Emerging Technologies · By Brian Fulcher
If you run AI models for customers, a chip can finish one calculation quickly and still leave the server doing a lot of work to deliver each answer. At SC26 in Chicago, Q.ANT reported a test of a different way to divide that work. A chip used light to handle one mathematical operation while an AMD graphics processing unit helped run the language model.
Prototype or demonstration. This label applies to the optical Fourier operation in the Gen 2+ language-model test.
Last reviewed November 17, 2026.

What changed
On November 17, 2026, Q.ANT announced the NPU Gen 2+ and reported the SC26 demonstration. The public summary of its technical paper with AMD identifies an AMD Instinct MI210 GPU and a language model with around one billion parameters. Inference is the work of running an already-trained model. Q.ANT says its optical processor performed discrete Fourier transforms in a model design that used them in place of a conventional attention layer.
How the optical part works
A discrete Fourier transform rewrites a series of numbers as a mixture of repeating patterns. That makes it useful in signal processing. It can also mix information across the pieces of text a model processes. The 2022 FNet research paper is an earlier example of Fourier mixing in a language model. Q.ANT’s test combined a GPU with a coprocessor that implements the transform optically.
The optical circuit uses how light waves combine to carry out that calculation. The model still has other computations to do, and values have to move between the electronic and optical parts of the system. The speed of the transform is one piece of the time it takes to deliver an answer.
What can be used today
IONOS announced a separate pilot on November 9, 2026. Approved applicants can test workloads on the earlier Gen 2 processor in a German data centre. Q.ANT says it is extending its developer toolkit for Gen 2+ and plans commercial systems combining the two processors in the second half of 2027. The launch account gives no current public purchase or hosted access to the specific Gen 2+ demonstration setup.
What the demo leaves open
The public summaries describe a working hybrid test and leave open how it compares with a GPU-only server on the same task at the same output quality. A Fourier-based model can behave differently from a model that uses conventional attention. The public material reviewed here gave no comparable whole-server energy per answer, latency or sustained throughput figure. The full technical paper is offered by request, so its methods and any additional figures were unavailable for this review.
Root Access Take
For infrastructure teams, the useful question is what a split between electronics and optics costs to operate. The comparison should include data transfers, conversion to and from light, and the software changes needed to use the accelerator. A successful demonstration gives teams a workload to measure. The case for deployment depends on the total service cost and whether users get comparable answers.
What to watch next
Look for independent tests on the same task at matched output quality, with whole-server power, response times and throughput reported together. It will also matter whether developers can run the Fourier operation on their own workloads and reproduce the results outside the conference setting.
Primary sources
Q.ANT’s November 17 SC26 announcement.
Q.ANT and AMD technical paper listing and public summary, November 2026.
IONOS announcement of the Gen 2 pilot, November 9, 2026.
FNet research paper, July 2022.
Related reading on Root Access Guy. Why AI networks are bringing light closer to the chip.