Atomic Machines CEO on Developing ‘Micro-Machines’

The company was founded by Jeff Holden, who was an early employee at Amazon. He was Uber's first chief product officer. Pleased to say he joins us now to talk a little bit more about exactly what he's trying to do. And, Jeff, I do want to kind of get to this idea. I mean, we talk about how I it's kind of really, uh, accelerated the creation of digital products itself. What are you trying to do with regard to creating physical products? And how is this manufacturing process different than what other people are doing? Yeah. No. Great. Great question. Thank you. Thanks for having me. Um, uh, the the matter compiler technology that we developed is, uh, is a completely new kind of factory. So, uh, what it does is it takes digital code as the input and outputs micro machines. Um, these micro machines, there are many different applications. And we have a first micro machine called Prime Switch, um, which I'll talk about, uh, presumably uh, in a bit. Um, but uh, the, the technology, um, is, you know, it's new to the world in the sense that we can build machines at the scale that nobody else can build them. Um, you know, chip fabs, chip, chip fabs, you know, uh, manufacturing facilities make, uh, you know, printers, you know their circuits, right? They can make circuit patterns and so forth. We make machines with, you know, actual moving mechanisms and at this very micro scale. So give me a sense here. I mean, so, I mean, uh, so I take this to the device, uh, I tell it what I want to do. I mean, what is the software do itself and what does the engineer still have to do? Him, her or herself here? I mean, what is the interplay between, uh, my input as a human being and what the actual machine can do? Yeah. Great question. So there's you can think of manufacturing as having two sides to it. There's a front end design problem and then there's the back end manufacturing problem. So the matter compiler takes, uh, digital code and turns it into a machine. So that's that's the back end part that becomes real time now. And you could make a wide range of machines by giving a different digital code. Now the question is where does that digital code come from. That's the design of the machine. And a person can create that. Or an artificial intelligence can create that. Or a combination can create that. It's basically expressing a sentence in the language of the matter compiler that tells it what to make. And so yes, we're building the front end generative AI as well. Um, that will actually do the design such that you can go from the idea of something to the thing in your hand in real time. We call that prompt to product manufacturing. So I mean, so I mean as a you sort of find ways to do this. I mean, obviously there are a lot of layers to this. I mean, is this going to be capable of designing something that a human engineer would not have actually been able to create themselves? Yes, I expect so. Whether it could do something that engineers could never get to, I don't know. But, um, but it will certainly be able to do it much faster. Um, and we were building the system in a way, um, that, uh, that accelerates as the system makes things, actually learns how to make things faster. Um, so, you know, so we, we expect to be quite surprised by what the AI will come up with. And ultimately we're excited about is the innovation explosion. This will open up for people. Um, letting anyone make a device who has a really good idea. Well, how many devices have you made? Or fundamentally different devices, I should say. Have you made a using this technology without substantial retooling, physical retooling afterwards? Yeah. So the matter compilers designed to do it, to behave this way. We're focused on just prime switch right now. Um, prime switch is a is a is a great example of the kind of device because it has a bunch of materials in it that you can't, you know, you can't, uh, you know, put into a chip fab, for example. Um, it has moving parts, it has assembly. They have to be joined together. So it's using all the capabilities of the matter compiler that would be needed to make, for example, a coolant device for that could bond straight to a chip and pump coolant. Um, you know, at chip scale or a micro robot that would enable neurosurgeons to access the entire area of the brain. So far, we've been focused just on prime switch right now while building, and that a compiler would be able to do the full gamut of micro machines. Well, so give me a sense, though. I mean, like how like what would be the build time and what's the yield out of it? Give me a sense. Yeah. Yeah. So you know so yields today are you know kind of in the in a double digit percentages like high double digit percentages. Um but but the really interesting thing about that is, um, is that the eye wraps around the manufacturing process. So the yield, uh, you can split the yield in two kind of the actual process yields and the and the kind of the expected yield. And what I mean by that is every single operation in the compiler goes through precision metrology. It's measured. And that is fed back to the I and the I learned from that and the I can correct errors. So we have a different kind of process control from normal manufacturing, typical manufacturing which usually uses statistical methods. We're using deterministic methods where the I only lets a piece advance if its actually correct. So the actual yields will come out of the machine will be very close to 100%. So for the power relay for for data centers that you have now to give me a sense. I mean, so where is that sitting? Is that, uh, inside the server? Is it like a power distribution unit? I mean, where is it? In the chain, if you will? Yeah. That's great. It basically belongs in, uh, every tray and every rack and every row and every data center. Um, what's happening is that data centers are, you know, as the AI chips are advancing, um, you know, the amount of power they need is advancing as well. And the and there's also a push to bring chips closer and closer together to optimize performance. Then you take those two things working in conjunction with each other, and you drive a lot more power into the rack. And with that power comes, you know, huge losses. As you know, you're trying to switch that power because switches are all over the place. Um, in the eye data center is you're trying to switch that power, you're having losses. And also you need switches that are very, very fast to handle the kinds of faults that can occur in these very high power environments. And that's the missing ingredient. All this time has been, um, this very fast, this very low resistance, uh, switch the prime switch now solves. So. But I mean, don't we already have kind of these power switching. Uh, doesn't that power switching technology already exist out there? Power switching technology exists. Um, but there's two forms of it. There's semiconductor power switching, and it's a semiconductor, not a conductor. So it ends up bleeding off a lot of its energy as heat. And the amount of heat that it bleeds off goes up with the square of the current going through to the power going through it. The other kind is electromechanical, and our device is also electromechanical, but but it's tinier than any electromechanical switch, uh, in existence. And so it can move much faster. Normally electromechanical switches are very slow, so they're low resistance, but they can't react quickly. If something goes wrong, there's a fault. And so this is a is a as we're going to 800 volt DC and AI data centers. This is becoming a very acute problem. And prime switch is becoming really critical. Do you actually have customers that are ready and willing to actually put this into actual production? It's one thing to say, okay, we have this technology, it works. It's another thing to actually put it into mission critical infrastructure and use it. Absolutely. Do you have. Absolutely. Yeah, we have multiple programs going on with customers right now that are incorporating Prime switch into equipment that will go to data centers.