Sinolink: Large-scale AI cluster drives accelerated development of CPO, optimistic about industry chain companies.

Sinolink has released a research report stating that the scale of AI clusters is expanding, the signal transmission rate is increasing, and CPO has obvious advantages in power consumption, cost, and low loss, with clear demand. The supply chain for CPO is gradually maturing. CPO is expected to see accelerated penetration, with core CPO supply chain companies expected to benefit fully. It is recommended to pay attention to core CPO optical component companies, CPO manufacturers, and CPO solution companies. Sinolink's main points are as follows: The development of AI is driving the expansion of training cluster size, with token consumption of inference side continuously increasing, driving demand for computing power and network. The bank believes that CPO can effectively help large-scale clusters reduce power consumption, increase interconnection density, and improve transmission stability at high speeds. The current demand for CPO due to the development of AI is urgent, and the supply side of CPO is also gradually maturing, leading to the potential for accelerated penetration of CPO. 1) Demand side: The expansion of AI clusters is driving the increase in network size, density, and speed, catalyzing the demand for CPO. According to Semianalysis calculations, in large-scale networking, the power savings of CPO are significant. In a 3-layer network cluster, even though the use of CPO in switch chips has increased power consumption due to the integration of optical components such as light engines, the power consumption generated by optical modules has significantly decreased, leading to a 23% reduction in total network power consumption. For a GB300 NVL72 cabinet in a 3-layer network, using a CPO solution can reduce network costs by 21% compared to the traditional DSP optical module solution, leading to a 3% reduction in total costs. If the network is compressed to a two-layer CPO solution, network costs can be reduced by 46% and total costs can be lowered by 7%. Additionally, in high-speed signal transmission, a CPO solution can reduce transmission distance and minimize signal loss compared to traditional solutions. 2) Supply side: The bank believes that the supply of CPO is also gradually maturing, with major switch chip manufacturers such as Broadcom and NVIDIA introducing CPO switch chips. The stability of CPO has also reached a high level. The bank believes that the current supply side of CPO can meet a large shipment volume, leading to an acceleration in the increase of penetration rate. The bank believes that core CPO supply chain companies are expected to continue benefiting from the trend of accelerated penetration of CPO. The bank categorizes core CPO supply chain companies into: 1) core optical component companies; 2) CPO manufacturers; 3) CPO solutions. Core CPO optical components include lasers, light engines, FAU, Shuffle Box, MPO, etc. CPO light sources mainly use CW lasers, benefiting from the continuous growth in optical communication demand. Major laser manufacturers like Lumentum, Coherent, and Sumitomo are expanding production. Lumentum and Coherent are expected to continue benefiting. Domestic manufacturers such as Yuanjie Semiconductor Technology are making progress, and Suzhou Dongshan Precision Manufacturing also has strong optical chip production capabilities. Currently, lasers use indium phosphide substrates, with major suppliers being AXTI, etc. FAU is mainly used to input and output light signals into and out of light engines. Currently, Tianfu Technology has successfully completed mass production of 1.6T light engines and research and development of CPO matching optical components. CPO requires high-precision micro-lens arrays, and traditional optical companies like Zhejiang Lante Optics are expected to enter CPO. MPO is used to connect optical fibers to external ports, connect switches to other switches or remote NICs, and T&S Communications is an MPO connector supplier indirectly supplying end customers through Corning. Shuffle Box is designed specifically for managing complex fiber interconnections. Manufacturers in the CPO manufacturing sector that stand to benefit include wafer manufacturing, outsourced packaging and testing, and CPO testing companies. In terms of wafer manufacturing, the EIC in CPO light engines requires advanced processes, and TSMC has a significant advantage in advanced processes. In the connection of PIC and EIC in light engines, NVIDIA and Broadcom have started using TSMC's COUPE technology. The bank believes that TSMC's COUPE technology is expected to solidify TSMC's position in the silicon photonics generation industry. Traditional packaging and testing companies such as ASE and Amkor are expanding into light engine packaging testing, and Suzhou TFC Optical Communication is actively involved in advanced packaging and is expected to benefit. After CPO light engines are integrated into switches, defects will lead to the scrapping of the entire module, increasing the importance and complexity of CPO testing. The bank believes that CPO testing equipment is expected to benefit significantly, including manufacturers with high complexity in optical alignment and those benefiting from more optical channels and bandwidth. Platform-based testing companies are expected to benefit from the industry beta of increased chip testing complexity brought by CPO. CPO solution side NVIDIA uses a more aggressive MRM modulator with advantages of low power consumption and small footprint, while Broadcom uses MZM modulators, which are more mature in manufacturing. Future modulators are expected to shift to thin-film lithium niobate solutions. The bank believes that NVIDIA's solution is expected to generate competitive advantages against Broadcom, driving NVIDIA's network revenue growth. The bank believes that the switch industry is expected to benefit from increased revenue due to higher integration of CPO. Risk warning CPO technology development falls short of expectations; AI model iteration stagnates; US macroeconomic pressure; AI hardware development and yield fall short of expectations; deterioration of US-China technology policy.