I went from Anthropic, OpenAI, etc. being the best businesses in the world to now I think I'm in the camp of there's an increasing chance that they are going to way underperform what people think. And there are three data points I'd love your reaction to. First is token consumption per customer seems to have exploded and now is starting to turn over in a number of things. We see this in the Ramp AI index, spending is down. We also see this in a number of these kind of metrics that look at, you know, whether it's OpenRouter or whatever, that that that's coming down. We also know Grok recently released their latest model and they showed a chart of performance and price. Everything is getting better performance and price is coming down. So it looks like everything's kind of diving down into the right.
Well, if you take those two numbers, number of tokens processed times price of token, that is ultimately the revenue of the model labs. So talking to Dan Ives this week and I said, "Look, if those two numbers are both coming down, Anthropic, like the latest charts are just kind of showing slight growth but to like plateau revenue." And so as businesses realize like maybe we need to own more of this ourselves and not just outsource everything to these two companies, it feels like there's some pressure there. Now, the overall market is growing. Maybe there's like breadth in AI where like the revenue flows, but could there be weakness in the model labs and maybe some pressure of them actually getting public? Yet the AI trade more broadly continues and so like you're almost getting like a reallocation of where value is accruing.
So let me, I'm going to disagree with a lot of what you said.
But not the facts. So the things you highlighted...
No, but you're falling into the trap that most people get into, which is, well, let's look at this data, let's go through. So let's let's...
No. So the first part of this year was about token maxing. Let's just throw tons of tokens inefficiently at everything and see what happens.
Okay, well, we knew that was going to end. So when you look at the Ramp index and all this stuff, the token cost is deflationary. It's going to continue to come down. There's models now that three months ago, I mean even OpenAI has the models from three months ago which are better today than three months ago or cheaper than they were back then. So everything is moving that way. So let's leave that alone. We've entered the agentic side.
The token numbers are going to explode. We are going to see... So all we had were enterprise agents.
Eight billion people on the planet didn't use anything. Well, now Grok, Instinct, Muse are just starting.
But how many of those... So this is where I think the nuance is. How many of those use OpenAI or Anthropic models?
Well, no, no, no, no. So let me let me get to the end because there... Let's go. What you said at the beginning, I am not some bull on OpenAI and Anthropic. But...
No. But here's the thing. Do I think their growth rate has already passed the fastest point? Absolutely. So the conversation, remember, two months ago was they're on an annualized run rate, which is still there for Anthropic, of 100 plus billion. Okay.
Nice business. Like, by the way, I'm not saying I'm not jealous of having that business.
Let's just get something straight for everyone who's like wants to fade these guys. The total revenue for this year that OpenAI and Anthropic will have at the end of this year will be greater than the entire software revenue. So let's let's just get something straight. Like, these are the most prof... These are the fastest-growing companies in history. Anthropic was a $50 billion company. It's now a multi-trillion in a year and a half. Like, you can't fade what's going on. Do I think they're going to get to a trillion dollars in revenue? Absolutely not.
But do I think people are underestimating what the biotech companies want to use, what the nanotechnology companies want to use, what everything in the physical sciences will pay for frontier intelligence? You're crazy if you think they're going to use cheaper models. Like in the hedge fund industry where I'm from, this has already happened. Like Jane Street, Millennium, they're all using Anthropic. So everyone's crazy if... But here's my thing is I'm going to argue that I don't know if frontier intelligence is the actual right label for OpenAI and Anthropic anymore. And I'll give you a... Let me give you two examples.
I think everyone starts with, I want to use AI. I think AI could be valuable. And so what you do is you outsource to what is the smartest thing I could use that's kind of off-the-shelf. And you go right to these guys.
There is no way in hell that at Silvia, with less than 10 people on an engineering team, we are able to build better models that are more accurate than these trillion-dollar companies for personal finance, and Jane Street can't figure out how to build a better model for their specific use case in terms of custom AI harnesses, data pipelines, all the model weights, and the training, etc.
And so what I think, and again, this is where we are now entering a world of, I call it Jordi thinking, probabilistic thinking. Of I can't say this with 100% truth, but I know this is not zero.
I think that OpenAI and Anthropic, for many businesses, are the perfect on-ramp. So you get this acceleration on a per-customer basis. And then the customers start to say, 'Well, can we get better intelligence? Can we get cheaper intelligence?' They start to try to optimize. That optimization point, almost every CEO I talk to in the private market is saying, 'We are starting to do this internally.' And so to me what that means is their future growth is about how do we go get more of the people who do not currently use AI to get onboarded, because they're almost like leaking their existing users, the per-token usage is going down. So they have to go get new customers to replace the leakage or the retention almost to some degree. Does that make sense?
Yeah, yeah, this gets back into the part I disagree with, is for everything that you've said, if I told you the most important point is that we're in the 1% right now of usage. And that Jevons paradox is a real thing, which I completely agree with. There's no way that OpenAI and Anthropic don't continue to grow. So there's a difference between growing as fast as they did for the early inception and continuing to grow. As someone...
I'm not saying they're not going to grow, by the way. I agree that they're going to keep growing. It's just the nuanced dynamics now...
are changing because you also have like Jev... Right, shows up. Like, there's a lot of stuff that's happening where you made a comment three months ago: the easy money in AI is over.
We are now, like, I'm fully on board with that and I think that it's becoming very obvious. You can't just be like, the big model labs get all the money.
Correct. And that part I agree with. The thing, the thing that as a user, and using my son as like an example, we both will crap on Gemini. I don't... When the new model comes out, well, this is good. It makes mistakes that I have OpenAI and Claude fix constantly.
So I think people that use it all the time and that are scared of mistakes and hallucinations and things like that know the models that are more likely to do this. I think when you make the decision to go to open source and you go to these, I personally believe on my own experience, it better be for stuff that isn't super important. That's why I brought up the Jane Street and Millennium. I don't agree that all these models are even close to the same. I've used... I stopped my Open Claw and my Hermes because I was able to do things in Claude and Anthropic that were different. And then when Grok came out and Muse came out and everything else, Muse is not as good a model, anything close to what OpenAI, but it doesn't need to be. It's just doing basic stuff for me. So I think we're saying the same thing. I just think...
I think the hardest thing for people to understand and the most important thing that I want to get across to people: you've heard Jevons paradox. We're at that point right now. What happened the last two months may have been an air pocket on one way, which is this token cost thing, but we're entering a very, very dangerous period now for the next year. The usage is going to skyrocket, and we don't have enough compute. So this gets back to the other thing OpenAI and Anthropic own.
I listened to the interview with the Instinct founder. Did you listen to it?
Yeah, Noah on Invest with the Best, yeah.
He specifically highlighted that he underestimated his compute needs.
He said, I think he said 40% of his time is being spent trying to accumulate compute, and he's having a hard time because he's growing 10% per day. And so he's like, how do I project out if I double my compute, you know, I basically use it all up in a couple of weeks.
And then how do I kind of project how much I'm going to need?
So here's the scary thing. We're entering the hype, the parabola of efficiency and compute as well with Vera Rubin. So Blackwells were great at the efficiency of the chips. Well, now we're getting to Vera Rubin. Who owns all of that? The big guys do. I mean, it's the speculation between SemiAnalysis and what I've heard on the All-In pod is that OpenAI and Anthropic combined will own up to 50% of all the compute coming online. I mean, isn't that a monopoly or a duopoly of some on compute? And so this is where I don't think you can make the assumptions that they won't actually go up again sometime next year just because they have the compute, which means two things. Number one, they can service more people and more companies. And number two, their models start getting better and better because you need more computing power to actually get to the level of the IQ that they eventually want to get to for the physical sciences.
