Author: Feng Ning
For over half a century, Moore's Law has been the underlying logic driving the iteration of global computing power: shrinking process nodes, stacking more transistors, and delivering increasingly powerful chip performance. However, this approach is now nearing its limits.
Today, the parameter scale of large models is expanding at an almost uncontrollable pace, and the cycle for surging computing power demands is being continuously compressed. Yet, the physical ceiling of electronic chips is already visible to the naked eye. The power wall, memory wall, and bandwidth wall—these three major hurdles have firmly stalled the evolutionary pace of traditional GPUs. The more high-end the computing chip, the higher its power consumption, the more difficult its heat dissipation, and the more severe the invalid loss in data movement.
Simply put: the path of boosting computing power by stacking transistors has reached a dead end. The industry urgently needs a completely new solution, and using light instead of electricity for computation has emerged as the most disruptive and highly anticipated answer at present.
At the recently concluded 2026 WAIC (World Artificial Intelligence Conference), the optical computing track went all out, with three leading domestic companies revealing their trump cards in a concentrated manner:
Lightbase Technology announced the successful tape-out of its first photonic in-memory computing chip with a 256*256 matrix scale. This chip currently boasts the highest computing density and computing precision among optical computing chips globally, primarily targeting high-computing scenarios such as AI inference.
Lightelligence launched the world's first optoelectronic hybrid computing integrated computing box—Tianshu · Optical Cube. This innovative product is built based on its independently developed PACE 2 accelerator card.
DeepThink Photonics jointly with partners exhibited a full-stack integrated solution of "spatial optical computing chip + large model applications".
After a wave of intensive exposure, everyone is asking the same question: with optical computing this hot, is it truly about to achieve mass production and commercialization, disrupting the computing power industry?
01. Three Types of "Optical Chips" with Entirely Different Missions
"Optical chip" is an overly generalized umbrella term. In the industrial context, there are at least three distinct categories with vastly different functions, namely optical communication chips, optical interconnect chips, and optical computing chips.
Among them, optical communication chips are responsible for long-distance (kilometer-level) signal transmission and are core components in fiber optic backbone networks and 5G fronthaul/backhaul. Typical representatives include laser chips (DFB, VCSEL), detector chips (APD, PIN), and modulator driver chips. Their core mission is to enhance transmission distance and signal integrity.
Optical interconnect chips handle short-distance (centimeter to meter-level) interconnections between chips, boards, or servers, addressing the issue of data movement efficiency. Typical forms include silicon photonic transceiver engines and microring modulator arrays. Their core mission is to increase bandwidth density and reduce energy consumption per bit transmitted.
Unlike the previous two, optical computing chips are positioned not as transmission channels, but as computing cores. An optical computing chip is a novel processor that uses photons as information carriers, completing data computation through physical processes such as the transmission, modulation, and interference of optical signals. Compared with traditional electronic chips, it boasts advantages such as natural parallelism (utilizing multi-dimensional information like wavelength, phase, and polarization), ultra-low latency (transmission close to the speed of light), and high energy efficiency (virtually no resistive loss).
Since it focuses on data processing, it is also frequently compared with GPUs.
02. Optical Computing Chips: Where is the First Pot of Gold?
Given the critical turning point in the optical computing industry, a core understanding urgently needs to be established: the goal of optical computing has never been to replace GPUs, but to "lighten the load" for GPUs.
The advantages of GPUs are unquestionable—their general-purpose parallel computing architecture enables them to handle diverse scenarios such as training, inference, rendering, and scientific computing with ease. However, this versatility comes at a cost: power density continues to rise, cooling system costs increase year by year, and the growth rate of inter-chip interconnect bandwidth lags far behind the rate of computing power improvement, severely constraining the effective computing power utilization rate in large-scale clusters. In other words, versatility trades for breadth, but it also locks in the efficiency ceiling.
To break through this bottleneck, the industry has formed two innovative paths: first, processing-in-memory (PIM) technology, which deeply integrates storage and computing through near-memory or in-memory computing, alleviating the latency and power consumption caused by the "memory wall" and data movement; second, non-GPU architecture innovations, such as SambaNova's streaming computing and Google TPU's ASIC route, which improve the peak utilization of computing units through transistor-level microarchitecture reconstruction. However, these two paths ultimately fail to break free from the fundamental constraints of digital integrated circuits (ICs).
Precisely for this reason, optoelectronic hybrid computing is regarded as the "key variable" for next-generation computing infrastructure. Its disruptive nature lies not in replacement, but in the division of labor: assigning the computing tasks most suitable for optics to light, and leaving complex logic and general control to electricity.
In the AI era dominated by Transformers, matrix multiplication accounts for over 70% of the total computing workload, which happens to be the core advantage scenario for optical computing chips. Its principle is concise and highly efficient: the input optical signal is split into multiple paths for parallel propagation, with each path assigned a corresponding weight via a tunable Mach-Zehnder Interferometer (MZI). The signals from each path naturally converge at the output end, and the accumulation operation is completed instantaneously in the optical domain. The latency of the entire computing process depends on the propagation speed of light in the waveguide—offering near-physical-limit immediacy—while the optical transmission process generates almost no Joule heat, resulting in power consumption far lower than that of traditional electrical chips.
At present, a consensus has basically been formed in the industry: the training stage relies on GPUs (due to complex logic and high precision requirements), while the inference stage can adopt optoelectronic hybrid solutions. In this sense, the first pot of gold for the commercialization of optical computing chips is mainly mined from the AI inference end. For example, in the field of AI with larger computing capacity and higher speed requirements, such as primary AI tasks like image recognition and speech recognition. Beyond that, weather monitoring and financial investment are also potential application targets.
03. Racing on Multiple Technical Routes: Who is Closest to Mass Production?
Currently, the domestic optical computing track is flourishing, with different companies betting on different technical routes. The progress of commercialization has also shown a clear gap, but it is mainly concentrated on optoelectronic hybrid computing rather than all-optical computing.
Optoelectronic hybrid computing uses electronic circuits for control, storage, and non-linear operations, while optical devices only handle linear acceleration tasks such as matrix multiplication and accumulation. Signals need to undergo repeated electro-optical and opto-electrical conversions at the chip boundaries or between computing layers. This solution can effectively leverage the mature CMOS ecosystem and is the mainstream approach capable of commercial deployment at present.
In industrialization practice, companies pursuing the optoelectronic hybrid route have achieved significant breakthroughs. Lightelligence has chosen the silicon-based MZI coherent optoelectronic hybrid computing route. In December 2021, the company launched its second-generation high-performance photonic computing processor, PACE, pioneering the integration of optical and electrical chips and proposing the "new paradigm of optoelectronic hybrid computing power" for the first time. From 2024 to 2025, the company's optical computing chip shipments ranked first globally for two consecutive years. It is reported that Lightelligence has reached strategic cooperation with SenseTime, StepFun, SWS, Zhongke Tiansuan, and Orient Space, covering multiple AI application scenarios, with the goal of achieving full-link synergy from underlying technology to upper-level solutions.
The products of the US company Lightmatter are divided into three parts: the photonic computing platform (Envise), chip interconnect products (Passage), and compatible software (Idiom). Envise is the world's first photonic computing platform. Each Envise processor features 256 RISC cores, providing 400Gbps inter-chip interconnect bandwidth, and supports the PCI-E 4.0 standard interface, offering excellent compatibility. The principle of the Envise processor is that light performs computations through waveguides, and adding a light source of a different color can correspondingly increase the computing speed. For example, if a red laser source can perform 1 million computations per second, adding another color of laser source can double the speed to 2 million, and so on; adding light sources requires almost no hardware modifications.
The Envise platform is not a pure photonic processor, but a hybrid optoelectronic heterogeneous system that combines mature CMOS processes with advanced silicon photonic integration technology, aiming to balance flexibility, programmability, and ultimate performance.
The aforementioned Lightbase Technology offers a leading photonic in-memory computing solution in terms of matrix scale among currently available public information. Distinct from the traditional MZI architecture, photonic in-memory computing integrates weight storage with optical computing units, theoretically alleviating the "memory wall," reducing the power consumption overhead caused by repeated weight refreshing, and demonstrating energy efficiency potential in continuous inference scenarios.
Unlike the on-chip integrated silicon photonics route, DeepThink Photonics focuses on free-space optical computing, relying on spatial light modulators to achieve highly parallel optical operations. This technical route may be more suitable for scenarios such as edge devices, automotive, and industrial perception, where constraints on chip integration size are relatively loose and ultra-low static power consumption is pursued, without the need for highly complex heterogeneous integration of silicon photonic wafers.
In terms of university research achievements, as early as 2024, the Tsinghua University team pioneered the distributed broad-spectrum intelligent optical computing architecture and developed the world's first large-scale interference-diffraction heterogeneous integrated chip, Taichi, achieving 160 TOPS/W of general-purpose intelligent computing. In June 2025, the team from the Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, innovatively solved the problem of "high-density information parallel processing on optical chips." By integrating independently developed multi-wavelength light source chips, large-bandwidth optical interaction chips, reconfigurable optical computing chips, high-precision optical matrix driver chips, and parallel optoelectronic hybrid computing algorithms, they successfully developed the ultra-high parallel optical computing integrated chip—"Meteor No. 1," realizing an optical computing prototype verification system with a parallelism greater than 100.
All-optical computing, on the other hand, pursues the retention of optical signals throughout the entire computing link, attempting to eliminate optoelectronic conversion losses and directly implement linear operations and non-linear transformations in the optical domain. However, constrained by the lack of high-performance integrable all-optical non-linear devices and the difficulty in scaling optical storage, it is currently only at the laboratory verification stage.
In December last year, the team from the School of Integrated Circuits at Shanghai Jiao Tong University achieved a breakthrough in the field of next-generation computing chips, realizing for the first time the all-optical computing chip LightGen, which supports large-scale semantic media generation models. According to reports, LightGen can fully implement the closed loop of "input—understanding—semantic manipulation—generation," completing high-resolution (≥512×512) image semantic generation, 3D generation (NeRF), high-definition video generation, and semantic control, while simultaneously supporting multiple large-scale generative tasks such as denoising, local and global feature transfer.
Although the optical computing track continues to yield results and attract a crowd of players, it currently remains only at the stage of small-scale pilot commercialization, far from reaching the level of large-scale mass production. It is fundamentally constrained by four major shortcomings. First, low process yield and high production costs result in an overall cost-effectiveness inferior to that of mature GPUs. Second, repeated optoelectronic conversions bring additional power consumption, offsetting the energy efficiency advantages of photonic computing. Third, the software ecosystem is fragmented, making adaptation difficult and deployment costly. Fourth, the supporting core industry chain is imperfect, with insufficient capacity and technology to support scaled deployment.
Looking back at the development history of the computing power industry, every breakthrough in the computing power revolution is an iterative upgrade of the underlying physical carrier. From vacuum tubes to transistors, from single-core electronic chips to multi-core GPUs, today, photons are opening a brand-new chapter in next-generation computing power.