Photonic Computing: The Final AI Hardware Frontier (long QCLS)

I’ve spent the last few months studying photonic (optical) computing. Optical hardware is suddenly becoming a hot topic on Wall Street. But let’s look at why optical computing may be the next shift in AI hardware.

GPUs are important because they have a lot of cores doing work in parallel. A typical Intel or AMD processor has 16 to 32 cores. A GPU has tens of thousands of core-like computing units. Individual GPU cores are not faster than CPU cores, but the massive parallelism makes the system dramatically faster in aggregate.

But what are GPUs actually doing? What is so important that we need to raise hundreds of billions of dollars to buy them? They’re doing an arithmetic algorithm: the matrix multiplication. A matrix is simply a grid of numbers. Multiplying matrices together is the core operation in neural networks—the systems that power modern AI. In modern AI workloads, GPUs spend 99% of their time (and energy) performing matrix multiplications. This is why the entire AI hardware economy now revolves around one primitive: matrix multiplication. Trillions of dollars in market value—from Nvidia to hyperscale data centers—ultimately exist to make this one operation faster and cheaper.

While researching and poorly trading quantum computing stocks, I learned a lot about algorithms (my partner studied this at MIT) and computing in general. Photonic computing would occasionally pop up in my research and I’d overlook it. Eventually, I saw a paper in Nature that suggested matrix multiplications are “free” in light. The implication is staggering: if the dominant workload in AI can be executed using physics instead of transistors, the economics of computing change completely.

That stopped me cold.

Free? As in computationally O(1)? Yes, with caveats. (For what it’s worth, GPUs do matrix multiplications—which I will call “matmuls” from here on out—in O(n³). The O(1) is a bit contrived, but I do believe that photonic computing can do matmuls in close to linear time, e.g. something like O(2n). Shouldn’t that be a game changer? On large matrices, say 1 million elements (1000x1000), you’re talking about a difference between 2 million operations and 1 billion operations.

In other words, instead of scaling AI with more silicon, more power, and more cooling, we may be able to scale it with physics itself. And the optical computation itself can be nearly free in energy.

But is it possible? We’ll see. Light naturally performs linear transformations. This is because optical interference naturally computes weighted sums—exactly the primitive that matrix multiplication requires. But it is not so great at nonlinear operations, which occur between layers in neural networks. But the bigger question is “how do we pipe data into the optical matmul engine?” This isn’t trivial: you’re limited by how fast you can convert an electronic signal (say an API request) into an optical one. Then you’re limited by the nonlinearity and storage requirements of architectures like transformers. True optical memory does not exist—yet. I think I have solutions to these problems but we’ll talk about that some other time.

If this architecture works at scale, it could represent the first fundamentally new computing substrate since the transistor.

God chose light to do matmuls! So, my partner Chelsea Voss and I have started working with Q/C Technologies (Nasdaq: QCLS; she is on the Board, I am an advisor). The company has its own disclosure procedure, so I won’t front-run any announcements here. But I have made no secret on my X account of my view that photonic computing is the next hardware trend. At some point, we’ll explain what that means for Q/C—but the broader point is that photonic computing may define the next era of AI hardware.

I am a shareholder of and an advisor to Q/C Technologies.