Will the AI Boom Create Enough Profits to Justify Today’s Valuations?

During their recent episode, Jake Taylor and Tobias Carlisle discussed Will the AI Boom Create Enough Profits to Justify Today’s Valuations? Here’s an excerpt from the episode:

[Jake] Well, yeah, we could talk about AI. I did a little back of the envelope math. And this was somewhat inspired by comments from a friend of the show, Chris Bloomstrand.

He was just talking about how much CapEx was going out and roughly what returns need to look like for that. What’s nice about this is we’re not getting into TAM or use cases or AGI timelines until terminators are running around. But this is a very narrow question, which is once all this AI infrastructure is built, what annual revenue does it need to generate in order to earn some acceptable return?

And so I’ve seen estimates from, call it, $3 to $7 trillion of CapEx that’s likely to go out in this build. So let’s just call it four to make the math easier. Useful life of this equipment, let’s call it 10 years, which I think is actually incredibly generous.

Leading edge compute is probably more like five years, maybe. But let’s give bulls a full decade’s worth of usefulness for this stuff. And then a required return of 20% return on capital.

And these companies historically have earned something low to mid-20s on their capital expenditure. So they’ve been great businesses. And so we’ll just kind of round down as the hurdle is 20% you got to get over.

And also, 20% is probably what it needs to be to justify today’s valuations. So maybe I’ll do a little quiz for you, Toby, just for fun. What is the amount of revenue that you need annually, on a run rate basis, for the AI system to basically pencil out with the spend and the return requirements?

[Tobias] You’re saying $4 trillion over a decade, but depreciating every decade. So $400 billion a year. Good.

You’re on the right track. 20% return over and above that, is that 1.2 trillion?

[Jake] Give this man a medal, everybody. Right. So 20% on $4 trillion is $800 billion.

And then you have $400 billion a year in depreciation that you need to come up with just to account for the assets wearing out. So yeah, $1.2 trillion every year. And by the way, this is actually a bit of a fantasy version, because that’s assuming a 100% contribution margin there.

That’s like, there’s no power, no cooling, no staff, no discounts, no idle capacity. And if we move that back down to like a 50% contribution margin, which is still pretty generous because I think electricity is like 20% of the cost there. Now the hurdle doubles, and we need $2.4 trillion in AI revenue every single year. And that’s roughly six apples of annual revenue at maturity, just to clear this bar.

[Tobias] So it’s not really like… $200 a month, times 12 months, $2,400. How many subscribers do you need?

[Jake] You need everybody, I think. Yeah.

[Tobias] And you need all of the Tesla cyber cab drivers to start AI subscriptions.

[Jake] Yeah. And we need it to be a big number too. You might not need to eat anymore.

You’re just going to have to have AI subscriptions. Now, there’s a second extension of the analysis that I did that I thought was kind of interesting. What we just talked about was only for the vendors.

Let’s look at it from the S&P 500’s perspective as an index fund. So let’s say that the vendors do get their $2.4 trillion in revenue. Like the hurdle gets cleared, we’re popping champagne bottles, it’s all good, right?

Now, does the S&P 500 actually get more profitable? When one S&P 500 company sells AI to another, the seller’s booking revenue and the buyer is booking an expense. So the index is really only keeping any net there.

And say half the AI bill is paid by other S&P 500 companies, which I don’t think is totally crazy. That’s $1.2 trillion a year that’s landing inside the index also as a cost. So say those customers are…

Let’s give them a little bit of credit. Let’s say that they’re crawling back a quarter of that through real productivity gains. So $300 billion back, $900 billion of drag.

Now we’re… Plus the vendor profit of $800 billion. We’re looking now at maybe a minus $100 billion for the index as a constituent.

And so the vendors can hit their return on capital targets, but the index may actually go backwards because of that. So I think the internet was totally revolutionary too. But it didn’t permanently raise returns for everyone who spent into it.

And with AI stocks, the accounting for something… I think right now it’s like 45% of the S&P’s market cap. I don’t know how you define AI stocks.

I think everyone’s got their own version. And it’s been, I think, 80% of the year-to-date gains so far. So basically, everyone is lining up on this bet that the vendor side of the trade is going to go really well.

I don’t know as an index, actually, if that’s going to work out. Because half of it, at least, probably is one man’s revenues, another man’s expense. And you, as the owner of both of these businesses, is just kind of treading water.

So we’re going to find out what happens there. I think it’s going to be really interesting.

[Tobias] What’s the $6 trillion in CyberCab revenue that’s coming in? Don’t forget that.

[Jake] Okay, Cathy, slow down a little there.