A hands-on guide to photonic computing
Q.ANT builds a processor that does maths with light instead of transistors. This site explains the product in plain terms, runs its core idea live in your browser, and documents its open source toolkit with examples you can run today, without the hardware.
The product
From a crystal to a rack server to a Python library. Every figure links to the Q.ANT source it came from.
A photonic integrated circuit on z-cut lithium niobate on insulator, a crystal whose optical behaviour changes with voltage. Q.ANT makes it on its own pilot line with the Institute for Microelectronics Stuttgart (IMS CHIPS).
Source: qant.com, Photonic computing · LRZ goes into operation, 22 Jul 2025
The photonic core on a full-length PCIe card. It sits next to the CPU and GPU in an ordinary server and takes work over PCIe.
Source: qant.com, Photonic computing · NPU Gen 2 brochure, June 2026 (PDF)
A 19 inch rack server with an x86 host running Debian or Ubuntu LTS, holding several NPU cards. This is what customers install.
Source: qant.com, Photonic computing · NPU Gen 2 brochure, June 2026 (PDF)
The library that sends work to the card, for Python and C/C++, written in Rust. Open source since 23 September 2026, with a CPU emulator so you can develop without the hardware.
Source: Toolkit on GitHub · Open source SDK release, 23 Sep 2026
How it computes
A digital chip counts. Every number is a pattern of switches, and arithmetic is millions of switching steps. It is exact, and it costs energy for every step.
An analog chip weighs. It sets up a physical system whose behaviour is the answer and measures it. One measurement replaces many steps, but the answer is approximate, like reading a scale.
Q.ANT's core uses light. A voltage shifts the phase of light in one arm of an interferometer, the two arms recombine, and the brightness that comes out follows a cosine of that phase. So a cosine, and things built from cosines, come out of the physics rather than from a long chain of arithmetic. The live example shows this step by step.
The card runs the forward pass only. There is no backward pass on the device, so you train on a normal computer and run inference on the card. And it works in bfloat16, about three significant digits, so precision is a design constraint, not an afterthought.
Native optical operations, per Q.ANT
In toolkit 2.3.1, sent to the card
In toolkit 2.3.1, computed on the host CPU
Read from the toolkit's source code. The activations, bias and pooling run on the host today, between calls to the card. See Key concepts.
Source: NPU Gen 2 brochure, June 2026 (PDF) · Toolkit on GitHub
The numbers
About 30 times less energy per operation, at the circuit level. Q.ANT publishes this as a chart without the measurement method, and the card itself computes in bfloat16 rather than 8-bit integers.
Source: qant.com, Photonic computing
30x energy · 50x performance
Q.ANT phrases these as "up to" and "per application". Its June 2026 release ties the 30x to equivalent matrix operations at the photonic circuit level. Whole-system figures are not published.
Source: qant.com, Photonic computing · Generative AI on photonic hardware, 23 Jun 2026
From benchmarking first- and second-generation cards at the Leibniz Supercomputing Centre. The brochure does not state the baseline or the method.
Where it is
Founded in Stuttgart in 2018 by Michael Förtsch, who is CEO. Access today is an on-premises server for evaluation, with cloud access through IONOS announced for the coming months.
First Native Processing Server goes into operation at the Leibniz Supercomputing Centre (LRZ), running AI inference next to CPUs and GPUs. Source
€62M Series A, co-led by Cherry Ventures, UVC Partners and imec.xpand. Source
Second-generation NPU announced. Forschungszentrum Jülich starts a four-year collaboration and buys an NPS. Source
US headquarters opens in Austin, Texas. Series A closes at USD 80M. Source
IONOS signs as the first commercial customer, with cloud access planned. Source
A diffusion model and a time-series model run on the photonic hardware. Source
The toolkit goes open source, with a CPU emulator and six Python examples. Source
No hardware needed
See a cosine come out of interference, build one KAN edge from cosines, then run real Python in the page.
Open the example → On your machineInstall the open source toolkit with its CPU emulator and work through the six official examples, each tested on the emulator.
Read the docs → Design and runDesign a Fourier network, train it in the browser, or export a bundle that runs on CPU, the emulator, or a card.
Open the playground →Ask, argue and compare results with other ML engineers in the forum.
Product facts checked against Q.ANT sources on 8 October 2026. This is an independent guide, not Q.ANT material.