[Relevant Companies Mentioned in This Article] Lumentum, NVIDIA, Broadcom, AMD, Marvell, Ayar Labs, Coherent, Yuanjie Technology, JX Metals, Sumitomo Electric, AXT
CPO Mass Production Delayed? Supply Chain and Testing Accelerate First
First, let's clarify a fact: the recent leaks from SemiAnalysis primarily focused on the issues with scale-up CPO optical interconnects and the mid-plane PCBs for Kyber NVL144. While this poured cold water on the market, it also prompted deeper reflection within the industry. Although CPO mass production was seemingly imminent, the reality is that NVSwitch will have to wait until the Feynman generation. This delayed expectation to 2028 provides a two-year window for the pluggable optical transceiver and NPO (Near-Package Optics) camps. However, according to Lumentum's latest disclosure, the CPO solution is advancing steadily along the established roadmap. The CW (Continuous Wave) laser co-developed with NVIDIA is expected to be delivered in mass quantities in the fourth quarter of this year, with further volume ramp-up in mid-2027.
TSMC (Taiwan Semiconductor Manufacturing Company) recently announced its PIC (Photonic Integrated Circuit) capacity, which currently stands at 5k pcs/month. It is expected to reach 10k pcs/month by the end of 2027 and at least 25,000 pcs/month by the end of 2028. Given the current tight capacity, major clients for the COUPE platform are basically locked in as NVIDIA, Broadcom, and AMD. Next year, capacity will be gradually opened to clients such as Marvell and Ayar Labs. In TSMC's CPO supply chain, TSMC handles the front-end Chip on Wafer process of CoWoS (Chip on Wafer on Substrate), while the back-end on Substrate process is undertaken by ASE (Advanced Semiconductor Engineering). In this model, TSMC first bonds the PIC and EIC (Electronic Integrated Circuit) and delivers them to ASE in wafer-level format. The stacking structure typically involves the PIC bonded at the bottom and the EIC on top. ASE then dices the wafer and mounts it onto an ABF IC substrate. Additionally, ASE is responsible for installing heat sinks and assembling optical components such as FAUs (Fiber Array Units) and couplers, ultimately delivering the complete OE (Optical Engine) products with pigtails to the testing stage.
Meanwhile, the author has also gathered some non-public information indicating that TSMC's current capacity layout spans multiple facilities. In addition to the Longtan fab, which possesses manufacturing capabilities for advanced packaging and bonding processes, newly built parks in Hsinchu and Kaohsiung have also been scheduled for production. In the future, Chiayi and Tainan will also be key layout locations. For the back-end process, ASE is distributed across two locations, advancing capacity and yield ramp-up. During the early stage of mass production, the production lines of SPIL (Siliconware Precision Industries) in Taichung will be used first, and subsequent capacity expansions are planned to be located in Kaohsiung. Beyond capacity planning, the yield rate for CPO has not yet reached the critical threshold of 90% at this stage. The EIC wafers primarily adopt a 6nm process node, while the PIC uses a 65nm process node. However, the excessively low overall yield rate consumes a tremendous amount of manpower and resources, and blindly expanding capacity would incur substantial costs.
Looking at the bottleneck at the testing level, the main constraint lies in the testing time per single wafer. Taking Insertion testing as an example, Insertion 2 wafer-level optoelectronic joint testing required approximately 22 hours in the fourth quarter of 2025, whereas it has now been shortened to 6 hours per wafer. With the increase in interface speeds and the upgrade of synchronous parallel probe cards, it is expected to be reduced to 3-4 hours per wafer. This testing phase is also an essential step in determining the product yield after the bonding of EIC and PIC. Meanwhile, leading manufacturers in Taiwan, such as WinWay, are simultaneously developing probe card equipment for 3nm and even 2nm nodes.
Another piece of market information indicating volume ramp-up in the silicon photonics market is that Aehr Test Systems stated in its earnings call for the fourth quarter of this year that its performance far exceeded market expectations. It explicitly mentioned that orders for its silicon photonics device reliability testing and fully automated wafer-level burn-in screening equipment continue to surge. Silicon photonics business accounts for 20% of its total revenue, while equipment related to AI computing chips accounts for approximately 70%. Furthermore, silicon photonics testing is gradually advancing from early validation to large-scale deployment.
Sub-sectors: EML and CW Lasers Reach an Inflection Point
Furthermore, according to the InP (Indium Phosphide) laser chip supply model released by Citi, the global supply-demand gap for EML (Electro-absorption Modulated Laser) lasers from 2026 to 2028 will be 33%, 20%, and 11%, respectively. It is evident that industry analysts have concluded there will be a continuous shortage over the next three years, although the gap will gradually narrow. Conversely, the supply-demand gap for CW lasers will progressively widen from 12% in 2026 to 55% in 2028. The supply of InP laser chips across the industry is expected to remain in a state of shortage over the next three years. On the supply side, it will grow from approximately 280 million units in 2025 to about 1.3 billion units in 2028, representing a three-year CAGR (Compound Annual Growth Rate) of approximately 66%.
According to the Citi report, AXT and Sumitomo Electric, at the upstream wafer level for EML and CW laser chips, hold a combined market share of 50%. Their capacity is expanding rapidly and is expected to increase by 3.5 times by 2028. Moreover, the process is gradually transitioning from the current 3-inch to 6-inch wafers. With the yield rate and die size remaining basically unchanged per wafer, the chip output will increase by more than four times. Even so, there remains a huge gap between the supply of InP laser chips and market demand; even a more than fourfold increase cannot keep up with the demand. This is particularly true for high-power CW lasers. As NPO and CPO volumes ramp up, the demand for high-power CW lasers is growing rapidly. The higher the power, the higher the requirements for temperature stability, manufacturing difficulty, and yield rate, resulting in fewer manufacturers capable of stable mass production. Domestic company Yuanjie Technology (previous article: "[Co-Packaged Optics] The Title of the Most Expensive Stock Changes Hands Repeatedly: Is Yuanjie Technology Hype or the Future?") is also a standout in this track, with its products covering 2.5G to 50G InP laser chips and high-power CW light source chips for silicon photonics.
In addition to Sumitomo Electric and AXT, Japan's JX Metals recently announced it will increase investment, committing up to 120 billion yen to expand InP substrate capacity to approximately 7 to 10 times its current level. It will gradually build new factories, with a significant volume ramp-up expected by the end of 2027. Earlier, U.S. companies Coherent and Lumentum announced they will each invest 2 billion dollars to build new 6-inch InP wafer capacity. As the only global supplier capable of mass-producing 200G/lane EMLs, Lumentum has also secured the core materials for 1.6T modules. According to analysts, 200G EMLs will see double-digit price increases in 2026, and capacity will remain tight until 2027.
Conclusion
SemiAnalysis's interpretation delays CPO to 2028, which the entire market interpreted as bearish, prompting a rapid reaction in stock prices. However, in reality, the expectations have not vanished; they have merely been postponed. The demand is right there, and newly built capacity remains the strongest note. The vigorous expansion across the entire supply chain, including the materials and testing equipment sectors, serves as the best testament to this.
Images inserted in the article are generated by generative AI
References:
"CPO Volume Ramp-up Delayed to 2028, But This Is Not Bearish: What Is Extended Is the Shelf Life of Pluggable Optical Modules" - STT
"Analysis of Yield Bottlenecks, Mass Production Delays, and Supply Chain Ecosystem Layout in TSMC's CoWoS Platform CPO Packaging Business"
"Review of Domestic EML/CW Lasers: Global Shortage Exceeds 30%, Who Is Mass Producing and Who Is Defaulting?"