Can the AI buildout pay for itself?

Very hard because on the demand side of the compute, it's going to depend on new applications that get created. So, for example, the demand for tokens depends not just on AI chatbots, but on the new agents that are coming along. You have these new video generation models, SORA. Now, XAI has just released something. So, we don't really know what the demand for tokens is going to be. I think it's going to be huge. I think there's a lot of applications that haven't even been invented yet. We are in the very early stages are going to drive demand.

I don't think that's the question. I think the question is, what is the ROI on input and output tokens per use case? What is the ROI of a random use of a model? And I think that that is a very good question. How many of those tokens are generally gross margin negative? How many of those are unprofitable? How many of those are profitable? And I think that's what they're trying to ask when they ask that question. They may just not know the technical intricacies to ask it that way. But I think that's what people are trying to debate when they're trying to guess. And I don't think we totally yet know.

The irony is, unlike many exciting times in technology, they're trying to capture yesterday's demand, right? The demand is way outstripping our ability to build data centers and to fill them with hardware, all right? And so, you know, we have a $25 billion backlog.

And we are not alone in that. The Open AI, Xanthropics, you go through this list of, Google wants more data centers, Microsoft wants more data centers, AWS wants more data centers, right? All of these players are not chasing sort of, if you build it, they will come. They're chasing, the demand is booked.

It goes into the box, and the first thing it does is, it pays the data center, it pays the power bill, it pays the interest and the principal, and then whatever's left flows back to us. Right? And so it is an incredibly well-structured, time-tested, pressure-tested vehicle to be able to borrow money against client paper and all of the other collateral around the deal, which is why CoreWeave, which is a company that many people haven't ever heard of, was able to go out and raise $35 billion in 18 months to build infrastructure at scale. But what's important to understand is, the economics in this box are such that within two and a half years of a five-year deal, we have paid for everything.

Come to us. If you have a deal, you can bring it to us, and we will give a residual value guarantee. And I think the way that that residual, which will further lower the cost of financing these, and that residual value guarantee can then be incorporated into those companies underwriting. And what that basically means is they're saying that after three years, four years, we guarantee that these GPUs can be rented at a certain rate. And then the risk that they are bearing is the risk between that value wherever the market rate is.

And the useful life is when you're actually realizing return value from that asset. Burry's point is incorrect. On Twitter he said, the idea of a useful life for depreciation being longer because chips from more than three to four years ago are fully booked confuses physical utilization with value creation. That is incorrect. There is value creation because they are generating revenue from those chips this year, six years later. So there is in fact a useful life for that chip that has extended into year six.

Hold on. I think the most important point of what you said there is, Chamath, this is a new asset and it's a new space. There might be some companies that are like, we're going to lose money on this and we're going to let you build, I don't know, Sora videos, right? That's just going to be a money loser for five years as we get ChatGPT to 2 billion users. We don't care about it. And those machines are going to have 90% of their utilization in the first three years. And then for the next seven years, it will be 10% of their value. They'll be doing some small jobs in the background that won't be as important. Accounting isn't built to do this kind of refined depreciation schedule.

Come to us. If you have a deal, you can bring it to us, and we will give a residual value guarantee. And I think the way that that residual, which will further lower the cost of financing these, and that residual value guarantee can then be incorporated into those companies underwriting. And what that basically means is they're saying that after three years, four years, we guarantee that these GPUs can be rented at a certain rate. And then the risk that they are bearing is the risk between that value wherever the market rate is.

I guess the thing we don't know is how much more efficient will the infrastructure get at producing these tokens. So that would be the offsetting factor is the chips are getting so much better. I mean, we're moving faster than Moore's law here. Every year, NVIDIA is releasing a new generation of chips, and it's doing inference at multiple times, the efficiency of the previous year on, I guess, a token per watt basis or whatever you want to measure it. So that's an offsetting factor and then there could be more efficiencies created in the model architectures with sparser architecture. So those will be offsetting factors is all the efficiencies we get. But then you have to, if I'm kind of arguing with myself, then you've got the Jevons Paradox, which is as the cost per token comes down more and more, that's going to enable AI to be used in more and more contexts where previously it wasn't economically efficient, now it will be. So that's going to fuel the demand.

And there's already a private credit problem right now. So you just layer on private credit, like Owl Capital getting all the refunds. And then you have these huge companies that have cashflow, but they're spending all their cashflow on CapEx and then they're borrowing on top of that. 50-year bonds. Right. That's planning for perfection.

And that's going to be hard. And we're building these data centers. And if there's a price performance curve on AI that minimizes the power requirements, there's going to be a lot of data centers that are going to be turned into pickleball courts.

The biggest risk to me is not on the demand side. The biggest risk is that you get a glut of compute and you get an overbuild. And the same way that we had dark fiber after the dot-com crash, if you had dark GPUs, that'd be a disaster for everyone. Especially if you built out your compute infrastructure expecting a spot price of 30 to 50 dollars a lot. As Elon said that they were expecting, right? So if all of a sudden there are too many people racing to provide this compute and now there's an oversupply and the market crashes, that'd be the risk factor.

And many analysts and many investors that are intelligent and do their homework will look at the cash flow statement and they will see the capex, they will see all the investments going out and they will calculate a number typically called free cash flow that will allow them to estimate the true cash generation of the business in a particular period and make an assessment of should they be valued on free cash flow or should they be valued on the GAAP standard of EBITDA and the investor has the choice on how they want to value the company and Burry is incorrect in thinking that they're hiding anything because it's all there, they're following GAAP standards and then investors make a market and they all decide what do I want to value this company on? Cash flow, EBITDA.

Oh, for sure. I was covering it. Definitely, 96, 97. I think that's that early. I think it's 97, 98. I do think we are seeing at the late stage a disconnection in reality and valuations where companies are being given far too much credit. They're being priced at 50 to 100 times top line revenue.

What we're seeing now is revenue and profits and growth in. in revenue and profits that we've never seen before. And that's kind of the anthropic story. It's the data center, infrastructure, networking equipment story. I mean, it is all real dollars flowing versus speculative utilization. You're exactly right. People were basing it on metrics like hits to a website and web pages. Here you have a very distinct difference.