Neurophos
Neurophos builds silicon photonics and programmable metasurface optical processing units that perform AI inference matrix multiplications with light, improving energy efficiency over GPUs.
Neurophos develops optical processing units that carry out the matrix multiplications at the heart of AI inference using light rather than electrons. Founded in 2020 and spun out of Duke University and Metacept Inc., the company builds on silicon photonics and programmable metasurfaces to attack the energy demands of AI data centres. Its T100 Optical Processing Unit performs exaflop-scale AI inference at a claimed 235 TOPS/W, against 7 TOPS/W for NVIDIA's B200, a 30-fold advantage in performance per watt.
The approach rests on active, programmable metasurfaces that shrink optical components by a factor of 10,000 compared with traditional silicon photonics. The chips are designed for CMOS semiconductor manufacturing, and the work spans silicon photonics, metamaterials, optical computing, matrix multiplication acceleration and energy-efficient computing. Target markets include AI infrastructure, data centres, hyperscale computing and semiconductors.
Neurophos raised $110m in a Series A round from investors including Gates Frontier, M12, Aramco Ventures, Bosch Ventures and Space Capital. The company describes itself as mission-driven, focused on the rising energy cost of AI inference and on rethinking the underlying hardware.