let's say a model got trained that costs a billion dollars last year, and then this year it produced $4 billion of revenue and cost $1 billion to inference from. So, you know, again, I'm using stylized numbers here, but, you know, be 75% gross margins and, you know, this 25% tax. So that model as a whole makes $2 billion. But at the same time we're spending $10 billion to train the next model because there's an exponential scale-up, and so the company loses money. Each model makes money, but the company loses money. The equilibrium I'm talking about is an equilibrium where we have the country of geniuses, we have the country of GPUs in the data center, but that model training scale-up has equilibrated more. Maybe it's still going up. We're still trying to predict the demand, but it's more leveled out.
There is another fun clip that we should watch from A Beautiful Mind. Jordi, have you seen A Beautiful Mind? It won the Oscar for Best Picture, I believe. It's about the mathematician John Nash. Have you seen A Beautiful Mind?
I have not. Wow. UNK status over there.
You would have. I was walking on the beach with Senra. Yeah. And we walked by an incredibly famous—one of the top movie directors of the last probably 10 years.
And Senra was like, 'Do you see that?' And I was like, 'See what? It's a guy with a dog.'
That's hilarious. I feel like your beach tours have been really star-studded lately. This is different from the previous one you mentioned, correct?
That's remarkable. Well, let's pull up the clip.
It's from A Beautiful Mind.
Nash, you might want to stop shuffling your papers for five seconds.
Yes, it's Eric Lyman in the Ramp biopic. We got our cast right here.
Oh, this is the original like looks-maxing movie. Anyone else feel she should be moving in slow motion?
Will she want a large wedding, you think? Should we say swords, gentlemen? Pistols at dawn. Have you remembered nothing? Recall the lessons of Adam Smith, the father of modern economics. In competition, individual ambition serves the common good. Exactly. Every man for himself, gentlemen. Yeah. And those who strike out are stuck with their friends. I'm not going to strike out. You can lead a blonde to water, but you can't make her drink. I don't think he said that. All right. Nobody move. She's looking over here. She's looking at Nash. Oh god. All right. He may have the upper hand now, but wait until he opens his mouth. Remember the last time? Oh, yeah. That was in the history books.
I think this is very, very stylized.
What are you talking about?
And completely apocryphal. Like he definitely thought of this theory, but not at a bar.
They block each other. Not a single one of us is going to get her. So then we go for her friends. But they will all give us the cold shoulder because nobody likes to be second choice. But what if no one goes for the blonde? We don't get in each other's way and we don't insult the other girls. It's the only way we win. That's the only way we all get laid. So he's describing the prisoner's dilemma. Adam Smith said the best result comes from everyone in the group doing what's best for himself. Right? That's what he said, right? Incomplete. Incomplete. Incomplete. Okay. Incomplete. Because the best result would come from everyone in the group doing what's best for himself and the group. Nash, this is some way for you to get the blonde on your own. You can go to hell. Governing dynamics, gentlemen. Governing dynamics. Adam Smith. He's wrong. There we go. Careful. Careful. Thank you.
Dave says, 'Just joined the stream. We watching a movie.'
Yeah. Lots of game theory going on in the AI wars right now. Everyone's trying to figure out how far to push it. There's a fair amount of risk. There's still the Cournot game of chicken around who will invest the most in advancing the frontier, but the end state looks a lot more durable than pure model commoditization and the perfectly competitive situation that many were predicting a few years ago.
Google Capital says, 'I thought Dario had a good point that software engineering is the only job where the full context needed to do the job is available to an AI agent via the codebase. And I didn't think Dario had a good answer for why automating other jobs will be as easy.' This got a bunch of people kind of reacting, kind of disagreeing generally that all of the full context needed to do the job is available. But I do think it's something we need to figure out. Yeah, we were debating this because there was a post that was just sort of like a Wojak reaction that was just making fun of this and it wasn't clear if they were saying that they were agreeing or disagreeing. But basically my take was, well, it's possible that a lot of the full context needed to do the job of a lot of different white-collar jobs is in fact logged. It's just logged in the final product, which is like a deck or a spreadsheet or a decision, and then a whole bunch of emails, a whole bunch of Slacks, and then a whole bunch of Zoom calls that's recorded. And so yes, if you're running a business where a lot of work gets done in smoky bars late at night and back-alley deal-making, sure, that's going to be harder to automate. But in the world where it's someone sitting in front of a computer and there's a screen recorder running, like you should be able to pull up most of the context. At the same time, you can't just snap your fingers and go back and get every decision that was made in the '80s that allowed Coca-Cola to become a dominant soda maker. But you can't get people on calls just being like, knowing the call's being recorded and used to train something to replace them. They're just like, 'I'll tell you offline. I'm not speaking this secret into the record.'
Golf this weekend. Debate around the posture. Dario, his posture was—
He's been in the gym a lot. He looks fantastic.
Um, and I love this sweater. The crew neck works really well. The pushed-up sleeves is a particular choice. Didn't translate into that Chad Wojak, but he looks fantastic here. A lot of fun on the timeline looking at the looks-mogging or whatever, or the looks-maxing. I don't even know. I can't frame-mogging. That's the—
Yeah. He kept bringing up the example of a video editor. Yeah. Saying, 'Yeah, but when will the models be good enough to edit videos well and pick out moments?' Yes. And give me two years and another 500 billion. We've tried every tool there is. They can't do it yet. Tricky. I don't know. I don't know what's—and it's not even that we're not trying the tools to replace the people on our team. We're trying to make them have higher output. One interesting thing is that there aren't a lot of open-source like Premiere profiles. Like I've edited a ton of videos for YouTube. There's a whole bunch of cuts in there. What I cut out, what I didn't. You could have that record, but it's not stored in GitHub. Like it's just—you can't necessarily train on it. You can train on the final product and understand, but you don't understand what actually got left on the cutting room floor. There's this whole concept of like 'kill your darlings.' Like when you're in the edit, like you need to be cutting more. You're like, 'Ah, I like that shot. It's so cinematic, so cool.' But does it actually advance the story? No. So you cut it down. I was watching The Matrix this weekend and there's this amazing shot of when Neo and Morpheus are going to visit the Oracle and they reach for the doorknob and the doorknob has this perfect reflection and the reflection shows Neo and Morpheus and they had to do this crazy VFX shot to hide the camera in what looks like Morpheus's coat because if you point a camera at a mirror you see the camera and you don't want to see the cameraman there. That ruins the shot. And so they did all this crazy stuff to cover up the camera. And I'd seen the behind-the-scenes and been like, 'Wow, that's really impressive.' And in my memory, I thought it was like, 'Oh, it's such an important shot. They probably lingered on that for like five seconds to really let it sink in.' Like they're pulling a trick on the audience. It's beautiful. It's like half a second. And they did all this work and then they knew that from a storytelling perspective, you don't want to hang out and watch a picture of a doorknob for five seconds. And so all these decisions, like they sort of get chronicled, but they don't get neatly organized in the way that a GitHub log does with pull request discussions and what happens. So it'll be difficult. So maybe two years and another 5 billion does it, but it's coming, so we'll keep monitoring it.
Andrew Reed says horses don't stop, they keep going.
Wait, did he actually say that? Yes. No way.
Yes. In response to 2026 being the year of the horse.
I love it. One of the greatest lyrics of all time.
Originally, to explain the joke, it's a Young Thug song and the actual lyric is 'hustlers don't stop, they keep going,' but it sounds like horses, and so people put 'horses don't stop, they keep going' and they show the AI-generated image of the horse bench pressing and it's incredibly inspiring.
