Neurophos raises $110M to build tiny optical processors for inferencing (techcrunch.com)

Neurophos, a photonics startup spun out of Duke University, has raised $110 million in a Series A funding round to advance its development of tiny optical processors designed for AI inferencing. Utilizing a novel “metasurface modulator,” Neurophos aims to revolutionize computing by creating an optical processing unit that performs matrix vector multiplication more efficiently than traditional GPUs and TPUs based on silicon. With thousands of these modulators on a chip, the startup claims its technology will significantly enhance processing speed and reduce power consumption, achieving a rate of 235 Peta Operations per Second (POPS) at just 675 watts, compared to Nvidia's B200 GPU which delivers 9 POPS at 1,000 watts. This innovation is critical for the AI/ML community, as the current demand for computational power in AI applications is outpacing energy efficiency improvements. The metasurface approach not only addresses these pressing needs but also promises to overcome historical challenges in the manufacturing of optical chips, utilizing standard silicon processes. With companies like Microsoft showing interest, Neurophos plans to deliver its first chips by mid-2028, positioning itself to potentially disrupt a market dominated by giants like Nvidia. This advancement is considered a pivotal step towards achieving sustained growth and efficiency in AI development.