Guoyuan: The current industrial scale of CPO technology is still relatively small, with vast potential for future penetration.

Guoyuan released a research report stating that although CPO technology currently has a small industry scale, its performance advantages are significant. In the long run, it may become the ultimate structure for data center optoelectronic conversion modules, with vast potential for penetration in the future. Currently, domestic suppliers participating in the CPO chain are still mainly concentrated in the upstream of the industry chain. In order to hedge against the current complex international situation, they have carried out extensive global capacity layout. Guoyuan is optimistic about Suzhou TFC Optical Communication (300394.SZ) and T&S Communications (300570.SZ) which have excellent customer resources. In addition, Yuanjie Semiconductor Technology (688498.SH) has an early advantage in the layout of CPO's CW light source and light source modules. Guoyuan's main points are as follows: The scaling law extends from parameters and computing power to parallel flows, and the interconnection bandwidth will also be driven. With the scaling law extending from parameter scaling in the pre-learning stage to inference time scaling in the post-learning stage, model performance significantly improves and application penetration accelerates. In order to achieve better model performance with as few additional parameters and hardware load as possible, parallel scaling has entered the academic field. According to the research team data of Qwen, opening 8 parallel flows is equivalent to tripling the number of parameters, but the actual increase in hardware costs is minimal. The parallel process during the training phase significantly increases the demand for "large-scale synchronization", requiring communication between different processors or nodes to synchronize model parameters or gradients and driving the demand for interconnection bandwidth within the cluster. Under the demand for cluster low power consumption and high speed, higher integration may lead to better solutions. On the one hand, according to data center TCO data, regardless of rack rate and external power load, electricity costs always maintain the highest percentage in costs. On the other hand, the penetration of parallel scaling in the model deployment process will also drive the demand for interconnection bandwidth within the cluster. Under the demand for low power consumption and high speed in the cluster, CPO (Co-Packaged Optics) technology has received attention. In terms of power consumption, the power consumption of CPO architecture has been reduced to 5-10pJ/bit (traditional pluggable solutions 15-20pJ/bit), Broadcom's CPO switch achieves 5.5W/800G, and Nvidia's solution reduces power consumption by 70%. In terms of speed, the signal transmission distance is shortened to millimeters, bandwidth density is increased, and latency is reduced by over 50%. According to LightCounting's prediction, due to the performance advantages of CPO, by 2027, CPO ports will account for nearly 30% of the total number of 800G and 1.6T ports. Risk warning: Risk of international trade friction, industry demand fluctuations, exchange rate fluctuations, market competition risks.