Optical computing

State of the art (2026)

State of the art (2026). Optical computing - performing the matrix multiplication itself in the photonic domain - remains demonstration-stage and unproven for production AI inference; Lightmatter, Lightelligence and Q.ANT have shown hardware, but none is a production accelerator at scale, and electronic silicon keeps improving faster than photonic compute can displace it. The live, commercial photonics story is optical interconnect (moving data between conventional chips with light), which is a different and faster-moving category tracked under Photonics and silicon photonics. We hold optical computing at low conviction and watch for a photonic accelerator that demonstrates both accuracy and scale on a real AI workload.

Core thesis

Optical computing — specifically photonic AI accelerators that use light instead of electrons to perform matrix multiplication — represents a potential paradigm shift in AI hardware. Matrix-vector multiplication, the core operation in neural network inference and training, can be performed at the speed of light using Mach-Zehnder interferometer arrays or micro-ring resonator meshes, consuming orders of magnitude less energy per operation than electronic GPUs. Lightmatter's Envise chip, Lightelligence's Hummingbird, and Luminous Computing (acquired by Alphabet) have demonstrated photonic matrix multiplication at competitive accuracy.

The energy argument is compelling. Training a frontier AI model like GPT-4 consumed an estimated 50-100 GWh of electricity. Nvidia's H100 GPU consumes 700W per chip. Data centers are projected to consume 4-5% of US electricity by 2030. If photonic processors can deliver equivalent TOPS (tera-operations per second) at 10-100x better energy efficiency, they could break the power wall that is becoming the binding constraint on AI scaling.

However, significant engineering challenges remain. Photonic systems excel at linear operations (matrix multiplication) but struggle with nonlinear activation functions, which require optical-to-electrical-to-optical (OEO) conversion, erasing much of the speed and efficiency advantage. Manufacturing photonic integrated circuits (PICs) at scale requires new fabrication processes distinct from conventional CMOS. Precision is another issue: photonic systems typically operate at 4-8 bit precision, whereas training requires 16-32 bits. Optical computing may find its first viable market in inference (where lower precision is acceptable) rather than training, and in specific workloads (transformer inference, recommendation systems) where the linear algebra is dominant.