Optical chips return as a bid to cut the energy cost of AI inference
IEEE Spectrum's Dina Genkina reported on 17 November 2026 that optical computing is getting another look as a way to run AI inference. The combined valuation of optical-computing companies is now $2.22 billion, up from $0.4 billion in 2022. Startups are appearing in the US, China, Germany and Israel. Most are targeting a 10- to 80-fold energy-efficiency gain over today's GPUs.
The logic starts with the workload. Inference accounts for most AI energy use, and Cornell's Peter McMahon says it has become power-limited. Its core operation is matrix-vector multiplication, which Joseph Goodman and colleagues proposed doing with light in 1978. Electronic multiplication energy scales as N squared for an N-by-N matrix. Optical energy scales roughly as N, because the multiplication step is passive once the matrix is fixed. In 2022 a team led by McMahon showed an optical neural network using under one photon of energy per multiply, about 2.5e-19 joules. A standard computer uses 1e-14 to 1e-12 joules. McMahon estimates optics begins to win at vector lengths between about 1,000 and 10,000.
This matters for abundance because the electricity bill is a main limit on how cheaply intelligence can be served. If inference takes far less energy, it can be offered to more people at lower cost, and it eases pressure on power supply.
The evidence is still early. Lumai of Oxford made its Iris Nova proof-of-concept server available for evaluation in April 2026. It says its next server will use about one-tenth of a GPU's energy for prefill. Neurophos of Austin targets 35 times the energy efficiency of Nvidia's Blackwell B200, with rack deliveries planned for 2028. These are company claims or targets, not independently verified. The 2022 figure excluded the cost of encoding data. McMahon cautions against extrapolating the 100,000-fold saving. Amdahl's law limits gains because nonlinear operations stay electronic. Whether optics can deliver at scale before electronics catch up remains an open question.
Source: IEEE Spectrum
Many Minded