An optical neural network performs the core arithmetic of machine learning - mostly multiplying vectors by matrices - using beams of light rather than electric current through transistors. Light has properties that make this attractive: many wavelengths (colors) can travel down the same channel at once without interfering, propagation is essentially instant, and the operations of multiplying and adding signals can fall out of the physics of how light passes through optical components, with little of the energy cost that digital multiply-accumulate hardware pays.
Two broad styles exist. One is free-space optics, exemplified by UCLA’s 2018 diffractive deep neural network, where light passes through a stack of patterned layers and the answer emerges from how the wavefront is bent. The other is integrated photonics, where waveguides, modulators, and interferometers etched onto a chip steer and combine light to compute. Companies such as Lightmatter have pushed the second approach toward production, reporting photonic processors that run standard networks like ResNet and BERT at accuracy close to digital systems while drawing tens of watts.
The promise is energy efficiency at a moment when the power draw of AI is a binding constraint. The catch is that light is good at the linear part of a neural network - the matrix multiplies - but the nonlinear activation steps, memory, and control still tend to need electronics, so practical systems are hybrids. Converting between optical and electronic domains costs energy and time, and precision is harder to hold than in digital arithmetic.
Why business readers should care: as inference cost dominates AI economics, photonics is one of the credible long-term bets on a cheaper substrate. It is moving from lab to product, but for now it complements GPUs rather than replacing them.