In the early morning of August 14, Nvidia officially announced the mass production of its first CPO switch, and a few hours later when the A-shares opened, the optical module sector experienced turbulence.
The star stock "Yizhongtian" had several days of continuous gains, but on the 19th, Zhongji Xuchuang and Xinyi Sheng led a sharp drop of 9 points. Subsequently, there were three days of declines followed by two days of gains, showing severe differentiation within the sector, marking the complete end of the "Believe in Light" market.
The essence of the CPO is to disassemble and redo the optical modules by adjusting the positions of certain components, removing and adding others, in order to reduce power consumption, lower costs, and increase transmission efficiency. This orderly transition within the optical module supply chain signifies a redistribution of profits.
China's mainland, having occupied a high ground in optical modules, thus faced a crack in its position.
Bears pointed their arrows at the downstream optical module integrators that earn the most, believing that the disassembled optical modules are directly targeted by the CPO, leading to market share erosion and could also affect upstream manufacturers with significantly fewer orders;
Bulls, on the other hand, argued that Nvidia’s rapid advancement of CPO mass production perfectly confirmed the downstream market's extremely high demand for high-speed optical modules. Morgan Stanley and Goldman Sachs both released reports in succession, significantly increasing their forecasts for optical module orders.
Both sides hold their ground, with neither able to persuade the other, condensing countless words into one question: Can optical modules still hit the market?
Not a Near Concern
The first clear industrial consensus is that within the next two years, CPO is unlikely to have a substantial impact on the revenue of optical module manufacturers.
Currently, Nvidia is rushing to advance CPO's implementation mainly to address the "electricity cost" issue of ultra-large-scale data centers.
According to Semianalysis estimates, a deployment of 200,000 GB300 NVL72 clusters results in an optical module power consumption of 17 megawatts, and running at full capacity throughout the year would consume 149 million kilowatt-hours of electricity.
In rough estimates, this is equivalent to the annual electricity consumption of a subway line in Shanghai, and at an electricity price of 0.8 yuan/kWh, the annual electricity costs would exceed 100 million yuan.
As the scale of the cluster reaches 500,000 or 1 million, the electricity costs skyrocket exponentially, adding more pressure to an already stressed power grid system.
CPO significantly reduces power consumption by moving the optical engine next to the switch chip; using Nvidia's CPO switch reduces the power consumption for every 800G bandwidth by more than 70% compared to the power consumption of an 800G DR4 optical module.
However, this does not mean that CPO is the sole solution for all data centers, as the cost-performance ratio must also be considered.
Taking the 200,000 GB300 NVL72 cluster (three-layer network architecture) as an example, the additional costs incurred by switching to CPO nearly entirely offset the reduced costs of optical modules and networking, resulting in only a 3% reduction in total cluster cost and a mere 2% decrease in total power consumption.
Not only did it lead to unnecessary busywork, but it also brought challenges such as increased onsite maintenance difficulties and loss of bargaining power, making it a lose-lose scenario.
Theoretically, the more GPUs and network levels in a cluster, the more the advantages of CPO will be highlighted, making it destined to be a choice for only a select few in the foreseeable future.
TrendForce’s March report predicts that by 2026, the penetration rate of CPO in AI data centers will be around 0.5%, specifically referring to its share in AI data centers, rather than all data centers, resulting in negligible impacts on optical module manufacturers' orders.
Zhongji Xuchuang also stated in an investor event on August 23 that almost all customers' optical module orders have been covered for the whole of 2026, and based on guidance and orders from major clients for 2027, it seems there will be no decrease compared to 2026;
The second-largest optical module company, Xinyi Sheng, was more cautious in its wording but also repeatedly emphasized that orders this year will continue to grow quarter by quarter, with continued prosperity and high growth into next year.
From the perspective of some industry insiders, optical module specifications up to 800G/1.6T still exhibit hard-to-replace cost-performance advantages. According to Zhongji Xuchuang and Xinyi Sheng, their 1.6T optical module products are in the stage of continuous volume increase, constituting the main force of current orders.
The real watershed is at 3.2T or even 6.4T, as the optical modules will encounter an objective problem: the switch's switching board can no longer accommodate more optical modules.
The red box shows the optical module interfaces on the switch; image source: Broadcom
Theoretically, the higher the speed and the more ports required, the more optical modules are needed, but the switching board space is limited, making non-external plug-and-play CPO an inevitable choice.
This also explains why initial industry speculation predicted that Nvidia's first CPO switch would be a 1.6T Quantum-X, but the first mass-produced was actually the 3.2T Spectrum-X, because the latter is where Nvidia truly hopes CPO can make an impact.
Similar to chip manufacturing, where advanced process chips below 7nm must use EUV lithography machines, this does not prevent ASML from shipping the highest volume of DUV lithography machines for mature processes each year.
In the eyes of most industry insiders, the coexistence of CPO and optical modules is highly likely, with CPO used to address the scale-up (horizontal expansion) issues of ultra-large-scale data centers, while optical modules will remain mainstream in large and smaller-scale data centers.
However, while there are no immediate concerns, there are long-term worries.
The evolution of new technologies is similar: they cut into high-end scenarios, reduce costs with scale, and then penetrate other scenarios, reaching ordinary households, eventually eliminating old technologies. Although it is too early to label optical modules as old technology, the momentum of CPO requires one to prepare for the future.
How to seize the opportunity to board the train during the ramp-up window of CPO mass production is a challenge that all optical module manufacturers must face.
But There Are Long-Term Worries
For the mainland optical module supply chain, CPO is not a monolith.
In rough terms, CPO disassembles the three core components of optical modules: the optical engine, DSP chip, and laser. The optical engine is placed next to the switch chip, the laser is mounted externally to the switch chip, and the DSP chip is abandoned.
This means that apart from the switch chip, which is Nvidia’s own, the main components that the optical module supply chain can leverage are just the optical engine and the laser.
First is the optical engine, which has the highest technological barriers and value. The optical engine includes an optical chip (PIC) and an electronic chip (EIC), with the complexity lying in the design and manufacturing of the optical chip and how to package both chips together. The former involves silicon photonics technology (using silicon chip manufacturing technology to produce optical chips), while the latter involves advanced packaging.
In terms of design, Broadcom and Marvell have an undisputed dominant position, while manufacturing is led by chip manufacturing giants like Tower and GlobalFoundries. TSMC, benefiting from Nvidia’s CPO, is likely to catch up, along with IDMs like Lumentum and Coherent that handle both design and manufacturing.
China’s mainland has varying degrees of shortcomings in optical chip design and manufacturing.
The design is slightly better, with companies like Optical Express Technology, Zhongji Xuchuang, and Xinyi Sheng having accumulated silicon photonics technology, theoretically having independent research and development capabilities, but currently, they do not have optical engine-related products in the market;
The manufacturing gap is more evident, as although there have been optical chip foundries like Yuxin Semiconductor and Xingyao Photonics emerging in recent years, their technology, capacity, and reliability fall far short compared to overseas giants.
The advanced packaging technology responsible for integrating optical chips and electronic chips currently relies heavily on TSMC, and Broadcom, which has jointly developed the CPO switch with Nvidia, is said to also rely on TSMC for its next-generation products.
Secondly, there are lasers. The optical engine controls light, while the laser generates it; the latter can be divided into design, manufacturing, and packaging stages as well. China’s mainland is somewhat more optimistic regarding technological reserves in lasers than with optical engines.
In terms of design and manufacturing, companies like Source Photonics and Optical Express Technology have some products that can compete with overseas counterparts, but most high-end products still lag behind first-tier manufacturers like Coherent and Lumentum, with some not even in production yet, holding a very small share in the global market.
The packaging of lasers is quite special and is the part of the supply chain in which the mainland holds the most leverage.
According to the initial conception of CPO, lasers were supposed to be moved next to the optical engine, but due to issues like heat dissipation, yield rates, and ease of replacement, the current mainstream solution is to mount them externally to the switch chip, creating independent laser modules (ELSFP).
The red box indicates the external laser module
This requires module integration and precision manufacturing technology; Tianfu Communication, currently the lowest-valued among the “Yizhongtian” stocks, has become one of the first five core suppliers announced by Nvidia for CPO.
Nvidia announced its five core suppliers for CPO mass production on the same day
According to Tianfu Communication's financial reports and investor event records, besides the laser module, another core component it supplies for CPO is FAU (fiber array), which simply means drawing out fibers from the optical engine manufactured by TSMC to connect with the outside world.
Another aspect frequently mentioned by institutions is CPO testing equipment.
Due to the addition of optical chips, optical tests need to be incorporated alongside the traditional chip testing steps, and traditional chip testing equipment cannot fully adapt to CPO, requiring a certain degree of redesign and modification.
The recent surge in A-shares can be attributed to Roboteq’s acquisition of ficonTEC, whose WLT-D2 uses a dual-sided electro-optical design, enabling simultaneous electrical and optical testing, thus becoming a hot commodity in the secondary market.
In summary, while mainland manufacturers do indeed gain a share from CPO, it is minor, mostly revolving around the peripheral areas of the value chain’s core layer.
The crux lies in the fact that compared to the downstream packaging and delivery capabilities of optical modules, the value chain of CPO clearly shifts towards the upstream components like optical engines and lasers, which are precisely the areas where mainland China has yet to complete its homework regarding optical modules, a long-standing ailment.
According to TrendForce's March projection, by the end of 2028, the penetration rate of CPO in AI data centers will linger below 10%, but then leap upward, reaching over 35% by the end of 2030.
Five months after the report was released, Nvidia announced the mass production of CPO switches, and Broadcom’s next-generation 51.2T Bailly CPO switch began small-scale deliveries in July, progressing faster than the general expectations in the industry.
Over the next two years, this will be a critical period for the ramp-up of CPO mass production, and potentially the last chance for mainland manufacturers to catch up.
Conclusion
In June this year, a groundbreaking ceremony was held for the expansion of Coherent’s Texas factory, where Jim Anderson, the CEO of Coherent, stood in the center on stage, with Jensen Huang bending low to his right.
Breakground ceremony for Coherent’s Texas factory; Jensen Huang is second from the left and Coherent’s CEO is in the center
The last time Jensen Huang attended a groundbreaking ceremony was three years ago at TSMC's Arizona factory. Now, Coherent's importance to Nvidia is comparable to that of TSMC.
Coherent’s laser technology leads globally and is a core supplier that both optical modules and CPO cannot bypass. It not only supplies Nvidia's CPO but also influences the shipments of high-speed optical modules for Zhongji Xuchuang and Xinyi Sheng.
In March of this year, Nvidia strategically invested $2 billion in Coherent, which was coincidentally at the same time as its investment in another laser giant, Lumentum.
Jensen Huang’s stance largely reflects the current elevated position of component manufacturers within the AI supply chain.
Rapidly changing demands accelerate the iteration of downstream product forms; the closer to upstream, the nearer one gets to certainty. The downstream integrators continue to change while the positions of upstream core component manufacturers remain solid as a rock.
Each irreplaceable component is reshaping the discourse power of the entire chain, redistributing the profits of the industry.
The mainland's absolute leading position in optical modules, which excels in integration and manufacturing capabilities, is a continuation and gift of this advantage. In the coming years, this advantage will continue to bring endless profits to the industry. However, it is time for these profits to be reinvested upstream.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
