Northeast: NPO as the new main line for AI optical interconnection, balancing performance benefits and the core path of engineering feasibility.

Northeast published a research report stating that NPO is not merely a transitional solution from CPO, but rather a core pathway that balances performance benefits and engineering feasibility at the current stage of the industry. As a new battleground for domestic computing power, leading cloud providers are accelerating the commercialization process. Domestic cloud providers, equipment manufacturers, chip manufacturers, and optical communication companies have jointly participated in the formulation of standards like OPENNPO, positioning Chinese enterprises as potential co-definers of the new optical interconnect architecture. NPO is a strategic move in the field of optical interconnect, with both domestic and overseas cloud providers embracing it, representing yet another significant opportunity for China's industrial chain, with continuous recommendations for related targets. The main viewpoints of Northeast are as follows: NPO balances performance, yield, and maintainability, making it the most feasible next-generation optical interconnect solution at present. As SerDes evolves from 112Gb/s to 224Gb/s, the long electrical channels of traditional pluggable architectures face constraints in loss, power consumption, and heat dissipation. The location of optoelectronic conversion continues to migrate closer to the switching chips and computing power chips. CPO integrates ASIC and the optical engine within the same package, achieving higher bandwidth density and energy efficiency limits, but poses high demands on advanced packaging, joint yield, thermal management, and system maintenance. NPO, however, deploys independently packaged optical engines near the ASIC, trading a smaller electrical performance gap for higher manufacturing yield and maintainability. Therefore, NPO is not simply a transitional solution from CPO, but a core pathway that balances performance benefits and engineering feasibility at the current stage of the industry. AI clusters are continually expanding, with Scale-up becoming the largest new application scenario for optical interconnect. Long context, multimodal, MoE, and complex reasoning continuously amplify the need for synchronous communication between chips. The competitive focus of computing power systems has shifted from peak performance of a single card to effective computing power of the cluster. Super nodes have evolved from dozens of cards to hundreds or even thousands of cards, and the Scale-up interconnection range has extended from within the machine and the rack to cross-rack connections, with its single XPU interconnection bandwidth being approximately 9-10 times that of Scale-out. Compared to Scale-out networks that have widely adopted optical modules, Scale-up still relies heavily on PCBs, copper cables, and high-speed connectors, resulting in a low penetration rate for optical interconnect. As high-speed copper interconnect comes closer to the limits of distance, power consumption, and wiring density, "optical in and copper out" will extend from between cabinets to within cabinets, devices, and closer to the chips, further unlocking the upper market limit for optical interconnect. NPO has become a new battleground for domestic computing power, with leading cloud providers accelerating the commercialization process. Domestic computing power needs to release effective computing power by expanding tightly coupled computing domains and improving system interconnection efficiency. NPO does not require advanced packaging of ASIC and optical engines, making it more suited to the current technical foundation and mass production capabilities in China. Huawei is forming a collaborative layout around Taos Law with the Hi-ONE high-density optical engine, OPENNPO standards, and super node systems; Alibaba is promoting the trial deployment of 3.2T NPO and evolving towards 6.4T UPO; Tencent is simultaneously validating 3.2T silicon photonics and VCSEL solutions, expecting to deploy the NPO super node scheme by Q4 2026, with ongoing acceleration in commercialization. NPO can fully leverage the existing optical communication industrial chain in China, providing significant opportunities for Chinese enterprises to deeply engage in the next generation of optical interconnect. NPO retains independent optical engines, PIC, EIC, lasers, FAU, high-density connectors, boards with optical fibers, and advanced packaging processes. The domestic manufacturers' accumulated experience in traditional pluggable optical modules, silicon photonics design, precision coupling, passive devices, optical chips, and large-scale manufacturing can be directly transferred. Currently, the mechanical, electrical, optical, management, and testing interfaces of NPO are still under a unified phase. Domestic cloud providers, equipment manufacturers, chip factories, and optical communication companies have jointly participated in the formulation of standards like OPENNPO, and Chinese enterprises are expected to further transition into co-definers of a new optical interconnect architecture. Investment Recommendations NPO is a strategic move in the field of optical interconnect, with both domestic and overseas cloud providers simultaneously embracing it, representing yet another significant opportunity for China's industrial chain, with continuous recommendations for related targets. (1) Optical engines: Zhongji Innolight, Eoptolink Technology Inc., Accelink Technologies, Hgtech; (2) High-density optical connections: T&S Communications, EverProX Technologies, Henan Shijia Photons Technology, Shenzhen Zesum Technology; (3) FAU, semiconductor optics: Suzhou TFC Optical Communication, Advanced Fiber Resources, Shenzhen JPT Opto-Electronics, Focuslight Technologies Inc.; (4) Light sources and upstream materials: Yuanjie Semiconductor Technology, Suzhou Dongshan Precision Manufacturing, Suzhou Everbright Photonics, Yunnan Lincang Xinyuan Germanium Industry; (5) Advanced packaging: JCET Group Co., Ltd.; (6) Domestic computing power and switching chips: Moore Threads, Muxi Co., Shengke Communications, ZTE Corporation. Risk Warning: AI computing power capital expenditures may fall short of expectations; NPO product validation and scaled deployment may progress slower than anticipated; the progress of domestic computing power and super node construction may not meet expectations; and market competition may intensify.