00:22:02 · Chris Power
[00:22:02] S1: I'll break the answer into two, which is factory development and then like factory design. And the first one is really easy, and I'll tell you a story. So one of the big things we do at our... One of the parts of Opus is called Mother Brain, and effectively we take over all the CAPEX software and reprogram. And you need some decent a... APIs into the CapEx, or at minimum you need a text file that you can read. But it's brutal engineering. Okay. So you're a small startup and you've got 6,000,000 in CapEx to spend, so we buy two types of machines. Makes sense. And you call the president of the German capex vendor, because your team send them an email saying, can we please have the API documents? And they think you're a hacker. Because no one's ever asked for the API to the machine before. And you explain it to them, and they they give you the API documents, and you read them and it looks all fine. When you get to the factory floor, what ends up happening is both brands of machines, the API that reports off the hardware is completely different, so it gives stochastic and consistent results. And one of the CAPEX brands, you hit all the APIs just to test it, and 50% of the API calls for a per... Return like niche implemented yet in German. And you realize that actually what you're doing, and you know, you you try to build software on top of it, and one of them takes twelve weeks, and the other one takes two weeks. So what you realize is actually what you're doing is dealing with a bunch of unknown unknowns. So you build every single plan against the fact that everyone's gonna fail against you, And you start building capability engineering systems around it. So you never expect like the PLC of the world robot until you've literally integrated it. You have no idea what reality looks like, which is fine as long as you build boatloads of hedges into your timeline and you plan around the fact that the industry is fucked. In terms of factory design, the biggest lesson is that people design factories around the success case. And the success case is like 90% yield, smooth flow, etcetera. I'm not talking about like maintenance downtime. Just one of the things we do which I think is you can only do with software powered factories, but makes complete sense, is you have a ton of slack capacity that's just based around failures.
00:24:32 · Chris Power
So the analogy I'll use is like, okay. So you've got a a five lane freeway and a bunch of Waymo's, and some of them are gonna go screw up at some point. Right? And you've seen traffic jams. Right? You hit the brakes, everything backs up. What you actually wanna do is as soon as someone has a blown tire, there's a... There's an off ramp or there's a a slip lane on the freeway where they can just wait, but it doesn't hold everything back. Most factories are designed with the CapEx, the people, and the... Let's call it business processes that if one of the cars blows a tire, there's no off ramp. And I'm not saying you trash the car. You need to send it back and fix the tire, but you're sacrificing this, like, kind of industrial engineering efficiency for like, you know what you would have wished when that car blew a tire? Man, I really, really wish I had an off ramp so I could put that to the side and fix it while not holding up five lanes of traffic at rush hour. Now what it requires is that every bit of... Every station inside the factory is so standardized that you can switch things out like a GPU, and that you have high enough efficiency that it doesn't hurt your margins or whatever. But that is that is one thing that we've learned very well. The second thing we've learned is that there is a limit to software complexity that doesn't right match the physical world, and I'll give you an example. So in in manufacturing, you've got, yep, production and inspection, and you've got some... Let's say they're components. And they're components that have stable processes at some yield rate. And the the important thing about yield is you can predict it. Because if you can predict yield and it's 20%, you just make 10, you get eight. And and it flows through. And then you have new product introduction where you're you're machining something, you're welding something, you're inspecting it, but you're a 100% sure that that's the right process. And you, you know, you you need to make one to... Okay. Now you end up with this big queuing system. So something goes from manufacturing to inspection. And what you wanna do in new product introduction is have the highest rate of the the lowest possible impedance between inspection data and manufacturing, because the thing that creates cost or late deliveries to the customer is the...
00:27:02 · Chris Power, Ty Morse
Not the cycle time of the manufacturing and the inspection, but the the back and forth. So if you make something I want inspected in an hour, so the feedback loop happens like this. So in software land, this is pretty easy like queuing theory. So if you scan an MPI component into the queue, it goes to the top of the inspection queue, and there's a... You know, this this is very easy software. In reality, it never works like that. So what you do is you have a physical set of CapEx that's just for the fast line. And there's this odd property where that's not just because humans are dumb. It's not because the the... This... It's... You literally need to separate your fast response part of the factory. It can be... There can be a rope between it. And the interesting about that is software gen... Software tends to design itself towards a physical layout of systems. So you need to kind of like bake the factory to the natural software workflows, and that you wanna match that as close as possible so it's interpretable. You know, and the rest of it is like, you know, you just you just gotta fucking go like hell. [00:28:14] S0: On kind of the same thread of sprinting through one way doors, what's been the best example of something that went absolutely horribly wrong should have killed the company. And somehow you're able to, like, pull a rabbit out of a hat and stop it.
00:28:28 · Chris Power
[00:28:28] S1: Okay. So one very clear example is this was a couple of years ago when we had we had a lot of money for a series a company, but in aggregate, a lot... Not a lot of money. And we're landing we're landing CapEx two months before customer contracts hit, which is what you do. You know? Ideally, you know, I I owe you a bit of compute, and I switch on the GPU two months before. You don't want it a day before, but... Yeah. And we we we bought all this new equipment, incredibly reputable vendor. And about a couple of weeks in, four months before we had a bunch of contract deliverables for a very early customer that was very important. But more importantly, we were coming up to a fundraiser, you couldn't miss a delivery quarter. Because the whole bet is, you know, obviously the thesis is real, but can you execute against it? And two of my top guys came running up to the office and they said like, I won't name the vendor, and there's, you know, there's some rust. Okay. And we think it's terminal. Okay. And we've called the vendor and they think, you know, it's not a problem. It's not a problem. Now our guys are like the best in the world. So you gotta assume that they're like, yeah, days at least, if not weeks ahead of, you know, the vendor's self management of... And it turns out that within within forty eight hours, we did assess the fact that a subcomponent on these in these bits of CapEx, they they had switched out the subcomponent vendor in this country that produced the CapEx. So you know, the car analogy would be we kept buying Toyota Camrys but they'd switched out the tire vendor, which is fine. People people do that all the time. And we worked... And because of the serial numbers, like we had 40 machines and the serial numbers were like 600 to 800 which means it's a bash problem. Right? It was propagating really fast. So we real... And, you know, the response plan from the vendor and the warranty was basically like, well, we'll get it done by Thanksgiving. And Thanksgiving would have killed the company because the the the fix it plan was we're gonna deconstruct machines one by one, we're gonna replace this stuff. They didn't really understand that you you had to replace it completely because it's rust, know. And the the guys were right. I spent forty eight hours on it myself and I I agreed with them.
00:30:53 · Chris Power, Ty Morse
And then I spent a week straight yelling at the president of this company, and told them that I would... If they didn't fix it in the next four weeks, the company would die, which was true. That's not a, you know, bullshit statement. Because the first thing is they didn't realize how bad the problem was and what the impact would be. You know, there's like eight weeks of downtime, and then there's... Eight weeks of downtime kills the revenue curve of the company because it it's like missing a sales quarter in SaaS. Like, suddenly your burn rate quadruples because you got all these salespeople and no ARR. And then force them to fly in 30 engineers from this foreign country and get the whole thing done in four weeks through pure just like, you know, I was gonna headlock these guys. So that's a real operating problem that would have killed the company. There are there are many others. It happens about once a quarter, But it's manufacturing. [00:31:52] S0: This is the game. What's the most recent company killing event? [00:31:57] S1: Nothing nothing in the last three months. But there's one every month. But I I think this is the point, is you're making... You're you're hard rolling sixes every quarter, and there's always a company killing event once a month. Because it's a matter of fact [00:32:12] S0: I recently interviewed Keller from Zipline. Great. And... Yeah. He's great. [00:32:16] S1: By by the way, I I I think Keller is probably the best deep tech CEO. Top five.
00:32:16 · Chris Power, Ty Morse
[00:32:16] S1: By by the way, I I I think Keller is probably the best deep tech CEO. Top five. [00:32:23] S0: I was recently interviewing him, and his entire timeline was basically, you know, like, five years from now or something. He wanted to have a million deliveries a day on his network. And then Uber came in, like, a month before the interview and basically pulled in the entire timeline by two and a half years and said, I'm gonna [00:32:43] S1: take [00:32:43] S0: all of [00:32:44] S1: All your... Yeah. [00:32:44] S0: Year and all your capacity. And this completely radically changed, like, how they are thinking about scaling their business because that's just one customer, and there's gonna be a hell of a lot more in two When and a half you're thinking about not only, like, pulling in timelines, but also resetting your expectations upwards [00:33:01] S1: Yep. [00:33:01] S0: So you go from, like, your projections were, I'm gonna make a billion dollars in revenue three years from now, and suddenly you're, like, actually, I'm gonna make, like, 5,000,000,000 [00:33:08] S1: Yep. [00:33:08] S0: Three years from now, and I need to make a billion next year. With these, like, long lead times on not only being able to, like, buy CapEx, but also get contracts for that. Yeah. How how do you, like, manage that?