首页文章详情

Raised 1.5 billion yuan in financing, Shanghai gave birth to a super unicorn: An 89-born doctor develops optical chips, the world's first.

铅笔道2025-09-04 20:38
Lightelligence raised 1.5 billion yuan in financing today.

A super unicorn has emerged in Shanghai recently: Lightelligence. It raised 1.5 billion yuan in financing today. Institutions such as the fund under China Mobile, Shanghai State-owned Investment Corporation, Guoxin Fund, and Pudong Venture Capital jointly invested.

A wave of computing power driven by light is accelerating towards us.

01

This company's business is a bit obscure: it is an optoelectronic hybrid computing power provider. As we all know, the chips in computers process information through electrical signals, but electrical signals have limited speed and are particularly power - consuming. While the speed of light is much faster than that of electricity, and its energy consumption is also low.

What Lightelligence does is to combine light and electricity to create more powerful "computing helpers".

Why is the signal processing speed of electricity slower than that of light? Because when electrical signals propagate in wires, electrons have to "jostle" between metal atoms to move forward, so the speed naturally cannot be fast, usually only traveling hundreds of thousands of kilometers per second.

Light signals are different. Light itself is an electromagnetic wave and does not need to "squeeze" through material atoms. It only needs to travel straight along the glass channel in the optical fiber, so it has an advantage in speed.

So, in what scenarios is optical computing needed? For example, large - scale AI model training, autonomous driving, and aerospace satellites.

Another question arises. Since both electricity and light have clear computing scenarios, what is optoelectronic hybrid computing power, and why is it needed? Because although light signals have high speed, they have weak flexibility and logical control, and it is difficult to independently complete complex operations.

Especially in computing power centers, optoelectronic hybrid computing power is more needed. And Lightelligence's core products are these: such as photonic computing processors, optical interconnection chips, and overall solutions. In simple terms, it is to build "computing power highways" and "photonic engines" for artificial intelligence data centers and supercomputing centers.

This startup story began in 2017. Shen Yichen, who had just obtained a doctorate in physics from the Massachusetts Institute of Technology, returned to China to start a business with a research result. He was born in 1989 and was 28 years old at that time.

This scientist from Hangzhou, Zhejiang, who had published a cover paper in "Nature Photonics", found that traditional electronic chips faced physical limits when processing neural network calculations - the slow transmission speed and high energy consumption of electrical signals made it difficult to meet the future AI computing power requirements.

The optical neural network architecture he proposed theoretically proved the feasibility of using photons for matrix operations, which became the technological starting point of Lightelligence. In 2018, Lightelligence was officially registered and established in Changyang Chuanggu, Shanghai.

In 2019, the company released the world's first photonic chip prototype board, verifying the technological feasibility.

In 2021, it launched the second - generation photonic computing processor PACE, which integrated more than ten thousand photonic devices. In specific neural network algorithms, its performance was hundreds of times higher than that of high - end GPUs.

In 2025, LightSphere X, jointly launched by Lightelligence, Biren Technology, and ZTE, is the world's first distributed optical interconnection and optical switching GPU super - node solution, which won the highest award "SAIL Award" at the World Artificial Intelligence Conference.

02

Lightelligence is in the silicon - based photonic computing chip industry.

During the embryonic stage (2010 - 2017), the academic community had realized the physical limits of electronic chips, but photonic computing was only in the theoretical stage and faced the pain points of low integration and insufficient reliability.

From 2018 to 2022, the demand for AI computing power exploded, and the industry entered the rising stage. Application - specific integrated circuit chips became a transitional solution, but the bottleneck of data transfer in large - scale clusters became prominent.

Since 2023, photonic interconnection technologies represented by co - packaged optics (CPO) have begun to emerge, solving the limitations of electrical interconnection in terms of bandwidth, power consumption, and latency.

In 2025, the global market size of photonic computing chips is expected to exceed 12 billion US dollars. Among them, the Chinese market accounts for a relatively large proportion, and its growth rate ranks first in the world.

Currently, the industry is facing a thorny problem: when the scale of the computing power cluster climbs above the kilocalorie level, the traditional electrical interconnection solution exposes many shortcomings.

In computing power clusters above the kilocalorie level, the problems of energy consumption and latency are quite severe. In terms of energy consumption, under the traditional electrical interconnection solution, the energy consumption for data transmission is extremely high. Taking a medium - sized intelligent computing center with thousands of GPUs as an example, the annual electricity consumption for data transmission alone is as high as millions of kilowatt - hours. According to the current commercial electricity price, the annual electricity bill alone exceeds one million yuan.

The latency issue is also not to be underestimated. During large - model training, data needs to be frequently transmitted between different computing nodes. Under the traditional electrical interconnection method, the data transmission latency can reach the millisecond level.

From the perspective of the competitive landscape, the global photonic computing chip market is concentrated at the top, but emerging forces are rising. According to a relevant market research report released by Research and Markets, the top 5 manufacturers account for 71% of the market share.

Among them, Chinese enterprises are developing quite rapidly.

According to relevant data, from 2019 to 2023, the global market share of Chinese CPO module products reached 5% in 2023. In 2025, Chinese manufacturers, relying on technological breakthroughs and production capacity advantages, have occupied more than 70% of the global high - end module market share.

03

Currently, the blue ocean in this field is concentrated in the high - speed product lines of 800G and above, and there is significant room for domestic substitution.

Data from LightCounting, a research institution in the optical communication industry, shows that in 2024, hyperscale cloud service providers heavily invested in AI infrastructure, making 400G/800G Ethernet optical modules sell like hotcakes. This investment boom continued in 2025, and Chinese cloud providers also followed suit.

In the field of 800G optical modules, there are few international companies that can commercialize 400G - 800Gb/s silicon photonic chips, but Chinese manufacturers are at the forefront. Zhongji Innolight revealed that in 2025, the demand of key customers for 800G optical modules could be several times higher than that in 2024, and the company has been expanding the production capacity of relevant products.

In China, one of Lightelligence's competitors is LightIC. LightIC focuses on photonic in - memory computing chips (computing - in - memory) based on phase - change materials, but its technical route is different from that of Lightelligence.

In addition, one of Lightelligence's differentiations is the "open ecosystem + distributed architecture".

Nvidia has a closed system, just like a highway built by a family only allowing its own cars to run; while Lightelligence's optical switching solution can be compatible with GPUs of different manufacturers regardless of the protocol, solving the incompatibility problem of "ten different protocols for ten domestic GPUs".

From the demand side, this field has entered the growth stage: the demand continues to rise, especially in fields such as large - scale AI model training, autonomous driving, and aerospace satellites, which have a strong demand for high - speed, low - latency, and low - power computing. The application scope of optical computing technology is becoming more and more extensive.

This article does not constitute any investment advice.

This article is from the WeChat official account "Pencil News" (ID: pencilnews). The author is Xi Wen. It is published by 36Kr with permission.