British startup Lumai has moved from lab demo to productization with Iris, a server that offloads matrix multiplication to a free-space optical engine. Input vectors are encoded into 1,024 laser sources, passed through a display that encodes model weights as pixel intensities, and focused by a lens that performs the addition step optically. No integrated photonics, no exotic materials -- just physics doing math.
The efficiency argument is genuine and not hand-wavy. Because the optical domain handles the heavy matrix-multiply and power draw is proportional to input vector size while compute scales with the square of matrix dimensions, efficiency improves as matrices get larger. Lumai handles 2048x2048 matrices in one pass, which is exactly where transformer inference spends most of its time. The claimed 50x GPU throughput and 90% power reduction for Llama-class models lines up with the fundamental physics; what remains unproven is production yield, reliability at scale, and whether the electrical/optical conversion overhead eats the efficiency gains at small batch sizes.
The roadmap is credible in structure if not yet in timing. Iris Nova (single engine, test clusters by end of 2026) leads to Iris Aura (multi-engine rack) and then Iris Tetra (cluster-scale). That is a sensible three-generation ramp. The detail worth tracking: Lumai uses a software orchestration layer to decide which workload fragments route to the optical engine versus the co-located CPU. That layer is where the real integration work lives, and it is not shipped yet.
Optical compute has a long history of promising 10-100x efficiency gains and failing to commercialize due to precision, thermal, and scalability problems. Lumai's approach via free-space optics rather than integrated photonics is genuinely differentiated, and the EE Times coverage is based on actual tested hardware running DeepSeek-class models -- not simulations. Whether Iris Nova ships in volume by Q4 2026 is the first real test.