Neurophos is an American semiconductor company developing photonic hardware for artificial intelligence inference. Founded in 2020 and spun out of Duke University, it designs an optical processing unit that performs matrix multiplications using light rather than electrons, employing silicon photonics combined with programmable metasurfaces. The company was founded by technical founders including CTO Tom Driscoll, alongside Patrick Bowen and Andrew Traverso, and has its origins in research associated with Duke University and Metacept Inc.
Its flagship product, the T100 Optical Processing Unit (OPU), is designed to deliver exaflop-scale AI inference performance on standard CMOS manufacturing lines. Neurophos reports an energy efficiency of 235 TOPS/W for the T100, against 7 TOPS/W for the NVIDIA B200 - equivalent to a roughly 30-fold advantage in performance per watt. The use of metamaterials is claimed to shrink optical components by a factor of 10,000 compared with traditional silicon photonics.
The company has raised $110m in Series A funding. Its stated objective is to address the energy consumption of AI data centres, a growing constraint for hyperscale computing and AI infrastructure. Engineering work spans silicon photonics, metamaterials and semiconductor design, with relevance to data-centre, hyperscale and semiconductor employers.
