On May 12th, after surpassing a market capitalization of one trillion yuan, Zhongji Innolight Co.,Ltd. reached another milestone—its stock price broke through 1,000 yuan, becoming the second stock in the history of the ChiNext board to achieve this. The previous ChiNext board stock to reach this level was Imeik Technology, a leading domestic player in the medical aesthetics sector, which supported a market cap of hundreds of billions with hyaluronic acid injection products. However, its stock price has since fallen by nearly 80% from its peak. The sector now holding this position is optical modules.
Since 2026, driven by the expansion of AI computing infrastructure, optical modules have become one of the hottest sectors in the A-share market. Three leading stocks, playfully nicknamed "Yizhongtian" by the market—Eoptolink Technology Inc.,Ltd., Zhongji Innolight Co.,Ltd., and Suzhou Tfc Optical Communication Co.,Ltd.—have seen particularly strong gains.
The phrase "You should stand in the light, not just stand there," a homophone pun among investors, reflects the ongoing "optical module rally" of 2026. It also hints at the underlying market sentiment: the strategic value of optical modules in the AI industry chain is being repriced. This represents a rare moment in the spotlight for "Made in China" within the global computing supply chain.
While domestic companies are still catching up in the field of core AI chips, optical modules represent one of the few hard-tech sectors where Chinese firms genuinely occupy the global first tier and secure orders from leading clients.
If we compare an AI data center to a precision machine, its operation relies on three core elements: computing power, storage power, and transmission power. In computing power GPUs, NVIDIA holds a near-monopoly. In high-bandwidth memory (HBM), Samsung, SK Hynix, and Micron dominate. In these two areas, Chinese companies are still in pursuit. However, in "transmission power," Chinese companies have already moved to the global forefront.
According to LightCounting and Dell'Oro 2025 data, Zhongji Innolight Co.,Ltd. holds over 28% of the global optical module market share, firmly securing the top global position. Chinese manufacturers account for more than half of the global top 10 optical module vendors.
However, the more lucrative the business, the more participants and competitors it attracts. As the stock prices of "Yizhongtian" and other optical module companies repeatedly hit new highs, a technological revolution targeting optical modules has been quietly brewing. Tech giants like NVIDIA and Alphabet are exploring next-generation optical interconnect solutions, such as Co-Packaged Optics (CPO) and Optical Circuit Switching (OCS), which differ from traditional "pluggable optical modules."
If the routes pursued by NVIDIA and others succeed in large-scale applications, the "Yizhongtian" companies currently in the spotlight may face not just erosion of market share but a potential redefinition of the entire sector. The celebration continues, but a transformation has already begun. How long can those "standing in the light" remain there?
01 Do NVIDIA and Others Aim to Revolutionize "Optical Modules"?
In recent years, the industry has diverged into several distinct technological paths centered on making optical interconnects "faster, more efficient, and closer."
Optical modules are primarily plugged into servers and switches, responsible for "sending" and "receiving" data, transmitting it at high speeds between different devices via optical fibers to handle the massive flow of information within data centers. What NVIDIA aims to do is embed this "entry/exit point" directly into the chip.
At the 2025 GTC conference, NVIDIA officially launched its silicon photonics-based CPO network switches—Spectrum-X Photonics and Quantum-X Photonics. The core idea is to integrate optical communication capabilities deeper, closer to the switch chip, to reduce power consumption and transmission losses in hyperscale AI clusters. In October of the same year, NVIDIA announced that Meta Platforms, Inc. and Oracle would adopt Spectrum-X Ethernet switches for building hyperscale AI factories. CPO evolved from a technical direction into an industrial reality.
In March 2026, NVIDIA announced a $2 billion investment each in optical communication companies Lumentum and Coherent, totaling $4 billion. This indicates that NVIDIA is not only developing CPO technology in-house but also directly binding upstream laser and optical component suppliers through capital, integrating the optical interconnect supply chain into its own ecosystem.
At the 2026 GTC, NVIDIA further clarified the role of CPO in next-generation AI data centers. Jensen Huang emphasized that as AI clusters continue to expand, data center internal connections will rely on both copper cables and optical interconnects; among these, CPO will handle critical connection tasks requiring higher bandwidth and lower power consumption.
However, what Jensen Huang did not mention is that CPO has its own challenges. Traditional pluggable optical modules can be replaced individually, but CPO deeply binds the optical engine to the switch chip. Once a failure occurs, the complexity and cost of repair increase significantly. Additionally, thermal management under high heat density, fiber coupling, high-precision packaging, and yield rates for mass production are all issues that must be resolved before CPO can be truly commercialized.
Alphabet, on the other hand, has taken a different, more unconventional path: OCS. The logic of Alphabet is: since photoelectric conversion itself is a source of loss, try to avoid conversion as much as possible. Using MEMS micro-mirrors to complete path switching directly in the optical domain, "reflecting" the light beam to its destination, bypassing the intermediate electrical signal stage. Alphabet has introduced its self-developed OCS technology in TPU v4, allowing connection paths between different TPUs to be reconfigured according to task needs. This enables data to flow more flexibly between chips during AI training, making clusters easier to scale while reducing some power consumption and improving overall utilization.
Although Alphabet uses OCS, it still requires fundamental components such as optical devices, fibers, and transceiver modules/optical engines. However, OCS is not suitable for all traffic types. It is better suited for large-scale, relatively regular, and schedulable data transmissions, such as certain cluster communications in AI training.
Beyond NVIDIA and Alphabet, giants like AMD, Broadcom, and Cisco have long been laying out plans around CPO and silicon photonics interconnects. More radical are a group of silicon photonics startups—Ayar Labs, Lightmatter, Celestial AI—targeting optical interconnects between GPUs and between chips and memory. If this path succeeds, the entire organization of AI computing systems could be rewritten. Of course, this is also the highest-risk, longest-cycle bet among all routes.
Different routes and bets are backed by different giants with different strategic logics. They will not determine a winner in the short term, but the direction is clear: the battlefield for optical interconnects is gradually shifting from the "front panel" to "inside the chip."
02 Outcome Uncertain, A Crisis and an Opportunity
If innovative routes like CPO achieve widespread application in the future, the most direct impact will be on companies like Zhongji Innolight Co.,Ltd. and Eoptolink Technology Inc.,Ltd., whose core business revolves around traditional pluggable optical modules. Their current revenue heavily depends on the continued procurement of 800G, 1.6T optical modules and related products by cloud providers like NVIDIA, Alphabet, and Microsoft.
In 2025 alone, Zhongji Innolight Co.,Ltd.'s revenue reached ¥38.2 billion, a year-on-year increase of 60%; Eoptolink Technology Inc.,Ltd.'s net profit surged by 236% year-on-year. The underlying logic of this rapid growth is based on the premise that "pluggable optical modules remain the mainstream."
However, the "Yizhongtian" companies are also alert to these changes. All three are deploying in CPO, albeit from different angles. Suzhou Tfc Optical Communication Co.,Ltd. positioned itself earliest, with its core products—Fiber Array Units (FAU) and silicon photonics engine packaging components—being highly relevant to the photoelectric packaging required for CPO. It has appeared as TFC on NVIDIA's Spectrum-X Photonics partner list. Zhongji Innolight Co.,Ltd. has already laid out plans in the CPO direction, with the penetration rate of silicon photonics technology continuously increasing. Eoptolink Technology Inc.,Ltd. stated during its April 2026 performance briefing that it has built a technological system covering various interconnect forms, including pluggable optical modules, LPO/LRO, XPO, NPO, and CPO.
In other words, this battle over technological routes is not just a threat for optical module giants; it also presents an opportunity to enter new tracks by riding the trend. The real risk lies in managing the transition pace—transforming too slowly risks erosion of existing market share, while betting too heavily may impact the cash flow from traditional business.
Furthermore, this impact is far from as decisive as some external narratives suggest. Several A-share listed companies have confirmed through communication with major cloud service providers that network planning for 2026-2027 still centers on pluggable solutions, with no cloud provider planning large-scale CPO deployment in the next two to three years. The lifecycle of pluggable optical modules is much longer than the market imagines.
03 How Long Can the "Optical Module Narrative" Last?
Morgan Stanley states that against the backdrop of continuous AI computing power upgrades, the industry story for optical communication is far from over. Currently, there are no signals to overturn the optimistic logic for optical communication. The next round of growth opportunities will still belong to companies with greater advantages in supply-demand dynamics, technology, and profit certainty.
As long as large models are still being trained and GPU clusters continue to expand, data must have a place to flow. But the real test is: maintaining market share under existing technological routes while securing a position when new architectures begin to be implemented and scaled—this is the key to the next phase of competition.
Moreover, we need to understand that beneath this prosperity, several issues require attention.
First is the price pressure brought by capacity expansion. According to Goldman Sachs and LightCounting estimates, 800G and 1.6T capacity will be concentratedly released around 2026. Once downstream AI computing investment slows, the supply-demand balance could quickly be disrupted, and price pressure would directly erode corporate profit margins.
Second, although high-end optical modules are mass-produced primarily by Chinese manufacturers, there remains external dependency for some upstream core components. Taking EML lasers and DSPs as examples, global major suppliers are still concentrated in the hands of American, Japanese, and other manufacturers. For module manufacturers, once upstream supply tightens or prices rise, pressure would first manifest in BOM (bill of materials) costs and gross margins.
However, companies like Zhongji Innolight Co.,Ltd. are increasing the proportion of silicon photonics solutions, attempting to reduce reliance on the traditional EML (Electro-absorption Modulated Laser) path.
Simultaneously, the high prosperity of the optical module sector is attracting more competitors, including "cross-border players." Besides the aforementioned entry of giants like NVIDIA and Alphabet, consumer electronics giants like Luxshare Precision are also participating. They possess mature capabilities in precision manufacturing, supply chain management, and scaled delivery, giving them a foundation to enter the optical module field. Additionally, upstream chip and equipment manufacturers are extending downward, trying to capture more value chain share.
Although high-speed optical modules still have barriers such as customer certification, yield control, reliability verification, and high-speed photoelectric design, making them difficult to replicate simply in the short term. What truly needs vigilance is that once more manufacturers pass major customer testing and enter bulk supply lists, coupled with the continuous release of new capacity, the industry could shift from supply shortage to supply-demand balance. At that point, price competition and gross margin pressure would likely rise accordingly.
In summary, for optical module manufacturers, the next step is not just about expanding production but also about finding new moats amidst technological route shifts, customer binding, upstream component autonomy, and cost control.
Comments