Dr. Michael Förtsch stood beneath a humming rack of shimmering fiber-optic circuits in Hamburg, where pulses of light—instead of electric currents—were generating images and predicting financial trends in real time. At ISC High Performance 2026, Q.ANT unveiled what many in the AI world have long awaited: a photonic processor running not just experimental code, but full-scale, production-ready generative AI and recurrent neural networks. This wasn’t a lab curiosity—it was a diffusion model synthesizing images and an xLSTM network forecasting time series data, both operating on light-based computation.
For decades, AI’s growth has been throttled by energy. As models grow larger, so do their power demands—often requiring megawatts of electricity and vast cooling infrastructures. Q.ANT’s breakthrough signals a pivot: by replacing transistors with photons, the Stuttgart-based company demonstrated that its second-generation Native Processing Unit (NPU) can handle the most mathematically intense AI tasks with up to 30 times the energy efficiency of classical silicon processors. This leap isn’t theoretical—it’s measurable at the circuit level.
The demonstration featured two landmark workloads. First, a diffusion model—akin to those powering tools like Stable Diffusion—performed image-to-image synthesis, a process reliant on thousands of parallel matrix operations. This marked the first time such a complex generative model has run on photonic hardware. Second, NXAI’s TiRex model, built on the xLSTM architecture, executed time series prediction for applications in finance, supply chains, and weather forecasting. Both models ran natively, proving that photonic computing can support not just one AI paradigm, but many.
The momentum extends beyond the booth. In April, independent developers at Daisytuner successfully compiled a PyTorch-based object detection model directly onto Q.ANT’s hardware—the first time a standard machine learning framework has been deployed on a photonic system. By May, commercial demand followed: German cloud giant IONOS placed the first orders for Q.ANT’s processors, integrating them into high-performance cloud infrastructure. Institutions like the Leibniz Supercomputing Centre Munich and Jülich Research Centre are now evaluating the technology for large-scale scientific computing.
"Q.ANT’s photonic architecture changes the energy calculus for AI infrastructure," says Förtsch. And experts agree. Professor Dr. Björn Ommer, a pioneer of stable diffusion models at LMU Munich, called the demonstration a "compelling sign" that alternative computing substrates could shape AI’s future. As AI’s energy footprint looms larger, photonics isn’t just a promising alternative—it’s emerging as a necessity. With light now doing the work of transistors, the next era of computing isn’t just faster. It’s sustainable.