"Stand in the light, not just stand there." This long-standing joke circulating in the capital market reflects the intense popularity of the "optical" theme.
From the explosion of AI computing power demand in early 2023 to the structural shortage game of upstream materials and high-end chips in 2026, the "optical" concept has been hot in the capital market for a long time. This narrative differs from traditional optical communication; the main discussions focus on orders for 800G optical modules, the commercialization pace of 1.6T, and the controversy over the technological route of CPO (Co-Packaged Optics). These topics all point to optical interconnects.
Optical interconnects are an exclusive subset of optical communication in the AI era, a dedicated high-growth branch spawned by this round of AI computing power expansion, and the core track with the strongest performance realization and most concentrated narrative in the current optical field.
Currently, the underlying layer of computing infrastructure is undergoing a generational shift from electricity to light. Electrons are steadily approaching physical limits, transistor density is moving towards the atomic scale, and bandwidth improvements in copper interconnects are accompanied by an exponential climb in power consumption. Traditional electrical paths have revealed an insurmountable ceiling in the face of the interconnection demands of tens of thousands of card clusters in the AI era. Relying on its natural physical characteristics of high speed, low loss, and high parallelism, light is regarded as the only viable path to break through this bottleneck. Optical interconnects are the direct product of this shift.
This is a critical technological turning point. Therefore, today we want to systematically explore the following—
What is the real status of each link in the optical interconnect industry chain?
Where is the most scarce link, and in which direction are profits migrating?
Are the market's core judgments on optical interconnects right or wrong?
(Recommended to bookmark and read slowly)
Understanding Optical Interconnects from an Industry Panorama
"Light" is becoming the universal foundation supporting the next-generation digital economy. However, looking at the four core tracks of the photonics industry, although they share the same technological origin, their industrial maturity, prosperity cycles, and realization capabilities are completely fragmented.
1. Optical Connection (Optical Communication/Optical Interconnect)
The core value of optical connection is very pure: utilizing photons to transmit, forward, and switch data. As the basic disk of the photonics industry, optical connection has entered a high-growth stage relying on AI computing power iteration, and is currently the only golden track with continuous large orders and deterministic performance.
Optical connection can be clearly divided into two major sectors.
The first is traditional optical communication, serving operators' wide-area networks, covering scenarios such as backbone networks, metropolitan area networks, 5G fronthaul, home broadband, and submarine optical cables. It features stable demand and a gentle cycle, representing the traditional basic disk of the industry.
The other is data center optical interconnect, which actually belongs to a refined subset of optical communication. It is high-end optical communication in AI scenarios, specifically used for short-to-medium distance interconnection of AI computing clusters, covering scenarios from a few meters to 2 kilometers. It includes 800G/1.6T/3.2T optical modules and CPO optical engines, undertaking massive data interaction between GPUs and switches, and between clusters, serving as the core infrastructure of the computing network.
2. Optical Sensing
Optical sensing mainly relies on photon characteristics to achieve physical signal perception, widely applied in scenarios such as automotive LiDAR, industrial inspection, medical optics, and environmental monitoring. The related track lacks the explosiveness of the AI track, but excels in stable cycles, rigid demand, and mature commercialization. Moreover, it can share underlying technologies such as lasers and optical packaging with optical connections, possessing continuous technology spillover value.
3. Optical Computing
This is the track most prone to over-hyping in the market. The core of optical computing is to utilize photon interference and phase modulation to complete core AI matrix operations, belonging to the computing power calculation layer. Its industrial positioning is an optoelectronic hybrid co-processor, not a replacement for GPUs. At this stage, it is only in the laboratory R&D and small-scale pilot stage. The market generally believes that optical computing will be difficult to achieve large-scale commercialization in the next 5 years, and the industry inflection point is expected to arrive only after 2030. Currently, it only has long-term narrative value, with no substantive commercial orders, belonging to a typical long-term layout track.
4. Frontier Photonics Technology
It mainly covers directions such as quantum photonics, optical storage, and AR photonic chips. This track is currently far from commercial implementation and is only suitable for basic scientific research layout.
Understanding Optical Interconnects within the Industry Chain
From the perspective of the complete full-stack AI computing system, the bottom layer consists of computing chips such as GPUs, ASICs, and HBMs, which is also the core of computing power. The middle layer is the high-speed transmission system of optical interconnects, connecting upwards to switches, complete machine cabinets, and IDC infrastructure, and ultimately extending to computing power scheduling and cloud service applications. Among them, the positioning of optical interconnects is precise and irreplaceable. It is the core hub connecting the underlying computing power with the upper-level network, and also the core bottleneck for the large-scale expansion of AI clusters.
The industry chain of optical interconnects itself can be roughly divided into five levels: core optical materials, high-speed optical chips, optical devices and optical components, integrated terminal forms, and optical fibers and connectors. Specifically—
The starting point of the optical interconnect industry chain is to solve the problem of "where light comes from and how fast it can run." Core optical materials determine the loss, speed, and modulation efficiency of a beam of light, and also determine the upper limit of the entire industry's performance.
Indium Phosphide (InP) is the core material for high-speed EML lasers, adapting to the mainstream scenarios of 800G and 1.6T. Affected by export controls, the gap in 6-inch substrates is as high as 70%, and the shortage pattern is expected to last until 2029. Thin-film lithium niobate (TFLN) targets ultra-high-speed iterations of 3.2T and above, with leading performance and an ecosystem under improvement. Silicon photonics relies on mature CMOS processes, has the highest integration degree, and is generally regarded as the mainstream direction for future photonic integration. The three are not in a substitution relationship, but play their respective roles at different speeds and scenarios.
With light available, the next step is to "write" electrical signals onto light and "read" received optical signals back into electrical signals. This action is completed by optical chips. Optical chips are the core bare semiconductor dies for electro-optical and opto-electrical conversion, and also the link with the highest barriers and tightest choke points in the entire industry chain.
In a single 800G optical module, the cost proportion of optical chips is as high as 55% to 70%, but the domestic production rate is less than 5%. The EML high-speed laser on the transmitting end is the core device for high-end scenarios of 200G and above, almost monopolized by overseas enterprises. The PD and APD detectors on the receiving end also face a gap in high-end capacity. The annual industry demand is about 150 million units, while domestic capacity is only 80 million units, and high-end products continue to be in short supply.
Optical chips cannot work independently; they need to rely on optical devices to complete beam shaping, isolation, and wavelength multiplexing, ensuring the stability of optical signals during high-speed transmission.
TOSA and ROSA are standardized optical components that represent the first integration of bare chips and passive devices, and also the optical core of traditional pluggable optical modules. Domestic enterprises have achieved global competitiveness in the low-to-mid-end passive device field, but there is still room for import substitution in high-end miniaturized and high-reliability products. The overall industry supply-demand gap is about 40%, and the certification cycle for high-end devices is as long as 2 to 3 years, forming a time barrier no less than chip manufacturing. Components like TOSA and ROSA will face long-term pressure from form changes and demand structure migration under the trend of CPO and PIC (Photonic Integrated Circuit) integration.
In addition, opto-electronic components, as the upstream precision processing link of passive devices, cover basic optical parts such as lenses, filters, and isolators. AI speed iteration increases the precision requirement by about 10 times, with a capacity gap of about 25%.
Packaging devices and chips together constitutes the current most mainstream integrated terminal form—the pluggable optical module. It completes the final conversion of optoelectronic signals and is the standardized interface between equipment and optical fibers.
According to the latest forecast by LightCounting in March 2026, the global high-speed optical module market size will reach $28.5 billion in 2026, a significant year-on-year increase of 60%. The compound growth rate from 2025 to 2031 will exceed 20%, and the market size is expected to break through $60 billion in 2031, with data center optical modules contributing over 70% of the market increment. China's capacity accounts for 70% of the global total, but it is mainly assembly and OEM, with lower technical barriers. The industry has already shown obvious involution, and gross profit margins continue to be under pressure.
On the electrical chip side, DSP (Digital Signal Processor) chips are responsible for signal anti-attenuation and recovery, accounting for about 30% of the cost of 800G optical modules. They rely almost entirely on overseas suppliers such as Marvell and Broadcom, with a domestic production rate close to zero. This is also the important industrial background for the rise of the LPO (Linear-drive Pluggable Optics) solution—removing the DSP can bypass this core chip that currently cannot be domestically produced.
The next-generation form is the CPO optical engine, which eliminates the independent housing and co-packages optical chips with switching chips, significantly shortening the electrical signal transmission distance, reducing power consumption, and improving integration density. Currently, CPO has moved from the laboratory to the commercial verification stage. Major global cloud vendors and chip companies are promoting standard formulation and pilot deployment; it is no longer just a long-term concept.
Optical signals are finally transmitted through optical fibers. Optical fibers are only responsible for photon transportation and do not participate in optoelectrical conversion. In data center optical interconnect scenarios, their value proportion is extremely low, belonging to supporting consumables, which is completely different from the core status of optical fibers in telecom transmission networks. Short-distance scenarios mainly use multimode optical fibers, while medium-to-long-distance AI cluster cross-cabinet scenarios use single-mode optical fibers.
The above five layers constitute the industry chain skeleton of optical interconnects, but beyond the skeleton, there is the common bottleneck of optical packaging and testing equipment that runs through multiple layers.
After completing wafer manufacturing, optical chips must undergo precise post-processes such as die attach, wire bonding, lid sealing, optical coupling, and burn-in testing to become usable devices. The high-end die attachers, couplers, and other equipment relied upon by these processes are about 90% dependent on imports, with delivery cycles as long as 18 months, directly limiting the expansion pace of the optical device and optical module links.
Looking at the optical interconnect industry chain, there are currently seven key directions with high attention, corresponding to the nodes with the tightest supply-demand and most concentrated value in the current industry: optical modules, optical devices, opto-electronic components, optical packaging and testing equipment, optical materials, DSP electrical chips, and high-speed optical chips.
These seven directions, from downstream assembly to upstream chips, see value, barriers, and scarcity climbing level by level, jointly constituting the most noteworthy supply tension map of the current optical interconnect industry chain. Among them, the optical chips at the very upstream are the core lifeline in the AI computing power era, and also the biggest variable in the redistribution of industrial profits in the next three years.
Penetrating the Optical Interconnect Market to Understand Long-Term Trends
Driven by high-density AI computing clusters, optical interconnects have been upgraded from peripheral supporting facilities in data centers to core bottlenecks, and the industrial logic and market cognition surrounding this track remain to be clarified.
First and foremost, this round of growth in the optical interconnect industry is essentially a rigid substitution forced by physical laws, rather than AI concept hype.
The so-called rigid substitution refers to the irreconcilable triangular contradiction formed by copper interconnects among transmission rate, transmission distance, and power consumption: the higher the rate, the shorter the transmission distance, and the more severe the heat generation. This is determined by electromagnetic physical laws and cannot be avoided through process optimization. When AI clusters expand from thousands of cards to tens of thousands, the data exchange volume between GPUs grows exponentially, and copper wires are already incompetent within a few meters.
In contrast, optical signals have extremely low transmission loss in optical fibers, are immune to electromagnetic interference, and a single optical fiber can carry dozens of signals simultaneously through wavelength division multiplexing, with an information density far higher than copper cables. There is no choice of "who is better" between the two; rather, in high-density computing clusters, the copper solution has failed, and light has become the only viable physical path. Therefore, the increase in the penetration rate of optical interconnects is certain, and the only variable lies in the pace of technological iteration and the periodic fluctuations in capital expenditures of cloud vendors.
Many believe that rigid substitution will bring a window for domestic substitution, but this needs to be viewed separately.
There is currently no substitution issue in the downstream. Among the global TOP 10 in optical module manufacturing, China accounts for 7, and China is already the undisputed leader. In emerging module technology routes, CPO and LPO started simultaneously at home and abroad, so there is no issue of catching up.
In the midstream, the localization progress of optical devices, components, and packaging and testing equipment varies. Domestic passive devices have achieved global competitiveness, with significant capacity and cost advantages, but there is still room for import substitution in high-end miniaturized products. Active components TOSA/ROSA are limited by the supply of upstream optical chips, and the localization rate in high-speed scenarios remains low; the breakthrough pace depends on the progress of optical chips. Optical packaging and testing equipment is about 90% dependent on imports, with delivery cycles as long as 18 months, which is the biggest shortcoming in the current midstream and an invisible bottleneck for capacity release.
The situation of upstream optical materials and optical chips is relatively complex.
Optical chips are actually the most core battlefield for domestic substitution, directly determining module performance and cost, and also the link with the highest barriers and tightest choke points. Currently, CW (Continuous Wave) light sources have been delivered in batches, 100G EML has been mass-produced, and 200G EML is under verification. In high-end categories such as EML, the global supply-demand imbalance has provided a 2- to 3-year window period.
At the material level, over 90% of the global capacity for indium phosphide substrates is monopolized by three overseas manufacturers, with a domestic production rate of less than 5%. However, the competitive landscape of different materials is completely different and cannot be confused. Thin-film lithium niobate presents a different picture: China is a global hub for lithium niobate manufacturing, with capacity accounting for 42% of the global total. Jinan Jingzheng occupies 78% of the global supply of thin-film lithium niobate wafers. In this specific material direction, China is already globally dominant. Silicon photonic materials have been laid out domestically, but large-scale mass production capacity is still catching up.
Zhongke Yucheng Investment believes that in the capital market and industrial narrative, "optical modules" and "optical chips" are discussed too much and too broadly, which to some extent covers up the real differences between different links in optical interconnects. The true difficulty of industrial breakthroughs often lies in those links located further upstream and less discussed; those are the real choke points. This deviation in attention and cognition also easily brings emotional fluctuations and cognitive misjudgments to the market.
For example, the over-optimism regarding the substitution pace of CPO. Some believe that CPO will quickly eliminate traditional pluggable optical modules, but the objective reality is that the two will coexist for a long time in the next five years. The technological maturity and ecological support of CPO are not yet perfect, and it currently only adapts to top-tier supercomputing cluster scenarios; existing data centers and small-to-medium computing clusters will continue to use pluggable modules for a long time, and the substitution pace is much slower than market expectations.
There is also the expectation that silicon photonics will completely replace indium phosphide—this expectation also deviates from the ultimate end of the industry. The actual situation is that multiple technological routes will coexist for a long time, and no single technology can dominate the whole situation. In short-distance scenarios, silicon photonics have a cost advantage; in medium-to-long-distance, high-computing-power, and high-speed scenarios, the performance of indium phosphide EML is irreplaceable; and thin-film lithium niobate is exclusive to ultra-high-speed coherent scenarios above 3.2T.
Finally, there is the view of AI optical interconnects as a non-cyclical, perpetually high-growth track. The growth attribute has indeed raised the long-term center of the industry, but the cycle has not disappeared. After the concentrated release of high-speed optical chip capacity from 2027 to 2028, optical modules will face the pressure of overcapacity and price wars, and gross profit margins are highly likely to decline. At the same time, export controls on high-end optical chips in North America and quarterly capital expenditure adjustments by cloud vendors will create fluctuations on the demand side.
Optical interconnects have reached today, and industrial consensus has been formed. Copper retreats and light advances, which is determined by physical limits. But consensus itself is hard to generate opportunities. In the view of Zhongke Yucheng Investment, what truly widens the cognitive gap is the structural differences under the consensus. In the same track, the supply-demand states of upstream, midstream, and downstream are completely different, and the competition logic is also completely different. And opportunities are often hidden in the differentiation within the track.
As a private equity fund company focusing on investing in cutting-edge technology in China, Zhongke Yucheng Investment takes investing in leaders, boutique projects, and excellence as its core philosophy. With cutting-edge technology projects from the Chinese Academy of Sciences and its institutes as core targets, and pursuing the ultimate "project exit rate" and "return on investment", it is committed to becoming a benchmark institution in the field of cutting-edge technology investment, co-creating value with outstanding scientists, entrepreneurs, and investors.