Q.ANT unveils photonic NPU Gen 2+ for AI workloads
Q.ANT’s new photonic processor adds optical Fourier transforms and a hybrid AI demonstration, but commercial systems are still planned for 2027.
Q.ANT has unveiled its NPU Gen 2+, a photonic processor that adds optical Discrete Fourier Transforms (DFTs) to the company’s accelerator technology. The announcement was made at SC26 in Chicago on 17 November, with Q.ANT presenting the hardware as a co-processor for conventional CPUs and GPUs rather than a replacement for them.1
DFTs are used across signal processing, image processing, simulation and data analysis. Q.ANT says performing the operation directly in light could reduce the amount of electronic computation needed for workloads that handle large volumes of data. The company frames that as a route to leaner AI models and lower energy use, but those benefits remain claims from the manufacturer rather than independently verified results.1
What Q.ANT is actually showing
The Gen 2+ is designed to extend the mathematical operations that Q.ANT’s photonic hardware can perform natively. The company says its optical component can carry out a DFT in a single operation, using the physical behaviour of light instead of a long sequence of electronic calculations.
That does not mean the entire AI workload runs optically. Q.ANT’s SC26 demonstration paired the NPU Gen 2+ with an AMD GPU in a hybrid system running a large language model. According to the company, the optical processor handled a computationally intensive attention layer while the GPU continued to provide conventional processing.1
That distinction matters. Photonic accelerators are being positioned as specialist hardware for selected operations, with the CPU and GPU still handling the rest of the application. Q.ANT says the Gen 2+ connects through PCIe and fits standard rack-mountable servers, which should make it easier to test alongside existing data-centre hardware.
A new software path, not a mass-market product
Q.ANT launched an open-source SDK in September with C and Python APIs, examples, documentation and a CPU-based simulator. The company says it is now extending that toolkit for the Gen 2+ so developers can experiment with Fourier-based algorithms without needing immediate access to optical hardware.1
Q.ANT also points to a photonic pilot with IONOS as a route to real hardware access. The company says commercial systems combining its Gen 2 and Gen 2+ processors are planned for the second half of 2027. That timing makes this an early platform announcement and roadmap milestone, not evidence that a broadly available AI accelerator has reached the market.1
The company’s own announcement also describes the processor as complementary to existing compute. In practical terms, the pitch is not “replace every GPU with light”; it is “move a narrow set of expensive mathematical operations to a dedicated optical co-processor”. Whether that division of labour delivers a worthwhile improvement will depend on software support, data movement, system integration and workloads that genuinely benefit from optical execution.
Our opinion
Q.ANT’s announcement is interesting because it focuses on a specific bottleneck instead of treating photonic computing as a vague alternative to electronic chips. A DFT-capable co-processor and a hybrid language-model demonstration give the technology a clearer test case than broad claims about replacing conventional hardware.
The important caveat is that the announcement does not provide an independent benchmark, a full system power comparison or enough detail to judge the claimed gains against a modern GPU. The Gen 2+ is best understood as a promising specialist platform with a long commercial runway ahead, rather than an immediate answer to the energy cost of AI. The SDK and planned 2027 systems will tell us more about whether the hardware can move from an impressive demonstration into useful, repeatable deployments.