The Information's TITV
00:05:45 · Paul Kedrosky, Akash Pasricha
[00:05:45] S1: Well, no, because again, think back to the size of Texas as a sovereign nation. Texas is sovereign to replace that. There is no place where you can replace that size of queue if you grant that something on the order of, you know, 25 or 30% of the projects within this so called sovereign nation of Texas in the data center sense is gonna see delays. So, no, it's not straightforward to see how you do that. And I mean... Obviously, this has all kinds of other consequences, which I've written about many other times. The data center is a de facto stimulus program in The United States. It was on the order of 25 to 40% of US GDP growth for umpteen quarters in a row now, is which one of the reasons why the prior... While the current administration embraced it so strongly that it was a de facto stimulus program that was in part bankrolled by the private sector and then steamrolled a lot of local interests in terms of the permitting and build out. So this has consequences in terms of the health of the economy writ large, not just, you know, Jensen Huang's bankroll. [00:06:37] S0: So let me ask you this. You know, the data center build out is one half of the story. The other half is businesses trying to find efficiencies from the AI tools that we have as of now. So, even if the models don't continue to get better, I mean, we have these models and there's probably still some pressure from shareholders to use them, do you think that that pressure decreases at all? Because even if the tools aren't getting more cutting edge, you still have the story of Oh, AI absolutely.
00:07:09 · Paul Kedrosky
[00:07:09] S1: Oh, of course. Listen, and this is the thing that people get wrong all the time, this whole idea that you can hold two contradictory ideas in your head at the same time. That the data center build out is on this precarious point doesn't mean that large language models don't work. That would make all of this much easier. Right? Because the problem is they actually work remarkably well even in their current incarnation. So as you say, without having to go to building out even more advanced models, corporations under huge pressure to not only find efficiencies, but more specifically, and you saw this with Revolut the other day, was to say, wait a minute. This is costing me a fortune, and I grant you it works, but I need to find a way to make it less costly. So which the big pressure you're going to see going forward is increasing on premises use of large language models, in terms of sovereignty of data and cost reduction, which is bad for cloud based large language models, frontier model companies, but tremendous in terms of the likely productivity consequences. And you saw that yesterday or the day before, I'm losing track, in the BLS data with respect to the latest productivity gains. And the only caution I'd make about that was that most of the productivity gains came from capital, not labor. So the returns are accruing to capital, not to labor, which is very, very dangerous moment in terms of where the gains are coming from AI. It's coming, it's just not accruing to labor.
00:08:20 · Akash Pasricha, Dan Shipper, Paul Kedrosky
[00:08:20] S0: So you put out this other note talking about how S and P five hundred's earnings growth ex MAG seven, [00:08:28] S2: the first [00:08:29] S0: time since 2022 outpaced the MAG seven. This, of course, keeps in mind that NVIDIA hasn't yet reported. Why do you think that is? What's the story here? [00:08:43] S1: Well, it's this... It's the intense cyclicality of of what... Of the build out that in the early years was relatively relatively straightforward to fund it all from cash flow. And so funding it from cash flow was was tremendous because I didn't have this overhang of debt obligations, which in turn was eating incrementally to my cash flows alongside of this what I was spending on data centers. So now I have an extra... I'm feeding an extra hungry animal, and the extra hungry animal is the... Is external financing of... Which is now something like 50 plus percent, 55% of the funding of data centers. So that's the problem. So increasingly, we have a lot of mouths to feed with respect to cash flows that are flowing out of these large and heretofore very profitable hyperscalers. And so that's beginning to cut into... And we're seeing this with negative cash flows almost across the board in the hyperscalers. And so it's really remarkable to see for the first time since the build out really began, that not only do many of them have negative cash flows, but for the first time, the rest of the S and P is outperforming them in terms of earnings growth.
00:09:39 · Akash Pasricha, Paul Kedrosky
[00:09:39] S0: Okay. Because you follow the data so closely, you mentioned the estimates around NVIDIA sales, how that could be affected. Have you seen anything around how CapEx growth could be affected? We had somebody on the show earlier this week talking about... I mean, the next year's CapEx probably unaffected. Maybe the year after, something like maybe 10%, 15% at risk here. Does that match what you're seeing, or what are you expecting? [00:10:06] S1: No. It's on that order of things. And I think, you know, Bloomberg Intelligence had a report out. I think was Goldman had a report out yesterday. And, again, all of this is belated because this was all kind of baked in the cake after the midterms, but nevertheless, at least they showed up to the party and began doing the numbers. And so, yeah, on the order of something, probably nothing obviously this year, but looking out next year into 2028, on the order of sort of 12 to 16% would be my estimate, which is obviously a titanic number given that we're dealing in trillions.
00:10:06 · Paul Kedrosky, Akash Pasricha
[00:10:06] S1: No. It's on that order of things. And I think, you know, Bloomberg Intelligence had a report out. I think was Goldman had a report out yesterday. And, again, all of this is belated because this was all kind of baked in the cake after the midterms, but nevertheless, at least they showed up to the party and began doing the numbers. And so, yeah, on the order of something, probably nothing obviously this year, but looking out next year into 2028, on the order of sort of 12 to 16% would be my estimate, which is obviously a titanic number given that we're dealing in trillions. [00:10:33] S0: Okay. And last question for you, Paul. If you just look at the story that these big tech companies tell broadly, they have been able to get away with this narrative that we will lose... We'll burn cash. We'll invest for the long term. This is the game we're playing, and we'll see the results three years out, let's say. Mean, in a world where they cannot tell that story anymore, because like you're saying, the estimates might come down, what do you think is the new, is this like a year of efficiency story that we start to hear again come through in the next year, or what do you think is the broader narrative they start to all adopt?
00:10:33 · Akash Pasricha, Paul Kedrosky
[00:10:33] S0: Okay. And last question for you, Paul. If you just look at the story that these big tech companies tell broadly, they have been able to get away with this narrative that we will lose... We'll burn cash. We'll invest for the long term. This is the game we're playing, and we'll see the results three years out, let's say. Mean, in a world where they cannot tell that story anymore, because like you're saying, the estimates might come down, what do you think is the new, is this like a year of efficiency story that we start to hear again come through in the next year, or what do you think is the broader narrative they start to all adopt? [00:11:13] S1: On premises. Everything gets pushed on premises because they're getting huge pressure from their customers to do AI on premises. So you start seeing these hybridized cloud models of saying, you know, it doesn't have to happen just in our cloud. It can happen on your premises. So it'd glad to talk about, you know, confidential data and secure computing and blah blah blah, all this geekery. But the bottom line is it's a response to this pressure that they need to find... Corporations need to find efficiency gains, large language models will try and do it by saying, hey. We can do that hybrid or on premises. And so that's gonna be, like, the litany for the next twelve months.