Optical AI Processor TeraFlow Deep Dive: When Photons Replace Electrons Running Through Chips

On August 11, 2028, photonic computing company Lightmatter announced that its third-generation optical AI processor, TeraFlow, has entered mass production. This chip uses photons (optical signals) rather than electrons to perform the matrix operations in AI inference, improving performance-per-watt by 15x.

TeraFlow's core is a matrix computing unit composed of a 128x128 optical interferometer array. Unlike traditional electronic chips where electrons flowing through copper wires generate heat, photons propagating through silicon waveguides produce almost no heat. This means TeraFlow can perform massively parallel computation at extremely low power consumption.

Lightmatter CEO Nicholas Harris said in an interview that TeraFlow was designed to solve AI inference's energy bottleneck. "Take a GPT-5-class large model — the annual electricity consumption of global inference compute already equals that of a medium-sized country. Switching to optical processors could reduce energy consumption by over 90%."

Google Cloud has announced plans to test TeraFlow chips in its 2029 data center deployments. Microsoft Azure and AWS are also in evaluation stages. Lightmatter said TeraFlow's production yield has reached 72%, with unit prices between $800 and $1,200, comparable to high-end GPUs.

However, optical computing has limitations. TeraFlow currently supports only inference, not training, because the weight updates and gradient backpropagation required during training remain technically challenging for optical systems. Harris acknowledged: "Optical training may be a goal for the next-generation product."