Neurophos, founded in 2020, develops silicon photonics and programmable metasurface optical processing units designed to perform the matrix multiplications underlying AI inference at the speed of light. The company's premise is that replacing electronic computing with optical computing addresses the power and performance bottlenecks facing AI workloads. It was spun out of Duke University and Metacept Inc., giving it research-driven origins in metamaterials and photonics.
Its flagship product, the T100 Optical Processing Unit (OPU), is a photonic AI inference chip that uses active programmable metasurfaces to carry out matrix multiplications optically. Neurophos reports energy efficiency of 235 TOPS/W, against 7 TOPS/W for NVIDIA's B200 - a 30x advantage in performance per watt - while delivering exaflop-scale AI inference performance at a fraction of GPU power consumption. The company also states that its programmable metasurface approach reduces optical component size by 10,000x compared with traditional silicon photonics.
The technology targets CMOS-compatible semiconductor manufacturing and is aimed at AI infrastructure and data centres, hyperscale cloud computing and the semiconductor and photonics sectors, with an emphasis on the energy efficiency of data centre AI workloads. Neurophos raised $110 million in a Series A round. Technical work spans silicon photonics, optical computing, metamaterials, AI inference hardware and matrix multiplication acceleration.
