Diode Computers raises a Andreessen Horowitz-led Series A to use AI to design and manufacture PCBs in the US

edge so it wouldn't be apples to apples but maybe I got to check the number of moves. That's the move. Okay. Well, we have our next guest here in the studio. Let's bring them in. Good soundboard, Jordy. What's going on the stream? How you doing? Welcome. Hello. Awesome to meet you. Oh, thank you. Great to have you.

Uh thought I was worried we had some audio issues. Turns out we don't. Uh can you give us an introduction? What are you building? Of course. Um I am building data computers alongside with my co-founder Lenny. And we use AI to automate the production of printed circuit boards. We design them and manufacture them. Wow.

Very cool. Um what's the state-of-the-art right now? Like what who are your competitors? So there's a a bunch of companies that really like try to build circuit boards. Circuit boards are at the heart of every single electronic product that you have. You you name it like aerospace, robotics, um medical devices.

These are all industries that traditionally would take months um to come up with designs and like volume manufacturing of these boards. And so our goal is to help them um and like scale up to production right away.

Uh our competitors are folks like Altium, Cadence, but also like uh current manufacturers uh that are like building circuit boards right now. And what's the latest news? Do you have anything good for us? I'm very proud to announce that we raised a series A by Andre and Hwitz. Um there [Music] fantastic.

When when did you start the company? I don't know. I I think I missed it. Um exactly a year ago. Like we hit the one-y year anniversary 3 days ago. And wait, am I looking at a virtual background or is this actually your office? No, this is the Brooklyn Navy Yard. We build circuit boards in Brooklyn, New York. Cool.

What uh what is the what is the core value proposition? Speed, quality, time. How do you think about like what are your customers asking for right now?

So the real like thing that uh was very interesting to me is that we have fewer and fewer people that can build circuit boards at the highest um like level of complexity and these people are working at insanely good companies like Tesla, SpaceX, Apple, Meta and they're very very hard for other hardware companies to kind of like take away.

Um and I like uh I used to be at one of these companies. I used to work at Apple. I used to work on custom silicon. So, our goal is basically to democratize the way that we currently like build circuit boards and allow uh like from startups to large enterprises to take advantage of a very high-end type of workflow.

And the only way we could think about how to do it is to teach large language models to actually generate schematics. And so, like the the advantage and the difference is that we use code to build our circuit boards, which is something that usually like is not done this way. Circuit boards are very visual.

Like you can think about them almost as like Figma. Um well we what we do is like more like web design like CSS and code that then gets rendered into like an actual circuit board. Um it's called artwork. Can you uh can is it like okay to take a circuit board schematic from an existing product.

I remember we had this founder on I think it was the Madic robot and he had was was that right or we had some founder on who had printed on their on on their circuit board like hi Amazon because they knew that Amazon was going to take it apart and like try and understand how they were doing what there was I might be me messing up the company names um but it seems like there's some proprietary information in the in the like like yes there's amazing data and amazing engineers at these big companies they're making these circuit boards But they're probably not okay with you just taking that as training data.

Or maybe it's legal. I don't know. What's the deal there? You're absolutely right. Um I think that Maddic is a fantastic company. Uh they and Whoop both print on their circuit boards. Like don't bother copies. That's the one. Yeah. Whoop. Fantastic detail. I I really love it.

Um yeah, you can make that you can make that a feature. Yeah. We we'll do it for everyone of our circuit boards. um data is actually like the most important part of this. So like there is no re like easily available data for this kind of problem.

Like whereas on code you can basically take GitHub and kind of like start training your models on very high-end um like data that you can use. Um part of like what we are doing at Dio is we're rebuilding the data set from the ground up.

Um, and so there's a lot of like uh uh data annotation, data cleaning, like generating new types of data that we can use to improve the models.

Um, and we are about to announce like next week a partnership with like a very large uh like software open source project and and our goal is basically to contribute back to the community, generate this data that like we're going to use to train like better models without going and like uh touching on proprietary data like we own our own designs.

Um and so while we assign the IP of the board that we design for our customers to them, we usually retain the rights to the individual components that we generate and those are like a competitive advantage for us. We are basically building an internal data structure. Yeah. Why why do you even need uh real data?

I feel like you could generate all of this in simulation. Like I when I hear people talk about bio and saying, "Oh, we don't have a simulated cell. " I'm like, "Okay, well that makes sense. Like humans are really complicated. Biology is really complex. " But like we created circuit boards in the first place.

we should be able to simulate those. Like why can't you just simulate all possible circuit boards and then do some sort of search through that? Why do you need hard data? You're absolutely right. We actually do do that. Um like we have reinforcement learning algorithms. There's two types of simulation that you can do.

One is like an electrical simulation which you can absolutely uh like get right. It's called spice. Um but then like the second order effects of like physics are quite complicated to model.

And so there's a really um like there's a good saying around like very talented electrical engineers which is like you never trust sim like you basically like you can try simulating it but then like manufacturing something and making sure it works is the ultimate like simulation test like the boards may be manufactured in a different way.

Uh the the like board house may not like respect your specifications. Uh and so we 100% do simulation. That's how we bootstrap the process. uh you couldn't like build the library of data that we've built just by hand like it's physically impossible. Sure.

Um so we we use like simulation to bootstrap but then the ability to manufacture the boards is really what creates differentiation and like brings home the fact that these like circuits are actually working. So what are you guys actually doing on the manufacturing side? Are you going to be scaling manufacturing?

Are you just making prototypes for customers and then they take those elsewhere to uh scale? Um, excellent question. So, our offer to the world is come to us, bring us what you want to design, we'll design it for you, optimize it for manufactur it, and then scale it with you.

Today, we have only small batch assembly in house. Like we actually have a electric lab right here that we use to assemble like circuit boards like these. Um and then we partner uh with uh manufacturing houses for like the largest higher volume production.

Um but the like eventual future is we are going to vertically integrate this manufacturing. Um and like this is what we want to offer to companies. If you look at the current offer for printed circuit boards in the US, it's very very hard to manufacture them uh like within the nation.

It's usually like anytime you want to scale, you need to go to China. Um, and we think that the way to solve that is basically by making all of our design look the same from a manufacturability standpoint. So, we can generate volume and like bring it to all of our clients at the same time.

Uh, we also do things like automatically matching the parts and the ordering, which is something that normally it's done manually and it's incredibly like long and annoying to do. And so to like keep a very complicated process short, we make ordering PCBs from the US as easy as ordering them from China.

Not quite as inexpensive yet, but we are going to go there. Where do you think you'll you'll set up your actual manufacturing hub? You're you're said you're in the Brooklyn uh Navyyard right now. Uh do you think you'll go out elsewhere or do you want to keep it in the in the great state of New York?

So I'm a big believer in uh like decentralized approaches for prototype level. So we're actually opening an assembly shop in San Francisco for our clients there and we're thinking about one in Austin where we have like some clients that I absolutely love um alongside the one in New York.

Um and I think that longterm for scaling production the best strategy will be to colllocate the warehouses for the components and the uh assembly. So, we're probably going to open a larger facility in either Arizona um or Ohio most likely. Awesome. Last question for me. IMO gold medal.

Uh what was your reaction to the news? And is is solving IMO level math useful to you?

Do you think there will be transfer learning from that model into the circuit board design domain or are we in the era of spiky intelligence where until you RL on that particular problem you're not getting uh generalizable results I think even if we were in the era of like spiky intelligence on like RL on a specific problem like circuit boards are an absolutely like good way of doing this like conceptually a circuit board is orders of magnitude simpler than uh silicone design.

Like before like building circuit boards for a living, I used to do like custom silicon designs. Um and those have been like represented as code for years. This is 100% an RL approachable problem. Um I do think that the IMU news is fantastic.

Like clearly um the there are incredible returns if you do train on like a very specific set of data. I think that the the biggest mode will be in the data. Like what we consider to be our competitive advantage is not really like training new foundation models.

We actually rely on like uh foundation model improvements in the wild. And our like entire uh model is based upon feeding those models with better and better data and being able to fine-tune and reinforcement learning train specific parts. Like one of them is like the model that spots mistakes.

That's how you tell uh like a good generated design versus a bad generated design. And then you can grow the model in reality and physics. That makes a ton of sense. Thank you so much for stopping by. We will talk to you soon. Congratulations on the on crossing the one-year milestone. Fantastic.

Getting an A done as well. Great stuff. Thank you so much day. Talk to you soon. Bye. Cheers. And up next we have Elon who uh I met and Delian was mentioning uh was building a company in France. Now he's building a company in America. We're going to dig into the differences between the two countries.

figure out which country is the best. Get to the bottom of it. Hey, how you doing?