Orient Securities released a research report stating that domestic AI models are set to complete a generational leap by 2026 through trillion-parameter MoE architecture, million-token context windows, and full-stack adaptation to domestic chips. Supply shortages are already reflected in rising prices, including for GPU rentals, DRAM and NAND contract prices, and cloud services. Significant potential exists for import substitution in the optical communication testing sector. As production lines for 1.6T silicon photonics expand, the process of substituting imported testing equipment is expected to accelerate. The firm recommends focusing on the core segments of domestic computing power.

Orient Securities' key views are as follows:

Domestic large language models are continuously iterating, with architectural efficiency advantages and the establishment of a commercial closed loop driving the transition. From the 2023-2024 period featuring the approval and scaling of hundred-billion-parameter general models, to 2025 where multiple evaluation metrics rapidly narrowed the gap with overseas counterparts, domestic models are projected to achieve a generational leap by 2026 via trillion-parameter MoE architecture, million-token context, and full-stack compatibility with domestic chips. This will accelerate the adoption of domestic models in areas like agent autonomous execution, long-context understanding, and complex reasoning, altering the global landscape of model call volumes. According to OpenRouter data, by mid-May 2026, the weekly call volume for Chinese large models reached 7.94 trillion tokens, 2.11 times that of the US during the same period. Commercially, Zhipu's API achieved an annualized revenue of $2.5 billion, with call volume still growing over 400% year-on-year after three price hikes. Minimax's ARR is approximately double its full-year 2025 revenue, and Anthropic's ARR in April was around $30 billion. The formation of this commercial closed loop for large models continues to unleash demand for computing power.

Computing power supply is tight, accelerating domestic chip substitution, with super-nodes reshaping the computing power landscape. Supply constraints are evident in price increases for GPU rentals, DRAM and NAND contract prices, and cloud services. Regarding domestic chips, rapid progress is being made due to combined factors including architectural innovation in design, advancements in foundry and packaging, and software stack coordination. Simultaneously, multiple manufacturers have completed adaptation for models like DeepSeek-V4. On the financial front, Cambricon's Q1 2026 revenue reached 2.885 billion yuan, with significant increases in prepayments and contract liabilities, confirming high activity on both the supply and demand sides. Concurrently, super-nodes are helping domestic computing power break through bottlenecks, with 2026 marking the first year of large-scale deployment.

The prosperity in the transmission layer is also commencing, with the import substitution for optical module testing equipment entering a critical window. The expansion of computing power clusters is driving demand for optical modules. Leading overseas cloud providers continue to raise their capital expenditure guidance, and major optical module manufacturers have clear intentions to expand production, driving procurement demand for upstream manufacturing equipment. The testing phase represents the highest barrier and the lowest domestic substitution rate in optical module manufacturing. As optical module speeds evolve from 800G to 1.6T, testing complexity increases, the number of test items grows, and the value per machine rises significantly. According to Frost & Sullivan data, in 2024, local companies held only a 16% share of the Chinese optical communication test instrument market, while Keysight, Anritsu, and Viavi collectively held over 60%, indicating substantial room for import substitution. As 1.6T silicon photonics production lines expand, the process of substituting imported testing equipment is expected to accelerate.

Risk warnings include AI development falling short of expectations, supply chain development falling short of expectations, and the pace of import substitution for optical module equipment falling short of expectations.