Big Technology Podcast
00:14:09 · Ranjan Roy, Alex Kantrowitz
[00:14:09] S1: This is exactly my argument. I I appreciate this. You just laid it out even more eloquently in an instructed way than I did. I mean, this is it. It's it's exactly the... It's a narrow set of companies that the entire trade is built around and anything... Again, we've talked about this even on Anthropic specifically, we're gonna get more into the current state of the IPO, but everything assumes perfect execution. That's on their own internal businesses, that's on growth, that's on revenue, that's on unit economics, but that is also on the data center story to be the infrastructure for this gigantic story and all this promise. So everything depends on perfect execution, and we're seeing even this little hiccup or... Can actually completely derail that story. [00:15:01] S0: Big hiccup. [00:15:02] S1: Mean, yeah. Yeah. [00:15:03] S0: Right. But this is what I'm this is. So this is why I'm going back and forth on it. And I haven't fully made up my mind on it yet is to just go back to is it really the end? Like I you know, I understand that I understand why the market will see a sign like that, and they wants to buy in. But I don't know if it's really the end or not because of all the different times we've seen this negative thing happen and it not actually manifest
00:15:26 · Ranjan Roy
[00:15:26] S1: the way people But so I'll push back on... In the past, the whole, like, idea around regulation and the tech clash was very limited to the actual singular companies themselves. And, like, can the FTC go after Meta? Can... And again, we've seen, I think what was it, New Mexico going after Meta for $25,000,000,000. Like we've seen more and more things happen. And again, Lina Khan and the FTC, there's certainly like actual action taken against large Big Tech. But by the same token, these companies have gotten significantly more powerful than they ever were before. This is a little different to me because this is the regulation is or the government involvement becomes so much more small and targeted in a locality and whether they're gonna actually allow a data center to be built. And that's gonna happen on a one by one basis across the country. So it's not these massive, like centralized federal and central company issues of whether Meta's lawyers can beat the FTC's lawyers. This is gonna happen throughout the country and I think that is what makes it far more unpredictable and actually in a interesting way makes it more likely that something can actually happen to these companies.
00:16:48 · Alex Kantrowitz
[00:16:48] S0: Now, here's some so as I've talked through this, let's talk through the some more evidence that I think would agree with your point. These are major planning projects, right? To be able to build a data center, you need to go through land approval, you need to buy the land, you need to find the land, you need to hire the company that's going to go build it for you if you're working with subcontractors. And then remember, we're dealing with, you know, power issues. So you have to figure that out, you have to figure out who the electricians are going to be who the tradesmen are going to be, to be able to put this together. So, you know, as much as you could say, and as much as I've brought up this idea that you can shift the data center from let's say, Ohio to Wyoming, and maybe you can. You also run into the fact that you're gonna have to do this from step one. And so, you know, the question is, you know, how many years do you end up delayed because of it? Now, let me bring something up to you, Ranjan. Some people have said that because of this, we're about to see the peak of AI CapEx. I just don't see that being true. I just think it might be drawn out across more years, than it is, than we've seen it today. And it it may... You know what? The funny thing is, this might be the thing that paces the frontier. You know? Well, no. But the At the end of the
00:18:06 · Ranjan Roy
[00:18:06] S1: day That's what I... That's why I'm not over... This does not make me more or less bearish or bullish on AI in general as a technology. I mean, we speak, I... Our our recording almost stopped because I have clawed in the background editing a separate video that I'm working on, and it's... I just got a notification from Riverside saying it's constraining my compute on my MacBook. And I can only imagine the amount of compute that that consuming. And like those kind of token heavy use cases are gonna continue to explode, and I strongly believe that. And... But it's not gonna happen in a straight line over the next two to three years in a way that a very small number of companies will actually be the beneficiaries in a way that has been predicted so far. And I think that that's gonna be the... That is sinking in and the market is bouncing back a bit today, but these things never go in a straight line down. Even the trade war and liberation day was like two days down, one day up, two days down. So so I think we we... Today in the last two or three days is not gonna be fully reflective. This is gonna take at least a few weeks to start to get a sense of like, how much has it set in, but I think it has set in, I definitely
00:19:27 · Alex Kantrowitz
[00:19:27] S0: do. You know, it's interesting. So if you have less room, they have been trying to expand, like crazy. I mean, when I was with Sundar, for instance, he said, I wish I had double the compute that I have today. And that's the case for with all these companies. They've been trying to expand because they're trying to do two or three things at the same time. And now they'll have trade offs among those. Right? So if you have the compute that you need, what do you do? You serve the current product with inference, right? So what you're doing with Claude, that is compute intensive, that is coming from one bucket of compute. While you're doing that, you have to already train the next models. Right? So that's another bucket of compute. And as you do both of those, you also have to do the proper test to make sure those models are safe. So you have product user inference, next generation research or training, and then, the security thing. And what happens just going to throw this out at you. Actually, this is very interesting trade off. If those data centers that were anticipated can't be built, what happens? Where do you make the compromise? Do you make the compromise on security to make it on inference, so you can make less money, and you have mad users and Anthropic will wish that will risk them going to open AI? Or on the third, the third bucket? Do you not get the chance to train the next models? And if you can't train the next models? Remember, this is a story of potential. So the potential of RSI or AGI or super intelligence, you compromise on that. Where do you compromise? It's not clear to me that there's an easy bucket to sacrifice.
00:21:02 · Ranjan Roy, Alex Kantrowitz
[00:21:02] S1: No, there's definitely no easy bucket. Because as you said, I think from a pure operational standpoint, the training of the next frontier model would be where I would sacrifice and just kind of build out and as I'm on team product, think that always makes more sense that there's still so much opportunity and room, but the investor story is not that. The investor story is the potential for the next frontier model and the next frontier model after that. So if you're solely trying to please investors and then you start falling behind on the benchmarks, it's a tough decision. Maybe, you know what, maybe here's my optimistic look at this. The same way innovation on the actual efficiency of compute side coming out of China and other like other areas where compute was more constrained was kind of the way things were approached. Maybe we start to see that in America right now and this election actually brings efficient computing innovation, and that's my platform. That's what I'm running on two years from now. I missed this one. [00:22:11] S0: Ranjan twenty twenty [00:22:12] S1: Efficient computing Well, [00:22:13] S0: that is what Sundar that is what Sundar said. He said the more... If you're a little bit constrained, you have to be creative and not just brute force it with compute. Yeah. So So I guess you're [00:22:23] S1: with him. That that can be our optimistic look at this, which is still not good for a lot of the economics of how the trades that have been built up. And and that's the irony