Optical computing chip companies are advancing toward mass production, positioning photonic processors as potential alternative to electronic GPUs for inference workloads.

For more than half a century, Moore's Law has driven global computing power advancement: as manufacturing processes shrink, transistor counts increase, and chip performance improves. But this development model is reaching its limit.

Today, large language model parameters are expanding at nearly uncontrollable rates, while computing demand cycles compress continuously—and the physical ceiling of electronic chips has become clearly visible. The power consumption wall, memory wall, and bandwidth wall are three fundamental bottlenecks constraining the evolution of traditional GPUs. Higher-end computing chips consume more power, make heat dissipation increasingly difficult, and suffer greater inefficiencies from data transmission losses.

The path of stacking transistors to increase computing power is no longer viable. The industry urgently needs a fundamentally new solution, and using light instead of electricity for computation has emerged as the most disruptive and promising answer.

At the 2026 WAIC conference, the optical computing sector demonstrated core technologies from three leading domestic enterprises:

Optical Core Technology announced the successful tape-out of a photonic in-memory computing chip with a 256×256 matrix scale—currently the world's highest computing power density and precision optical computing chip, designed for high-computation scenarios including AI inference.

Lightelligence launched the world's first all-in-one optoelectronic hybrid computing box, Tian Shu Optical Cube, built on its self-developed PACE 2 acceleration card.

MemryX exhibited a full-stack integrated solution combining "spatial optical computing chip + large model applications" with its partners.

After intensive announcements, a recurring question emerges: Will optical computing truly achieve mass production and disrupt the computing power industry?

"Optical chip" is an umbrella term masking three fundamentally different categories: optical communication chips, optical interconnection chips, and optical computing chips.

Optical communication chips handle long-distance (kilometer-level) signal transmission and form the core of optical fiber backbone networks and 5G front-haul/back-haul infrastructure. Representatives include laser chips (DFB, VCSEL), detector chips (APD, PIN), and modulator driver chips. Their mission is maximizing transmission distance and signal integrity.

Optical interconnection chips enable short-distance (centimeter to meter level) connections between chips, boards, or servers, addressing data transmission efficiency. Typical forms include silicon optical transceivers and micro-ring modulator arrays. Their mission is increasing bandwidth density and reducing energy consumption per transmitted bit.

Optical computing chips differ fundamentally: they are not transmission channels but computing cores. Using photons as information carriers, they complete computations through optical signal transmission, modulation, and interference. Compared to traditional electronic chips, they offer natural parallelism (exploiting wavelength, phase, and polarization dimensions), ultra-low latency (near-speed-of-light transmission), and high energy efficiency (virtually no resistance loss). Often compared to GPUs, they focus on data processing.

A critical distinction must be established: optical computing's goal is never to replace GPUs, but to "reduce their burden."

GPUs excel undeniably. Their general parallel computing architecture performs well across training, inference, rendering, and scientific computing. But this versatility carries costs: power consumption density continues rising, heat dissipation system costs increase annually, and inter-chip interconnection bandwidth growth lags far behind computing power improvements. This severely restricts effective computing power utilization in large-scale clusters. Versatility offers breadth but locks the efficiency ceiling.

Breaking through this bottleneck has spawned two innovation paths: processing-in-memory technology, which integrates storage and computing through near-memory or in-memory approaches to alleviate memory wall latency and power consumption; and non-GPU architecture innovation—such as SambaNova's stream computing and Google TPU's ASIC route—which improves peak computing unit utilization through transistor-level microarchitecture redesign. Yet both remain constrained by digital integrated circuit fundamentals.

This is why optoelectronic hybrid computing is regarded as the "key variable" for next-generation computing infrastructure. Its disruption lies not in substitution but division of labor: assign computations ideal for optics to light, leaving complex logic and general control to electronics.

In the Transformer-dominated AI era, matrix multiplication accounts for over 70% of total computation—precisely optical computing's core strength. The principle is elegant: input optical signals split into multiple parallel paths, each weighted by tunable Mach-Zehnder Interferometers (MZI), then converge naturally at the output where accumulation completes instantaneously in the optical domain. Processing delay depends on light propagation speed in waveguides—approaching physical limits—and optical transmission generates virtually no Joule heat, delivering far lower power consumption than electronic chips.

Industry consensus has formed: GPUs handle training (requiring complex logic and high precision); optoelectronic hybrid solutions suit the inference stage. Thus the first major commercialization opportunity lies in AI inference—particularly for high-capacity, high-rate applications like image and speech recognition, plus meteorological monitoring and financial analysis.

The domestic optical computing sector is flourishing with enterprises pursuing different technical routes at varying implementation stages, though most focus on optoelectronic hybrid rather than all-optical computing.

Optoelectronic hybrid computing assigns control, storage, and nonlinear operations to electronic circuits while optical devices handle linear acceleration tasks like matrix multiplication and addition. Signals undergo repeated electro-optical and optoelectronic conversion at chip boundaries or between computing layers. This approach leverages the mature CMOS ecosystem and represents the mainstream commercializable solution today.

Companies pursuing the optoelectronic hybrid route have achieved significant breakthroughs. Lightelligence chose the silicon-based MZI coherent optoelectronic hybrid computing path, launching its second-generation PACE high-performance photonic computing processor in December 2021, pioneering integrated optical and electronic chip architecture and introducing the "new paradigm of optoelectronic hybrid computing power." From 2024 to 2025, its optical computing chip shipment volume ranked first globally for two consecutive years. Lightelligence has established strategic partnerships with SenseTime, StepStar, Shenwan Hongyuan, Zhongke Tiansuan, and OriSpace, spanning multiple AI applications.