In the first half of this year, the performance of the optical communication sector in China's A-share market can only be described as "crazy." Over the past six months, the optical communication index has doubled, rising from 7,500 points to over 15,000 points.
Stock Price Changes of the Optical Communication Index Over the Past Half Year
Zhongji Innolight, a leading optical module manufacturer and constituent stock of the optical communication concept, reported a revenue of 19.496 billion RMB in the first quarter of this year, a year-on-year increase of 192%. Eoptolink, another optical module manufacturer, saw its revenue increase by 106% year-on-year in the first quarter of 2026. Yuanjie Technology, engaged in the R&D and production of indium phosphide laser chips, has seen its stock price surge by more than three times within half a year, with its market capitalization breaking through 200 billion RMB. As stock prices soar, a series of technical terms containing the word "light," such as "optical chips," "optical modules," "silicon photonics," and "CPO (Co-Packaged Optics)," have frequently appeared in the public eye.
Why have these "light"-related technologies suddenly exploded collectively?
The answer lies not in the candlestick charts of the capital market, but within the data centers. Here, a fundamental architectural revolution is underway.
What Role Does "Light" Play in Data Centers?
The explosion in popularity of "light" technology in data centers stems from a massive "data traffic jam."
Driven by AI demands, the computing and data exchange volumes in AI data centers are growing explosively. If a data center is considered a factory, an AI data center means this factory must have stronger production capacity to output more products, as well as stronger transportation capacity to deliver products promptly. In this factory, the "producers" are the computing chips. AI training and inference clusters are often composed of thousands or even tens of thousands of computing chips. With the iteration of chip products, their production capacity is steadily improving. However, the key factor restricting the further improvement of the factory's production efficiency falls on the "transporters"—the communication links.
In AI data centers, there are multiple "departments" acting as "couriers."
Four Levels of Data Transmission in AI Data Centers
At the chip level, chips on the market use electrons as the carrier for information transmission; whereas optical computing chips use photons instead of electrons as the information carrier. By manipulating optical signals for data processing and transmission, they can break through the physical limits of traditional electronic chips (such as GPUs/CPUs) in terms of computing power, energy consumption, and heat generation. Constrained by technological maturity and popularity, optical computing chips are not yet the mainstream solution.
At the board level, data transmission occurs between computing cards, which means connecting multiple computing chips to make them operate like a single chip. At this level, the traditional solution uses PCB (Printed Circuit Board) copper foil routing. However, electrical signal transmission will attenuate and generate heat if the distance exceeds a dozen centimeters. The CPO (Co-Packaged Optics) currently being explored by the industry directly solders the optical engine (electrical-to-optical device) right next to the GPU. Data is immediately converted into optical signals and transmitted as soon as it leaves the chip's "door," completely eliminating electrical losses on the motherboard.
At the server level, data transmission occurs between different server boards. This level is responsible for expanding a single machine's 8 cards into a "super large node" with dozens or hundreds of cards. At this level, high-speed optical modules or LPO (Linear Pluggable Optics) are replacing copper cables to play their roles.
At the data center level, the role of data communication is to connect hundreds or thousands of racks into a supercomputing cluster. Data is routed and distributed through multi-layer massive physical switches. To eliminate the high power consumption and long-tail latency caused by the secondary "optical-electrical-optical" conversion in traditional electronic switches, the network core layer is introducing OCS (Optical Circuit Switches).
In this infrastructure revolution brought about by the sharp increase in AI computing power demands, to improve the daily throughput of data centers, a revolution of "light advancing and copper retreating" is quietly taking place within this giant AI factory.
CPO: Packing "Light" into Chip Packaging
In traditional data centers, optical modules are "pluggable"—they are independent of the switching chips and plugged in via electrical interfaces. This architecture is simple and flexible but has hit a ceiling in the AI era: power consumption is too high, density is too low, and signals attenuate over long transmission distances.
The concept of CPO (Co-Packaged Optics) is straightforward: directly package the optical engine next to the switching chip, shortening the electrical interconnection distance from several centimeters to a few millimeters, thereby significantly reducing power consumption and latency. Silicon photonics is the process foundation for realizing CPO—it uses mature CMOS semiconductor processes to manufacture photonic devices, enabling optical chips to be produced on a large scale and at low cost, just like electronic chips.
In March 2025, NVIDIA released the Spectrum-X and Quantum-X network switches based on silicon photonics CPO technology at the GTC conference, marking a milestone event for CPO moving from the laboratory to productization. The performance data released by NVIDIA is quite stunning: overall energy efficiency improved by 3.5 times, signal integrity improved by 63 times, network resilience improved by 10 times, and the number of lasers reduced to a quarter of the original. The total throughput of the Spectrum-X Ethernet platform reaches 400 Tb/s, and the Quantum-X InfiniBand platform is 2 times faster than the previous generation. Its core foundry partner is TSMC (Taiwan Semiconductor Manufacturing Company), responsible for silicon photonics chip manufacturing and Advanced Packaging.
However, CPO is still in the early stages of commercialization. Traditional pluggable optical modules still have advantages in flexibility and maintainability and will not be completely replaced in the short term. The industry research firm Yole Group judges that CPO will achieve large-scale deployment between 2028 and 2030, at which time the network architecture of data centers will be completely rewritten. A rate of 200G per lane is expected to become mainstream between 2026 and 2027, paving the way for the CPO versions of 800G and 1.6T optical modules.
From the perspective of the industrial chain, this track has gathered a group of heavyweight players: Zhongji Innolight (TeraHop), Cisco, Broadcom, and Marvell are the leaders in vertical integration; startups such as Ayar Labs, Lightmatter, and Celestial AI focus on segmented directions; TSMC, GlobalFoundries, and Intel are responsible for wafer foundry. Zhongji Innolight has shipped millions of optical modules for AI interconnects and is becoming an unignorable Chinese force in the global CPO track.
Optical Chips: The Most Scarce Link in the Entire Chain
The optical chips mentioned here are not the "optical computing chips" mentioned above. They are not responsible for computing but act more like converters—used to complete the conversion between optical and electrical signals. Simply put, when an electrical signal comes in, the optical chip converts it into an optical signal and sends it out; when the optical signal returns, the optical chip converts it back into an electrical signal. It is the "heart" of the optical module.
In the high-speed optical modules of AI data centers, the most core optical chips are EML (Electro-absorption Modulated Lasers) and CW (Continuous Wave) lasers. EML integrates the laser and modulator on a single chip, featuring high speed and long transmission distance, making it the core light source for 800G and 1.6T optical modules. CW lasers are mainly used in silicon photonics solutions, with a shorter process chain and more guaranteed yield.
But right now, this "heart" is in short supply.
The global EML market size in 2025 is approximately 4.18 billion RMB, with China accounting for one-third, about 1.37 billion RMB. The scale sounds not small, but the demand growth rate far exceeds the capacity expansion rate—the delivery time for high-end EML lasers has been scheduled until after 2027, and building a new production line takes at least two to three years.
TrendForce stated bluntly in its April 2026 report: EML and CW lasers are the primary bottlenecks restricting the capacity expansion of optical modules. This shortage has directly transmitted to the downstream—in the first quarter of 2026, the total prepayments of the three leading optical module manufacturers, Zhongji Innolight, Eoptolink, and Tianfu Communication, surged from 172 million RMB to 2.267 billion RMB, a 13-fold increase. The main purpose of prepayments is to lock in upstream capacity. This figure shows that downstream manufacturers have reached the point where "they cannot get goods without paying in advance."
The global high-end EML chip market has long been dominated by a few overseas giants—Lumentum and Coherent in the US, and Mitsubishi Electric and Sumitomo Electric in Japan.
Chinese manufacturers are catching up. In the2025 Global Top 10 Optical Module Suppliers list released by LightCounting, 7 are Chinese companies, although the localization rate of high-end optical chips remains low. The good news is that a group of domestic companies is achieving group breakthroughs: Yuanjie Technology's CW products have been delivered in batches, and its profit in the first quarter of 2026 is close to the level of the whole year of 2025; Vertilite's 100G EML has been mass-produced, and 200G is under customer validation; Dongshan Precision's 100G PAM4 EML has been used for over 10 million units cumulatively, and 200G has also entered the mass production stage.
The industry judges that the overseas supply shortage coupled with domestic technological breakthroughs provides domestic manufacturers with a window period of two to three years.
Optical Modules: The "Hard Currency" of AI Computing Power
Optical modules package optical chips, driver circuits, and optical components into a pluggable small box, responsible for transmitting optical signals between GPUs, switches, and servers. It is currently the most deployed optical device in data centers and the most commercially mature link in the "light" industrial chain.
How large is the optical module market? According to data from TrendForce and LightCounting, the global AI-specific optical transceiver module market will reach 16.5 billion USD in 2025, and is expected to surge to 26 billion USD in 2026, a year-on-year increase of over 57%. The traffic in North American hyperscale data centers is still growing at a rate of over 30% annually. Cloud giants such as Google, Microsoft, and Meta continue to expand their GPU and AI server deployments, and the demand for optical modules is expected to become increasingly strong.
From the perspective of technical nodes, the current optical module market is in a critical period of transition from 800G to 1.6T. 800G is the absolute main force in 2025, with shipments of about 18 million to 19.9 million units; 1.6T will enter the volume production stage in 2026. Chinese companies are quite far ahead in this track. Zhongji Innolight's revenue in 2025 was 38.24 billion RMB, a year-on-year increase of 60%, with a net profit attributable to shareholders of 10.8 billion RMB; Eoptolink's revenue was 24.842 billion RMB, a year-on-year increase of 187%; Tianfu Communication's revenue was 5.163 billion RMB, a year-on-year increase of 59%. In the first quarter of 2026, Zhongji Innolight's single-quarter revenue approached 20 billion RMB, a year-on-year increase of 192%.
Optical Interconnect and OCS: Rewriting Data Center Networks at the "Speed of Light"
The CPO, optical chips, and optical modules mentioned earlier all solve the transmission problems "inside the equipment" and "between equipment," belonging to point-to-point connection technologies. But a data center is a network composed of tens of thousands of servers. Data not only needs to be transmitted between two devices but also scheduled across the entire cluster. This requires switching equipment in the network backbone layer.
In traditional data center networks, the core switch is an electrical switch—data must undergo an "optical-electrical-optical" conversion within it, which is both power-consuming and time-consuming. The concept of Optical Circuit Switches (OCS) is: signals are switched directly in the optical domain. OCS is the "traffic police" of the network backbone layer. It achieves dynamic reconstruction of the network topology through programmable optical path switching. It is at the same level as traditional electrical switches, but with significantly reduced power consumption and latency.
Google has already deployed OCS on a large scale in its Jupiter data center network. In a paper published at ACM SIGCOMM in 2022, Google disclosed that the Jupiter network achieved a 5-fold increase in speed and capacity by introducing OCS and software-defined networking, but the power consumption of OCS is only one-tenth to one-hundredth of that of traditional electrical switching. In April 2026, a white paper released by the Open Compute Project (OCP) presented a larger picture: the global OCS market is expected to exceed 14 billion USD in 2027. Overseas giants such as Meta, Microsoft, and Amazon are also actively exploring this technology.
In March this year, at the 2026 Appliance & Electronics World Expo (AWE), Shanghai INESA (Group) Co., Ltd., together with Lightelligence, Biren Technology, and ZTE, officially released the Guangyue Super Node 128-card. Adopting Lightelligence's globally first silicon photonics OCS optical switching chip, it has become the first domestic optical interconnect and optical switching GPU super node solution.
Progress in academia is equally active. Nagoya University proposed a 4096×4096 port OCS architecture with a total capacity of 819.2 Tb/s; the University of Cambridge demonstrated a training-phase-aware optical network solution, achieving a 37.5% communication acceleration through dynamic OCS reconfiguration; Alibaba Cloud verified the CPU-GPU resource pooling solution of PCIe 5.0-over-optics plus dynamic OCS; Shanghai Jiao Tong University also proposed the ocs-DRP solution.
In this data center infrastructure revolution triggered by AI demands, the maturity of different "light" technical nodes is not consistent. Optical modules are the most mature link, while CPO and OCS optical switching technologies are still in the early stages of market promotion, with broad market prospects in the future.
This revolution named after "light" has just begun.