Lightmatter is a deeply technical product. In layman's terms, what do you guys do?

We started the company in 2017, and the goal at that point was to figure out a path toward making sure computers get faster and more efficient, even in a time period where Moore's Law and Dennard scaling, these economic and energy-based scaling laws, are completely dead for standard electronics. Our take on how to drive computing forward is through using light. So at Lightmatter, we build silicon photonic products, and today our major focus area is linking these massive AI data centers together. In these data centers, you look at xAI, 100,000 GPUs. How are they connected? That's the kind of thing we work on. We link together ultra-fast GPUs using light.

You just unveiled the Passage M1000 and L200, two photonic interconnects aimed at solving one of AI's biggest bottlenecks. Can you break down what those new products unlock and why they matter right now?

If you walk into a data center today, let's say you go into xAI or you look at the deployments that NVIDIA is doing, what you'll see is these very small, what's known as scale-up domains. A scale-up domain is a tightly connected set of GPUs that are used to do a computation. When you do a computation within that group of up to 100 GPUs, it's very fast. But as soon as you go outside of that little mini brain of 100 GPUs, the speed tapers off dramatically.

L200 and M1000, two products we launched earlier this week, have the goal of making these entire 100,000-, million-GPU supercomputers act like a single brain: rather than many, many brains that compute in parallel and then synchronize results, one giant brain.

L200 is the fastest 3D co-packaged optics in the world. It's 64 trillion bits per second per chip, and it's really exciting. It's the world's first 3D co-packaged optics. And we didn't stop there. We were looking at the ultimate future and where photonics goes.

Passage is generally a technology that lives underneath GPUs and switches. How many components inside the package it lives underneath defines whether it's one of our 3D CPO products or one of our M-series substrates. M1000 is 114 trillion bits per second. It lives underneath an array of 34 chips, and it supports 1,500 watts. Just to put that in context, since I'm saying a lot of numbers here, the cables that connect North America and Europe are about 200 terabits per second. So a single M1000 from Lightmatter, our reference platform, is about the same bandwidth as the cable we're using to talk to Europe. It's pretty incredible all around.

We're talking about some massive performance improvements here. If we take a step back, what breakthroughs enable Lightmatter to make photonic computing at scale viable?

We do two things at Lightmatter. Generally, we're looking at the future of computing and interconnect. On the computing side, that's Envise, which is in our research division. On the interconnect side, we have Passage. The reason Lightmatter tech is so fast, and so alien in a way, is that we came into this, my co-founder Darius and myself, having done our doctorates at MIT in physics and electrical engineering, where we were building quantum computers using light. So when we look at the problem of communication and of computation, we see them completely differently. We've assembled an absolutely world-class team of physicists, engineers and business people to help bring this tech to market.

We've really pushed the envelope on packaging technology. Today, packaging is the differentiator between how good products are for semiconductor companies. Process nodes are less important than packaging technology, so we've invested like crazy in that. The reason we're so fast, ultimately, is that we're doing 3D stacks of traditional chips, like advanced-node chips, on top of Passage substrates. There's no limitation on how much of the chip can talk. Normally, computer chips can only talk around the perimeter of the die. We enable the entire surface area to communicate in 3D, and that gets you about a 100x increase in bandwidth.

Did you and your co-founder start working on quantum when you were at MIT? Was that the DNA or genesis of Lightmatter?

Yeah, that's right. We were working on building photonic quantum computers. We worked on quantum information theory and quantum computing, and what we would do is generate single photons, process them using photonic processors that we invented there at MIT, and then use superconductors to detect them. That was all really fun and exciting. I enjoyed that time period, and it was enormously productive. But that's a hard journey, and we both learned that's probably not the thing we want to work on for this 10 years. There's something more proximal, which is helping this AI thing scale using light.

Let's talk about this AI thing. You started seven or eight years ago, and AI arguably didn't have its seminal moment until ChatGPT, even though transformers came out in 2017 or 2018. Walk me through the sequence of events over the past seven or eight years in your founding journey, and also the maturity of the technology you've been building leading up to today.

We were building photonic quantum computers at MIT, and while we were there, in 2015, we saw the AI boom starting to take off. I'm an experimentalist, and one of the things my peers at MIT and I were noticing is that this deep learning thing had the ability to take a box with a lot of knobs and figure out how to tune all the knobs, without you having to understand what's going on, to achieve a desired outcome. When you build physical systems, it's often the case that there are a lot of knobs and you don't know the optimal way to tune them. So AI was taking off, and people were completely focused on how they could leverage it to drive science forward.

In 2017, the Transformer paper comes out and Google releases the Tensor Processing Unit. I think this is when the AI hardware boom really takes off, and it's when we were founded. We started out looking at how to accelerate the core math behind deep learning using light, and we quickly realized that's an important problem, but we needed to focus on how the chips communicate. We started to see that you weren't just using one chip to run AI models. You were starting to use 100, and then 1,000, and now 100,000, with a million coming soon. So the problem of how to link chips efficiently and at extremely high speed, so it seems like one giant chip rather than a bunch of chips synchronizing through a straw, was what we were after.

In terms of product readiness, we're already shipping chips to customers. We partner with big hyperscalers and semiconductor companies. It's been eight years of setting up the supply chain. We have partners like GlobalFoundries, Amkor and ASE, and we announced a partnership with Alphawave this week during the L200 launch. The tech is ready for high-volume manufacturing, and like I said, we're already shipping. But that's been an eight-year journey. It was totally easy.

I can only imagine. So we're eight years in. If we were having this conversation eight years from now, what do you hope, think or expect, both for where Lightmatter will be and for the technology itself?

One of the peculiar realities of building supercomputers is that the bigger they get, the larger the portion of the problem networking is. Today, if you look at an AI supercomputer, 75% of the spend is on the GPUs and 25% is on networking. If you go to 10 million or 100 million GPUs, it may be 50-50, or networking could be even bigger. So eight years from now, we're probably in this crossover regime where networking is the principal thing people are buying. I think about HAL 9000 and these sorts of things. We're definitely headed toward planet-scale computing. Today, data centers are using as much power as New York City, and New York City is seven and a half gigawatts.

Why do you think people aren't talking about this today? Is it just because human beings by our very nature are locked into the present, and we're not spending a lot of time thinking about where networking comes into the equation?

Well, Elon's been pretty good about calling it out, and I've been in agreement with him the whole time. He said transformers are going to be a problem soon, and then it's going to be power delivery. So some people are aware, and I think he knows because he's building a lot of these buildings. People are talking about it. Think about Microsoft bringing Three Mile Island, I think it's called, back online. It's a 900 megawatt (835 megawatt) nuclear reactor designed to power AI data centers.

It's just funny to think about where these things are going. If you look at Earth from space through a thermal imager to see what's hot, volcanoes would be the brightest objects. You'd see the deserts during the sunlight hours. You'd see the megacities, New York, Tokyo, places like that. And the new object that's going to be even brighter than the megacities is these data centers, because they're much smaller than a city. They're probably one one-hundredth of the area of a city, but they're using the same amount of power. So it's a pretty interesting future, and there's a lot of power to save by improving the networking piece. Like I said, networking will dominate as we get to bigger and bigger data centers.

What I admire most about people like you and companies like Lightmatter is not only the long-term mindset, but the ability to put your head down and work on an extremely difficult technical problem for the better part of a decade, all of which leads up to a moment like today. Congrats again on the product announcement, but more importantly on the technology and what you're all building and working toward. Nick, thanks so much for joining us today. It's great to see you.

Yeah, thank you, Eric.