Citic Securities has released a research report maintaining an "Outperform" rating on the communications sector. As AI large language models evolve towards Mixture-of-Experts architectures, the extreme demands of computing clusters on Scale-Up networks make NPO technology a key solution for all-optical interconnectivity. Currently, tech giants like Alibaba and Tencent are accelerating their investments in this direction. This technological evolution is profoundly reshaping the industrial chain, granting early adopters significant technological premiums. Driven by the sustained high growth of the AI industry and the active push by major players to implement NPO, leading domestic optical communication companies are expected to leverage their deep expertise in high-speed interconnectivity to gain a first-mover advantage, achieving simultaneous growth in value and profitability.

Citic Securities' main views are as follows: AI large language models are driving a Scale-Up network transformation, where the balanced advantages of NPO technology are becoming prominent. Amid the current wave of generative AI, large model architectures are gradually shifting from Dense to MoE, placing unprecedented extreme demands on Scale-Up interconnect networks regarding bandwidth density and communication latency. As underlying SerDes technology advances towards 224G and even 448G, signal attenuation from traditional copper wires or long-distance PCB traces is approaching physical limits. Compared to high-power traditional pluggable optical modules and the challenging CPO approach with its high maintenance difficulty and manufacturing barriers, NPO demonstrates strong "greatest common divisor" characteristics. By moving the optical engine closer to the switch chip on the same substrate, NPO significantly reduces electrical signal transmission distance and power consumption while maintaining the independent replaceability of the optical engine, balancing signal integrity and maintainability.

Tech giants are accelerating their involvement, with Alibaba's UPN512 architecture leading the way in all-optical interconnect implementation. Leading CSP manufacturers are actively promoting the adoption of NPO-related solutions. Alibaba has officially released its "UPN512 Technical Architecture White Paper," aiming to build a fully interconnected system of 512 xPUs using a single-layer Ethernet optical interconnect, which is expected to reduce optical interconnect costs by over 30%. As the core of this architecture, Alibaba has successfully powered up a 3.2T NPO module. This module supports silicon photonics and VCSEL technology, with a typical TDECQ of only 1.9dB and power consumption of just 20W. It has been率先 applied in a new generation of domestic four-chip switches. Furthermore, Tencent is actively exploring NPO evolution based on silicon photonics technology and, together with Alibaba, has initiated a standardization project within the ODCC. They anticipate related prototypes will achieve system-level power-on by Q3 2026, marking a new phase of engineering implementation for all-optical interconnect.

Domestic optical module manufacturers are transitioning towards becoming "optical interconnect solution providers," facing a historic opportunity. The accelerated布局 by internet giants like Alibaba and Tencent signals that NPO technology is entering the stage of large-scale commercial use. This will profoundly reshape the business models and supply chain structure of the optical communications industry. NPO requires optical module manufacturers to possess strong silicon photonics integration capabilities, high-precision packaging processes, and deep collaboration with chip manufacturers, pushing them to transition from simply "providing modules" to offering higher-barrier "optical interconnect solutions." Leading domestic companies are seizing this historic opportunity: Zhongji Innolight has demonstrated an OpenSocket NPO solution expected to see large-scale deployment by 2027; Eoptolink has introduced a VCSEL-based NPO optical engine to improve spatial utilization; and HG Tech's 3.2T NPO optical engine is also anticipated to achieve large-scale commercial application by 2026.

Risk factors include: AI investment falling short of expectations; intensifying market competition; slower-than-expected progress of new products like 1.6T; technology roadmap risks; geopolitical risks; and excessively rapid capacity expansion.