From decades of promise to 100× gains: Unlocking optical compute for AI inference

Optical computing has spanned over six decades of elegant concepts that have failed to deliver beyond niche demonstrations. These efforts were invariably constrained by fundamental barriers: data movement limitations, DAC/ADC power consumption, modulator density, switching speeds, quantization, and temperature sensitivity. The underlying device physics and system architectures have kept photonic accelerators orders of magnitude behind digital electronics for real-world AI workloads.

In this talk, we dissect these core barriers and show how Neurophos has overcome them with a metamaterial-enabled optical processing unit (OPU) that integrates more than one million micron-scale optical processing elements on a single chip. This capability is powered by a breakthrough in metamaterial optical modulators—delivering 10,000× miniaturization relative to conventional photonic designs—unlocking incredible density, low loss, and CMOS-compatible manufacturability.

The outcome is ~100× higher performance and energy efficiency than leading GPUs for general-purpose AI inference. We will detail the core physics innovations behind this leap and present the roadmap to datacenter-ready, exaOPS-scale photonic accelerators.