A couple of weeks ago, I spent a day at Q.ANT‘s office in Stuttgart. I talked with engineers, the product team and the marketing team, and I got to see photonic computing at work up close. I left more intrigued by the technology than when I walked in, and that curiosity is what led me to write this piece.
I want more ML engineers to be part of the conversation about photonic computing. Where does it fit? Who should be using it? What becomes possible when some of the math runs on light? I don’t think any one company can answer those questions alone, and I don’t think the answers will come from press releases. They will come from people trying it on real workloads and comparing notes.
A Co-Processor, Not a Replacement for CPUs and GPUs
First, a clear point. Photonic computing is not here to replace CPUs or GPUs. It doesn’t do everything, and it isn’t trying to. It works as a co-processor that sits next to them and takes on specific kinds of work. Its value is in the problems where conventional chips burn a lot of energy and compute, and probably in some problems we haven’t thought to try yet.
How Light Computes a Cosine Instead of Counting Steps
A digital chip counts. Every number is a pattern of switches, and every calculation is many switching steps. It is exact, and every step costs energy.
Q.ANT’s processor works differently. It sends light through tiny channels on a chip made from thin-film lithium niobate, a crystal whose optical behavior changes with voltage. A voltage shifts the phase of light in one path, two paths of light meet and interfere, and the brightness that comes out follows a cosine curve. So the chip doesn’t calculate certain functions step by step. The physics produces them.
In practice, the processor handles matrix multiplication, element-wise multiplication and cosine natively. Q.ANT reports up to 30 times better energy efficiency for equivalent matrix operations at the circuit level. It ships as a PCIe card inside a standard rack server, the Native Processing Server, which has been running at the Leibniz Supercomputing Centre since 2025 and which IONOS plans to offer through its cloud.
Two things matter for anyone building on it. It runs inference only, so you train on normal hardware and run the forward pass on the card. And it computes in bfloat16, so precision is a design constraint from day one.
From PyTorch to Photonic Hardware: The ML Engineer’s Workflow
The biggest barrier to trying this used to be hardware access. That changed in September, when Q.ANT open sourced its native computing toolkit under the Apache 2.0 license. It has Python and C/C++ APIs and, most importantly, a CPU emulator. You can write and test code on your laptop with no card and no drivers.
The workflow is familiar. You train a model in PyTorch as usual. You map its layers onto the operations the card supports, such as linear layers, convolutions, batch norm and KAN layers. You test everything on the emulator, then run it on hardware once you have access. In the current version, some operations like ReLU, softmax and pooling still run on the host CPU between calls to the card, which is worth knowing when you plan a model.
Why Kolmogorov-Arnold Networks Fit the Chip
The KAN detail is the one I keep coming back to. Kolmogorov-Arnold Networks put learnable functions on the connections between neurons, and those functions can be built from cosines. That is exactly what the chip produces natively. It is a rare case of a model architecture and a piece of hardware fitting each other, instead of one being forced onto the other.
Why Photonic Computing Needs Outside Experimenters
Every new kind of computing hits the same wall. The hardware works, but nobody yet knows which problems it is best for. GPUs were built for graphics long before anyone used them to train neural networks. The breakthrough uses came from people outside the company who experimented.
Photonic computing is at that early point now. It needs ML engineers who will benchmark it, share what works and what doesn’t, and figure out which workloads and which users benefit most. That is why I want to start an open research discussion around it. Not a sales pitch, but a shared notebook: real tests, real numbers and honest results, including the negative ones.
Where to Start if You Are New to Photonic Computing
I built Beyond the Transistor as a starting point. It explains the terms, walks through the tech stack and lets you run the core idea in your browser. There is a live cosine example, docs for the open source toolkit, a playground where you can design and train a small network, and a forum for questions and results.
If you work on ML problems that are expensive on today’s hardware, I would like to hear from you. Try the emulator, break something and bring your results to the forum. The most interesting uses of this technology probably haven’t been found yet.
Read Next: The AI Slowdown Is a Consolidation Play Dressed as Safety
Disclaimer: This article is written in a private capacity. Beyond the Transistor, its forum and the research initiative described here are independent activities. They are not affiliated with, endorsed by or related to Q.ANT GmbH or any of its team members. Product details are based on Q.ANT’s public materials.