$Zhongji Innolight Co.,Ltd.(300308)$
$Eoptolink Technology Inc.,Ltd.(300502)$
$Yuanjie Semiconductor Technology Co.,Ltd.(688498)$
$Suzhou Tfc Optical Communication Co.,Ltd.(300394)$
$Luxshare Precision Industry Co.,Ltd.(002475)$
This USB drive-like device has become a leader in the capital markets, driving the ChiNext Index to a new high in nearly eleven years and pulling the main stage of AI computing power hardware back to China.
In 2025, among the global top ten optical module vendors, seven are Chinese companies. For high-speed optical modules above 800G, China’s two leading vendors alone hold more than 60% of the market share. However, the dividends of the global optical module market have not been evenly distributed to every participant. Some vendors are feasting on big profits, some can only sip the broth, and a few are even seeing their performance decline against the trend.
Why is there such differentiation? Which companies in the optical module industry chain can continue to feast? With traditional manufacturing giants like Luxshare Precision (002475.SZ, ~CN¥67) entering the field, will the optical module sector follow the path of lithium batteries and photovoltaics and suffer from severe overcapacity? More importantly, how will CPO reshape the entire industry’s profit structure?
These questions touch the core of the industry’s development. In this video, we will decode them in depth.
Over the past five years, the data throughput of computing chips has grown exponentially, but the data transmission capabilities of networking equipment have not kept up. What is now bottlenecking the entire computing power system is no longer the GPU, but the optical module.
We know that GPUs and CPUs perform computation and data transfer inside servers using electrical signals. But to transmit those signals outward over long distances, they must be converted into optical signals and sent through optical fibers. The optical module is the key device that converts electrical signals into optical signals.
For a long time, the biggest buyers of optical modules were traditional telecom operators. Telecom networks mainly handle communication between people and the network. Streaming videos or making calls could be handled by low-speed optical modules below 100G. Because the technology is mature and barriers are low, the telecom market long ago became a brutal red ocean of cost and price competition.
The logic of AI data centers is completely different. In a 10,000-GPU cluster, the entities sending and receiving data are no longer humans but thousands of GPUs devouring data at insane speeds. A 100G “sewer pipe” simply cannot handle this level of data flood. It requires mandatory use of 400G, 800G, and 1.6T high-speed optical modules.
Once the shift to high-speed modules happens, the gap between optical module vendors becomes fully apparent. Moving from low-speed to high-speed appears on the surface to be just a speed upgrade, but in reality, the manufacturing threshold rises to an entirely different level.
To help everyone understand this more intuitively, let’s open up an optical module and take a look inside.
The internal structure of an optical module mainly consists of two parts: the electrical signal area near the server or switch side, and the optical signal area near the fiber side.
The core components in the electrical area are the PCB circuit board and the Digital Signal Processing chip (DSP) on it. The DSP is responsible for using algorithms to restore distorted electrical/optical signals back into accurate data after transmission, while also optimizing the signal on the transmitting end to ensure reliability.
The optical signal area consists of three components:
The Transmitter Optical Sub-Assembly (TOSA) — the cylindrical or square metal housing in the image. Inside it is encapsulated the scarcest resource in the entire industry chain: the laser chip. Its core job is to convert the electrical signals from the server or switch into optical signals.
The Receiver Optical Sub-Assembly (ROSA). Its core is a high-sensitivity photodetector chip paired with optical components. It converts optical signals coming from the fiber back into electrical signals to feed to the circuit board.
The passive optical devices between the transceiver components and the front-end fiber interface. Passive devices are pure physical optical parts that function without electricity — such as ultra-precise micro-lenses, isolators, and wavelength division multiplexers. They rely on physical refraction and reflection to combine different colored lights into one fiber or precisely focus the beam into a fiber core thinner than a human hair. Although they don’t require power, this part extremely tests a vendor’s micron- or even nanometer-level optical alignment and precision packaging capabilities.
This naturally brings us to active devices. The laser chips in the TOSA and the detector chips in the ROSA we mentioned earlier are active devices. They must be driven by external power and serve as the core engines that convert between electrical and optical energy.
In simple terms, what an optical module vendor does is assemble the PCB and DSP from the electrical area with the transceiver components and passive devices from the optical area into the complete module we see, then sell it to cloud vendors or telecom operators.
The optical module industry is now very mature, with clear division of labor across segments. There are specialized upstream optical component makers, DSP chip designers, core optical chip designers and manufacturers, and professional firms that package optical components into passive devices and optical chips into transceiver sub-assemblies or optical engines.
Under this extreme division of labor, optical module vendors don’t need to become all-rounders who manufacture their own laser chips, DSPs, or various optical components. They can meet most needs simply by purchasing externally.
This is especially true in the traditional low-speed telecom optical module market, where many vendors simply buy parts from specialized suppliers, assemble them, and sell the finished modules.
However, it is precisely this seemingly simplest “assembly” step that has determined the competitive landscape in the high-speed optical module era — deciding who gets to feast and who only gets the broth.
In terms of revenue scale in 2025, Zhongji Innolight (300308.SZ, ~CN¥1,200–1,250) and Eoptolink (300502.SZ, ~CN¥540–550) are far ahead in a cliff-like lead, and the gap continues to widen. Yet back around 2018, Chinese optical module vendors were basically all starting from the same line. It was only after the introduction of 400G modules that Zhongji Innolight and Eoptolink rapidly pulled away from the others.
So the question is: why can these two pull so far ahead when they are all doing assembly work?
Many people think it’s because they were the first to lock in major U.S. cloud customers. But that’s only the surface appearance.
When North American cloud giants fill their AI data centers with expensive NVIDIA GPUs, they cannot afford to be held back by optical modules that cost just a few hundred dollars. So when modules upgraded to 400G, the cloud giants’ supplier selection became extremely stringent.
In summary, microscopic packaging capability and macroscopic system-level delivery capability are the key factors the cloud giants evaluate — and the real reason for the gap between optical module vendors.
First, packaging capability. Many people think assembling an optical module is just soldering a circuit board. In the 400G/800G era, it is actually an extremely complex systems engineering task, usually divided into first-level (primary) packaging and second-level (secondary) packaging.
The process of assembling purchased optical sub-assemblies with DSP chips mainly tests secondary packaging capability. In the high-speed era, you must stuff a high-heat DSP chip and a heat-sensitive laser assembly into a USB-drive-sized metal box. This requires high-frequency circuit design to prevent electrical signals from interfering with each other, and sophisticated thermal-fluid dynamics design to perfectly dissipate heat. If cooling fails, the module will crash within minutes once plugged into the server.
Zhongji Innolight and Eoptolink were the earliest to overcome the challenges of high-frequency circuit design and extreme thermal management in secondary packaging, which earned them entry tickets to major customer testing.
However, secondary packaging alone is not enough. The more fundamental first-level packaging capability is even more critical. First-level packaging means precisely assembling extremely fragile bare optical chips with micro-lenses and other passive devices into optical transceiver sub-assemblies or optical engines. This is like carving on a hair — micron-level optical alignment that requires expensive automation equipment and profound materials and process expertise.
Most optical module assembly plants cannot do this well, or cannot achieve the extremely high yields needed for mass production. China’s Tianfu Communication (TFC, 300394.SZ, ~CN¥300–320) core business is precisely this first-level packaging. This is why, even though Tianfu’s revenue scale is much smaller than Zhongji and Eoptolink, the capital market still ranks it alongside them as the “Yi-Zhong-Tian” first tier of optical communications. Tianfu’s first-level packaging yield directly determines the basic quality of the entire module — it is the indispensable upstream “water seller” in the high-speed era. We’ll discuss this more in the CPO section later.
After covering microscopic packaging, let’s look at macroscopic system-level delivery capability.
Many have the misconception that as long as you can make the optical module, you can just take samples and sell them to big customers. In the 100G era, optical modules were like standardized USB drives — plug and play. But in the 800G and 1.6T era, this plug-and-play logic starts to fail. Optical modules have become highly customized devices.
North American cloud giants run the latest NVIDIA GPUs and Broadcom’s top switch chips in their AI clusters. How the DSP firmware algorithms inside the optical module are written to achieve ultra-low latency interconnection with NVIDIA NICs, and how the power consumption curve is designed to perfectly match a specific cloud vendor’s liquid-cooled data center environment — this kind of deep system-level co-tuning cannot be done after the product is already made.
Why has Zhongji Innolight secured the top spot? Because long before North American tech giants release their next-generation computing chips, their R&D teams are already working in the same labs with these giants, following the next-generation roadmap for underlying customized development. This joint R&D entry ticket is not something second- or third-tier vendors can easily obtain.
After customized development comes the test of large-scale delivery. Cloud giants buy optical modules to connect GPUs or ASICs worth tens of thousands of dollars each. They cannot accept expensive computing power sitting idle due to network failures or parts shortages. What they need is not lab samples, but the ability — after winning the bid — to mass-produce hundreds of thousands of modules the next month, with yields still above 99% even in the extreme high-temperature and high-humidity environments of the data center. If delivery fails and AI cluster construction is delayed, the cloud giants’ losses are measured in tens of millions of dollars.
It is precisely because Zhongji Innolight and Eoptolink crossed the thresholds of first- and second-level packaging and system-level delivery early that they secured seats at the North American cloud giants’ table. By the time other vendors saw the AI wave coming and rushed to build production lines and send samples, the table was already packed with top players.
As the profitability of the optical module business became evident, traditional manufacturing giants like Luxshare Precision also announced they would enter the field. So the question arises: will the optical module industry repeat the overcapacity mistakes of lithium batteries and photovoltaics?
This is the question the capital market cares about most. The overcapacity in photovoltaics and lithium batteries was essentially overcapacity of homogeneous production. As long as you were willing to spend money on equipment, everyone could produce battery cells or solar wafers with similar conversion efficiency, leading only to brutal price wars.
The logic of optical modules is completely different. Its effective capacity is not determined by how many production lines module vendors build, but is tightly bottlenecked by upstream “water sellers.”
Even if cross-industry giants build countless cleanrooms, without core electrical and optical chips, those lines cannot be converted into effective capacity. In the hardware cost of high-speed optical modules, opto-electronic chips take away more than 50% of the profit.
We already mentioned the electrical chips: high-heat digital chips like DSPs are basically oligopolized by U.S. giants Broadcom and Marvell. On the optical side, there is also an oligopoly.
To help everyone understand the impact of optical chips on the entire industry, let’s first build a basic framework.
In optical modules, optical chips are divided into laser chips for the transmitter and photodetector chips for the receiver. Since photodetectors are relatively mature, we won’t expand on them. This video focuses mainly on transmitter-side optical chips.
Current mainstream technology routes for optical chips are mainly DML, CW, EML, and silicon photonics.
At 100G low speeds, DML lasers were mainly used. The principle is simple — like turning a flashlight on and off rapidly to send Morse code (electrical signals converted to light). But when computing power explodes and speeds are forced to 400G, 800G and above, the switch has to toggle billions of times per second. The traditional DML approach no longer works because the light becomes severely distorted.
The industry introduced two high-end solutions:
Separate the flashlight and the shutter. Keep a Continuous Wave (CW) light source constantly on, then add a high-speed physical shutter (modulator) in front. Current modulators mainly include Thin-Film Lithium Niobate (TFLN) modulators (which have potential to break speed limits) and modulators highly integrated on silicon photonics chips.
Electro-absorption Modulated Laser (EML). This integrates the constant light source and ultra-fast shutter into one tiny piece of material. One side continuously emits light, the other side rapidly absorbs or transmits it. It is currently the mainstream, highest-demand, and highest-yield-threshold solution for 800G and 1.6T modules.
Silicon photonics is a brand-new route we will discuss in detail later.
In current 800G modules, the transmitter sub-assembly packs eight 100G or four 200G high-speed laser chips in parallel. This makes high-end 100G/200G EML chips, high-power CW light sources, and future TFLN modulators extremely scarce and difficult to produce.
This is not only because multiple chips must work perfectly in sync at extreme speeds, but because they are tightly constrained by underlying materials science and epitaxial growth processes. For example, the Indium Phosphide (InP) substrates used in high-end lasers have extremely difficult-to-control crystal growth and processing yields. This is true “mystical” craft that relies on veteran engineers’ experience rather than simply buying better equipment.
As a result, global high-end optical chip capacity and pricing power have long been firmly held by a few overseas oligopolies such as Sumitomo, Lumentum, Coherent, and Broadcom. NVIDIA’s multi-billion-dollar investments in Lumentum and Coherent in early 2026 were fundamentally aimed at locking up their limited optical chip capacity in advance.
For leading optical module vendors to achieve mass production and expansion, they must pay deposits early to secure capacity, otherwise they risk supply disruptions. This is why upstream optical chip companies command such high valuations in the capital markets.
In reality, only a few companies have truly broken through the R&D of high-end optical chips and achieved stable mass production. Currently, Yuanjie Technology (688498.SS, ~CN¥1,400–1,600) is the first domestic company to achieve mass production of high-end 100G PAM4 EML chips, and its 100mW CW laser chips are also in volume shipment. Additionally, Dongshan Precision acquired a Silicon Valley company with 20 years of history in optical chips (Source Photonics), whose 100G PAM4 EML chips have also achieved volume delivery.
At present, Yuanjie Technology and Dongshan Precision are the only two domestic companies that have achieved stable mass production of high-end laser chips. Other domestic firms have optical chip businesses, but there is no official information indicating they have reached mass production of 100G+ high-end chips.
Therefore, to judge whether a company’s valuation has a bubble or real substance, the key is whether its strongest products have been purchased in volume by top or second-tier customers and translated into actual revenue.
At this point, you can understand why traditional manufacturing giants’ entry can at most turn the low-speed optical module market into a red ocean, but in the truly profitable 800G/1.6T high-end battlefield, as long as the upstream barriers for core electrical chips, high-end optical chips, and modulators remain, the industry is unlikely to suffer severe overcapacity just because downstream module vendors expand production.
(Of course, if North American cloud giants suddenly slash capex collectively, all AI hardware could face relative overcapacity — but that’s a separate topic.)
Since traditional Indium Phosphide materials are so difficult and capacity-constrained, will global AI computing power be choked by a few materials companies? Tech giants will not sit idly by.
To break this artisanal materials science bottleneck, the silicon photonics route emerged.
@Daily_Discussion @Tiger_comments @TigerPM @TigerStars @TigerObserver @MillionaireTiger
Comments