The Information's TITV
00:55:44 · Stephanie Palazzolo, Nicole Su
[00:55:44] S0: And I feel like, you know, speaking of physical AI and the world of, you know, robotics and all that craziness, I feel like there is a bigger debate now between, you know, like there have been some studies that have shown that actually the general language models are actually pretty good at maybe directing robots better than people expected. I guess I'm curious, do you feel like there's still a strong argument for, you know, robotic specific models versus everything just being controlled by, you know, general language models? [00:56:16] S3: I think the reasoning part, I think, is totally transferable. I think it's more about interaction with the physical world and especially when it comes to specific form factors, because I'm seeing really good progress when it comes to robotics model architecture evolution, right? We started with the BLM and VLA and our world models allow people to predict the future, but at the same time, like we are missing previously, like inputs like tactile data, but I've seen some of these leading research teams have started publishing paper about how they actually were able to integrate all this tactile data into their foundational model that has all the elements of LLM for reasoning, L VLA for, like, interaction between vision and action, but also word model for prediction. So I'm seeing... I think it's gonna be a combination of all these model architecture to work together to unlock the physical AI world. Yeah. So kind of like
00:57:07 · Stephanie Palazzolo, Nicole Su
[00:57:07] S0: what you're saying, maybe today's reasoning models would be good at deciding what next step to take. But when it comes to, you know, how hard should I grip something or like what does it mean for a robot to pick up a glass of water off the table? That is something that's so niche that it's not really captured yet in language models, it seems. I see. You know, switching gears a little bit, we talked a bit about apps and physical AI. You know, one area that's gotten a lot of investment is infrastructure, data centers, chips, that whole area, right? So it feels like up until maybe the last couple of months, that was a very obvious area of investment. But I think maybe more recently, there's been questions about, you know, how much debt can these companies take on? There's been IPOs getting delayed. [00:57:48] S3: Do you [00:57:49] S0: feel like there's a turning point coming for these startups when it comes to how enthusiastic investors are about them?
00:57:49 · Stephanie Palazzolo, Nicole Su, Jessica Lessin
[00:57:49] S0: feel like there's a turning point coming for these startups when it comes to how enthusiastic investors are about them? [00:57:57] S3: I'm definitely seeing a change in terms of valuation levels for these type of companies. Obviously, early days when there is definitely secrecy in terms of the investable opportunities in the public market, some of these companies, you know, definitely rode the boom of AI infra build out. I think now it's really more about the heightened or increased investor awareness when it comes to the risks associated with cost of capital as well as the ability to execute. So I think many of these businesses, the fundamentals didn't change, right? But it only changes how investors started changing their way to value these companies. [00:58:31] S1: And [00:58:31] S0: I guess what are investors so worried about now when it comes to these, you know, I mean, there's a whole bunch of them. There's obviously the Neo Clouds. There's the companies that are building the data centers. Like, is there something specific that they're worried about now versus maybe a couple months ago?
00:58:45 · Nicole Su
[00:58:45] S3: I think it's what's really interesting, like from my seat, I'm definitely seeing I think we all see there is a huge demand for compute. Like, we don't need to debate that. And then there is also a huge wave of new companies being formed and trying to become neo cloud and who all need a capital injection. So you can see there's a big gap of, yes, there's a huge need to compute, in shortage for sure. And yes, there's a bunch of people who really want to start building out this compute. They have access to a specific piece of powered land or they have something they have unique access. But the really missing link is that we don't really have the capital today to breach this. I think investors historically has become very used to financing customer offtake contracts. The reason why they are willing to lend is because of the quality of their customers, right? Most of the time, these are hyperscalers. However, we're in this next wave of, hey, a lot of AI compute demand is coming from AI native companies or new labs where like there is no really truly rated credit for any of these companies. So I think lenders are definitely become more and more cautious when it comes to companies who are actually serving that bunch of companies. So I'm seeing that, I guess, the missing link is very much like the investors' willingness to back these type of companies and the level of risks they are willing to take, and then the costs come come with that.
01:00:05 · Stephanie Palazzolo, Nicole Su, Jessica Lessin, Anne Wojcicki
[01:00:05] S0: I see. So it's kind of understanding how risky are these businesses because we don't we don't know how risky their customers are because these are kind of new types of companies that haven't really existed before. Mhmm. Do you think we're gonna see a spat of fire sales or some of these companies going bankrupt in the next year? [01:00:24] S3: I think it's... That's a bit [01:00:25] S1: of extreme. Like I said, I don't [01:00:27] S3: think the fundamentals have changed. Right? There's still huge demand for compute. There is a shortage, and some of these public companies are very well run. I think that the problem is more about the group of companies I just talked about earlier, like this newer neo clouds who are trying to look for financing to really build on the next batch of compute capacity. I don't think anybody is going to go bankrupt because the compute they currently have is more about, like, can they even find enough capital to build out what we need for the future. [01:00:56] S0: Yeah. That makes a [01:00:57] S5: lot of sense. Mhmm. [01:00:58] S0: You know, obviously, the big news of this week was the Anthropic s one that dropped on Monday. I don't know if you can speak to it directly, but, you know, looking at that, I was... It was very interesting to see investors' reactions and the types of metrics that they're really paying attention to. I mean, generally, like what are you hearing from investors in terms of what metrics are they looking for in these S1s of these big model developers that are heading to the public market soon?
01:01:22 · Nicole Su
[01:01:22] S3: Yeah. I can certainly talk this in a broader general terms for Frontier Labs. Me personally, right, if the first number I would focus on is very much gross margin, but I also understand, I need to understand what does gross margin really mean for these companies because the definition could be very different. I think we had this huge debate around like the difference between two Frontier Labs and how they define revenue, right, that obviously would translate into gross margin. But then I think more importantly, would just go straight to, if I want to cut out all the noise and differences, I would go straight down to free cash flow, right? How much cash are these guys burning? And it because if they are bringing less cash, is it because they underinvested previously, or is it because truly they have a more efficient model when it comes to... Business model and comes to finance their growth? So it's very much like, Hey, let's look at gross margin, how healthy it is, but obviously be very clear on differences in definition, at same time also looking at cash flow and see, you know, what's the burn and what's not just the today, but the future and then the likelihood of, you know, getting to profitability in a more sustainable way. Totally. And, you know, we've read