I designed an optical neural network that's 700× more efficient than H100, what am I missing?[D]

I spent a few hours designing an alternative to silicon-based AI inference using optical computing. Looking for reality checks on whether this is directionally viable or fundamentally broken. The Core Idea: Instead of electrons flipping bits, use light passing through programmable optical materials to perform neural network inference. The math happens in the wave propagation itself diffraction through layers is the computation. The Architecture: 5 layers of electronically-tunable light filte