On November 17, Q.ANT unveiled the NPU Gen 2+, the newest iteration of its Native Processing Unit, at SC26 in Chicago. The headline addition is an optical component on the photonic chip that performs Discrete Fourier Transforms directly in light — extending the set of mathematics the processor can carry out natively beyond the matrix multiplication and non-linear operations the Gen 2 line, launched in 2025, was built around. And in the demonstration that anchored the announcement, Q.ANT coupled the Gen 2+ with an AMD GPU to run a large language model, which the company describes as the first time a model of this kind has run with optics in the loop.
That is two announcements in one, and they deserve to be judged separately. The first is a capability claim: a new native operation, performed in a single pass of light rather than a long sequence of electronic steps. The second is a systems claim: a photonic co-processor and a GPU running one model together, with optics replacing one attention layer. The first is concrete and checkable in principle. The second is a demonstration, and the company has published no performance, energy, or model-quality figures for it. The gap between them is where the honest reading of this week lives.
What the Fourier transform adds
The Discrete Fourier Transform is one of the most heavily used operations in computing. It converts a signal between the time domain and the frequency domain, and it underpins signal and image processing, simulation, and data analysis. On digital hardware it is expensive at high data volumes: a fast Fourier transform decomposes into a long sequence of multiply-and-add steps, each costing time and energy.
Q.ANT’s claim is that the Gen 2+ does the transform in a single operation, using the physical properties of light. An optical component on the chip performs the DFT directly, rather than the electronics decomposing it. The company’s chief executive, Michael Förtsch, put the argument in algorithmic terms: “With the DFT performed natively, developers now have a new tool to reshape the mathematics of AI models.” The point is not that optics does the same arithmetic faster; it is that a new native instruction lets developers write different algorithms — models that use fewer parameters and less data because the expensive function is now cheap. The press release says the Gen 2+ “demonstrates the potential algorithmic and performance advantages of photonic computing already now.” Note the word “potential.” The company is describing an opening, not a result.
The DFT is a defensible choice of operation to add: a well-defined function with a precise answer, so a photonic implementation can be checked against a known result, and genuinely useful across AI and HPC — beyond transformers, Fourier methods appear in spectral solvers, frequency-domain convolutional networks, and scientific simulation.
But the addition also raises the question every analog optical system faces, and the announcement does not answer it: precision. Q.ANT’s own earlier hardware converts between digital and analog at 16-bit precision, and photodetector noise has to be accounted for — the Leibniz Supercomputing Centre’s evaluation of the Gen 2 said so explicitly. A Fourier transform computed in an analog optical component will carry an error budget that a digital fast Fourier transform does not, and nothing this week gives a number for it.
The LLM demo: what happened, and what did not
The demonstration is the part of the announcement that will be quoted most, so it is worth stating exactly what Q.ANT says happened. At SC26, the company ran a large language model in a hybrid system coupling the NPU Gen 2+ with an AMD GPU. Functions performed in light replaced one computationally intensive attention layer. Q.ANT describes this as the first time a language model has run with optics in the loop, and says the demonstration shows how optical functions “can replace a computationally intensive attention layer and address a major computational bottleneck in generative AI.”
What the announcement does not say is just as important. It does not name the model, its parameter count, the task, the throughput or latency, the energy consumed, or whether output quality changed when one attention layer moved to optics. There is no comparison against the same model running on the GPU alone. As one trade report noted, the demonstration establishes a system configuration rather than a quantified performance advantage. The demo proves the parts can be connected and a model can run. It proves nothing about whether the result is faster, cheaper, or better.
Getting a photonic processor and a GPU to cooperate on a live language model — moving data between domains, keeping the model coherent, producing output — is real integration work, and the first time a vendor has shown it publicly is a genuine milestone for the field. But a demonstration on vendor-chosen terms, with no numbers, shows the system exists and runs. It does not show it is worth buying.
The choice of attention as the target is deliberate and well aimed. Attention dominates the cost of running a transformer: modelling the relationships between every token in a sequence scales quadratically with sequence length, and it is the reason long-context inference is expensive. Q.ANT’s vice president of software, Utz Bacher, described the work explicitly as “Fourier-based attention layers in language models.” There is precedent for the idea, independent of Q.ANT: in 2021, researchers at Google showed that self-attention in a transformer could be replaced by unparameterised Fourier transforms — the FNet architecture — and still reach competitive accuracy on standard language benchmarks, at a fraction of the computational cost. What is new here is not the algorithm; it is the hardware claim that the Fourier transform can be performed natively, in light, inside a working system.
Where Gen 2+ sits on the roadmap
The Gen 2+ is an extension, not a new platform, and Q.ANT’s own framing says so. It builds on the NPU Gen 2, launched in 2025, which performs matrix multiplication and non-linear operations on thin-film lithium niobate (TFLN) photonic integrated circuits. It keeps the same integration path — a PCIe card in a standard rack-mountable server — and the same positioning: a co-processor alongside CPUs and GPUs, not a replacement for them. “We have shown that our processors work as co-processors with GPUs,” Förtsch said.
That positioning is the sensible one, and it is also the constrained one. A co-processor that handles a defined slice of a workload leaves the rest of the model on silicon, which means data must move between the optical and electronic domains, and that movement costs energy, latency, and precision. LRZ researcher Dr. Ehab Saleh noted that “the co-processors cannot perform all operations using light,” so “a continuous data exchange between photonic and classical processing units is required.” The Gen 2+ widens the slice of the workload that can stay in light. It does not remove the boundary.
The strongest independent evidence about this hardware family remains the Leibniz Supercomputing Centre’s evaluation of the Gen 2, published on April 21 and covered by this desk last week. LRZ’s headline figures: the Gen 2 ran up to 50 times faster than the Gen 1 in the best case, on matrix multiplications, and ResNet-18 ran image recognition 25 times faster, with power consumption down roughly 50 to 84 percent between generations. The baseline is the story: LRZ compared the second generation with the first — nothing with a GPU or a CPU, and no measured energy-to-solution figure against a datacenter GPU node. The 50-times figure is a generational improvement, not a competitive one. And the Gen 2+ has no independent evaluation at all yet.
One discrepancy on the record is worth carrying into this week. Q.ANT’s May 19 release said the Gen 2 “demonstrated up to a 100x performance increase” in LRZ’s evaluation. LRZ’s April article gives a best case of 50 times; no LRZ document contains the 100x figure. This week’s announcement, which contains no performance figures at all, at least avoids repeating the error. Two more German centres hold Gen 2 hardware with no published results yet — Jülich and the Karlsruhe Institute of Technology, which has committed to publishing its results as open source. KIT is the most likely source of a public, reproducible benchmark for this hardware family.
The software is the product, and it is getting wider
The hardware announcement came with a software one, and the software may matter more. Q.ANT launched its open-source SDK in September 2026 — September 23, by the company’s earlier account — supplying C and Python APIs, sample applications, documentation, and a CPU-based simulation environment so developers can build and test code without touching optical hardware. It followed that with the Photonic Algorithms Library, Q.PAL. This week, Q.ANT said it is extending the SDK to support Gen 2+ capabilities, letting developers experiment with Fourier-based algorithms.
That is the right sequence, and it is the sequence the field has historically failed at. Hardware nobody can program is hardware nobody buys. An open-source SDK with a CPU emulator lowers the cost of trying the hardware to zero — a developer can write Fourier-based attention code today, on an ordinary machine, and decide whether the algorithm is worth running on optics before ever touching a card.
The access path to real hardware is also widening, though narrowly. The IONOS pilot announced on November 9 — covered by this desk last week — puts Gen 2 processors on bare-metal servers in a German datacenter, free to approved developers with a concrete use case. Whether the Gen 2+ joins the pilot is not stated. For now, the Gen 2+ is a device you can simulate in software and see demonstrated at a conference — not one you can rent. Q.ANT plans to offer commercial systems combining Gen 2 and Gen 2+ processors in the second half of 2027, roughly twenty months from now, and the company is explicit that the two generations will be sold together. Further out, Q.ANT says it is studying how to stack multiple optical functions onto a single chip. That is a roadmap statement, not a result.
The honest ledger
First, the DFT addition is real capability, and the right kind: a precise, checkable mathematical operation with a known answer. The physics argument — a transform in a single pass of light rather than a sequence of electronic steps — is coherent, and the algorithmic argument has independent precedent in published research.
Second, the LLM demonstration is a working hybrid system, and the first of its kind in public. But it is a demonstration without numbers: no model named, no throughput, no energy, no accuracy, no GPU-only comparison. It proves the system runs. It does not prove it helps.
Third, the precision question is unanswered and will decide whether Fourier-based attention on optics is a research curiosity or a product. Analog optical computation carries an error budget; the Gen 2’s 16-bit conversion and photodetector noise were documented by LRZ. Nothing this week says what error the optical DFT carries. The DFT’s determinism means the error can be measured precisely.
Fourth, there is still no independent evaluation of any Q.ANT processor against a GPU on a named workload. LRZ’s 50-times figure is generational, not competitive. The 100x figure in Q.ANT’s May release appears in no LRZ document. The Gen 2+ adds a new operation with no measured numbers of any kind.
Fifth, the software story is genuinely advancing, and it may matter most in a year. An open-source SDK, extended for Fourier-based algorithms, with a CPU emulator and a bare-metal pilot behind it, is the infrastructure a developer community needs. If developers find that Fourier-based attention is worth building, the hardware has a reason to exist.
What to watch next
Three things will tell readers whether this week mattered. First, whether Q.ANT publishes anything quantitative about the demo — the model, the workload, the throughput, the energy, the accuracy against a GPU-only baseline. A company that publishes the numbers is inviting scrutiny; one that moves on has treated the demo as a marketing event.
Second, whether the Gen 2+ reaches independent hands. KIT has committed to publishing its results as open source, and the IONOS pilot is the route for outside developers to run their own workloads. Either would produce the first measurement of the optical DFT that did not come from the vendor.
Third, what developers do with the extended SDK. The Fourier transform is a tool, not a result; its value depends on whether anyone builds a better model with it.
This is the second photonic compute demonstration at SC26 this desk has covered in as many days — LightSolver’s LPU went public on Monday, Q.ANT’s Gen 2+ on Tuesday. The conference has become the field’s proving ground.
Until then, read this announcement narrowly and literally. A photonic chip gained a new native operation, and a language model ran with optics in the loop beside an AMD GPU, for the first time in public. Everything else — speed, energy, precision, and whether any of it matters to a customer — is still to be measured. The demo is the beginning of the evidence, not the evidence itself.
Sources and further reading
Primary sources: Q.ANT, “Q.ANT Unveils NPU Gen 2+: Photonic Fourier Transforms Power Large Language Model,” GlobeNewswire, November 17, 2026 (Gen 2+ unveiling at SC26, optical DFTs, hybrid LLM demonstration with an AMD GPU, SDK extension, commercial systems in the second half of 2027, Förtsch and Bacher quotations). Independent evaluation: Leibniz Supercomputing Centre, “LRZ Puts Photonic Computers to Real World Test,” April 21, 2026 (Gen 2 vs Gen 1: up to ~50x faster on matrix multiplications, 25x faster CNN image recognition on ResNet-18, power down ~50–84%, 16-bit conversion, photodetector noise; no GPU or CPU baseline). Company figures flagged as unverified: the “up to a 100x performance increase” figure in Q.ANT’s May 19, 2026 release, which appears in no LRZ document; the up-to-30x energy-efficiency claim from internal benchmarking; the Gen 2 datasheet figures (150 W per NPU, 8 GOPS sustained on nonlinear functions, 100,000 GOps roadmap target for 2028). Context: Supercomputing News, “Photonic Computing 2026: Supercomputing Centers Test Q.ANT”; GazeWire, November 18, 2026.