00:00:00 · Ty Morse, Chris Power
[00:00:00] S0: Today, I'm sitting down with Chris Power, the founder and CEO of Hadrian. Hadrian has opened six factories since, like, 2022, latest one being 2,200,000 square feet. I wanna start this off with what does it actually require to scale manufacturing in The United States? [00:00:16] S1: A lot. Put... Putting aside the raw horsepower and pain, I think actually it is GPU style stations that can be modularly configured inside a factory, which gives you standards to automate against, and then a very complex organization that can build real manufacturing autonomy. And the third thing is a very careful balance of physical AI software and a new American workforce. [00:00:48] S0: Do you wanna go into what does it mean to have like a GPU as a factory? [00:00:54] S1: So most production lines are set up linearly like a CPU. Like, you can do more things, but they're running linearly. So, you know, there's there's like a sixty minute component that then goes into a two hour assembly process, and you're kind of just running things like this linearly. The GPU is more like everything is so standardized and automated that you can run a high mix, low volume factory, like, as as efficiently as a low mix, high volume factory and manage peak capacity like you manage GPUs. It's very complicated to do, but in... It's the only way you get like real high mix low volume performance at a factory scale.
00:00:54 · Chris Power, Ty Morse
[00:00:54] S1: So most production lines are set up linearly like a CPU. Like, you can do more things, but they're running linearly. So, you know, there's there's like a sixty minute component that then goes into a two hour assembly process, and you're kind of just running things like this linearly. The GPU is more like everything is so standardized and automated that you can run a high mix, low volume factory, like, as as efficiently as a low mix, high volume factory and manage peak capacity like you manage GPUs. It's very complicated to do, but in... It's the only way you get like real high mix low volume performance at a factory scale. [00:01:42] S0: One of the things that you talked about earlier today was this idea that you have to basically... If you want a contract, you have to kind of build ahead of the contract, and you have to build the factory to show that you're serious enough to actually, you know, own it. How does that work?
00:01:58 · Chris Power
[00:01:58] S1: Well, let me give you an analogy. So AWS, you know, developed the largest cloud computing capacity ever, and then and then second started selling it to the market. But they already had the largest, yep, data store for Amazon engineering and Amazon retail as the first customer. So it's pretty believable that like, hey, you're running AWS internally for Amazon. You can start selling it to like Stripe. Contract manufacturing in The United States is this big chicken and the egg problem where it makes sense to anybody that, hey, you are designing a new low cost interceptor. Why would any good startup build their own vertically integrated factories at a worse unit economic position than Hadrian? It doesn't make sense. What what does make sense is that there is no starting point of capacity. So Shenzhen, for example, if you start today a new humanoid robot and you have a circuit board, you're not gonna start your own circuit board factory because there's a bunch of capacity so it's obvious to you that you tap into it and you can trust it because it's at scale. The United States has no contract manufacturing capacity. So if you want anyone to take a big bet on you, you have to basically build at least 50% more like 1.2 x the percent of your future sales capacity ahead of having those contracts. Otherwise, you're gonna be so far behind the eight ball of both optics and execution. Execution being the most important thing because manufacturing, what we sell in manufacturing is trust that you're never gonna scare. Okay. So we're building a data center. Okay. We've never built a data center before. How how much easier is it for you to rent GPUs from me in my data center when it's already half built, as opposed to like on a slideshow. So American real industrialization is largely the former problem, not because people are bullshitting, just because the capacity doesn't exist. So you have to build it ahead of demand, which takes a lot of balls and capital.
00:04:21 · Ty Morse
[00:04:21] S0: There's this funny thing that you mentioned earlier today, which was this idea that to actually get some of the contracts that you're going for, it just takes two years from the start line. And I think this is kind of unique where you both have to project out here's what the contract might be and build the capacity for it and also not know necessarily when it's going to come. How do you kind of like math that in your brain?
00:04:45 · Chris Power, Ty Morse
[00:04:45] S1: So so the the math is burn rate and execution risk. Execution risk is the easier one because once you understand that any factory build out or, you know, I I buy a three d printer or CNC machine or robot. There are so many things that can go wrong that if you say it's gonna take eight months to get here, I know from experience that it's gonna be somewhere between eight and a half weeks to ten and a half weeks. Point number one. Point number two is if I told you to build a drone factory and a missile factory, you would think that that is binary risk in the sense that to win either of those contracts, you would have to deploy CapEx and human capital, but, like, both to have an equal chance of winning those contracts. If... The big if. If you have a flexible workforce and flexible CapEx via software, the risk mitigation you can do is that 80 of your CapEx, you can switch from program to program, like data centers. You don't know whether the consumer ChatGPT is gonna take more demand than anthropic enterprise compute, but you do know that they're gonna share the same GPU cluster. You can get manufacturing 80% of the way there to the flexible risk grade. If you are detail oriented enough to break down every product into the manufacturing method, And underneath the manufacturing method, the specific CapEx, which is extremely complicated. But if you are capable of doing that, then you can make bets that are not risky to you that are extremely risky to everyone else. [00:06:39] S0: Do you wanna go through, like, one of those bets that is extremely risky?
00:06:39 · Ty Morse, Chris Power
[00:06:39] S0: Do you wanna go through, like, one of those bets that is extremely risky? [00:06:44] S1: If you were to build a... If you were to just look at building a new $300,000,000 casting facility and you had no contracts, And the CapEx takes two years to get here, which means you're in production in year three, best case. If you are looking at that linearly for castings for turbine blades for argument's sake. You'd have to be damn sure that you... The demand is still gonna be there in year four and... Okay. If you break down the CapEx that is linked to the process that is a subset of the output, output, which is a... Let's call it a Monel casting. And you do the same thing for munitions, 40% of which are castings that are never gonna be additive because it wobbles too much. And there are some unique bits of CapEx for sure, but 80% of it can be shared. And you are willing to have a 20% larger CapEx bill that buys you a 100% flexibility, what you're actually doing is deploying 300,000,000 CapEx against three demand signals, not one linear one. But that is down to the level of detail of, hey, both tonnage of pour by alloy and for, like, a molding line, you know, envelope size of average component by floor space. All this can be done. It's just stupidly detail oriented. It's very precise.