Conversation with Ayar Labs CEO Mark Wade on Silicon Photonics, Co-Packaged Optics, and Next-Gen AI Infrastructure

Ben Bajarin (0:08)

Hey, welcome to another episode of The Circuit. I am Ben Bajarin.

Jay Goldberg (0:13)

Greetings, programs. I'm Jay Goldberg.

Ben Bajarin (0:16)

Well, we've got some very special guests. We are joined by Mark Wade, CEO of IR Labs, and Jesse Leiter, who is, you can give your title, I think, Head of Finance, IR. Why don't you guys say hi real quick, and then we'll jump into it.

Mark Wade (0:33)

Yeah. Hey, everyone. My name is Mark Wade. I'm the co-founder and CEO of IR Labs.

Ben and Jay, thanks for having us on.

Jesse Leiter (0:42)

Yeah. Thanks, Jay and Ben. My name is Jesse Leiter, and VP of Capital Strategy and Investor Relations. The role has changed. It has strategic finance element to it before, Ben, so you're not completely off on that. But yeah, we're really excited to be here, and I think Jay and I have been speaking for maybe a little bit less than a year now. He'd mentioned the podcast.

Ben, I know you've been talking to Mark more recently, and so yeah, excited to be here speaking to your audience and on your platform.

Ben Bajarin (1:09)

Yeah. It's great to have you guys. This is a very hot topic. I know that you guys not being a public company, there's always people asking us like, hey, what's your thoughts on IR Labs? Like we're trying to get more information. So I think this will be super informative for our audience. And my context is, because you guys haven't heard me say this, but Jay and I joke about this for a long time, like The Circuit is slowly becoming more of a networking show, which is the irony of this is, back in the day, we all covered networking, but it was relative to Telco and it was super boring. Like, great, you've got more switches, SureFibers, great, this is all interesting. And when we started this podcast, it was largely compute, because what was going on, compute trends, and then network became this whole thing, right? Interconnect and whatnot. And so now we spent so much more time also talking, but it's relevant because the interconnect is a wall. It's a problem to solve.

It's one of those big hurdles to get over. In fact, both GPUs and even memory is scaling faster than the interconnect wall is, right? And so I think that's super interesting. So at a high level, like as we start, you know, Mark, why don't you just give our audience, in case they aren't familiar with you guys, just the high level pitch for IR Labs?

Mark Wade (2:31)

Yeah. So let me try to start with a one or two sentence framing, which is that large scale AI data centers are limited in their revenue scaling by the performance of the interconnect in their compute systems. And so that's a claim that says, hey, if we want to go beyond a path to continuing to scale AI data center revenues, to continue to grow the supply of token throughput, to improve the tokens per second per dollar per watt on these kind of tokenomic, and while simultaneously improving the performance of these frontier leading edge AI models and our architectures and things like the reasoning models, we have to improve the interconnect. So that's the kind of high level motivation, which is that we view solving this interconnect bottleneck as one of the core major hardware level bottlenecks that needs to be solved to keep the AI roadmaps running. Now, if we double-click into that a little bit, the bottlenecks are coming from these electrical interconnects that exist in the so-called scale-up fabric. Scale-up fabric connects a number of GPUs and AI accelerators together, and you've got this issue in electrical interconnects, which is that the faster you want to go, so the more data you want to send between these GPUs and accelerators, the shorter distance you can go. So we call that the bandwidth distance trade-off.

The only thing that we know how to solve this, or the only way that we know how to solve this from first principles from the grounds up, is to swap over to optical connectivity. So that's where we say, hey, we've got to get a new generation of deeply integrated, you know, CMOS scale optical connectivity into the scale-up fabric, solve this bottleneck from the ground up, you know, where the real physics problem is happening, and then we enter the promised land beyond.

So, that's the quick snapshot of what IR Labs is trying to do.

Jay Goldberg (4:44)

So, I've covered optical for a long time, all through the telecom ups and downs, and ups and more downs.