[Investment strategy: The AI chip market is forming a pattern of parallel development of GPUs and custom ASICs] On September 8, the investment strategy stated that competition for overseas large models has heated up significantly again this week: OpenAI released GPT-6 Astra on September 4, Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 on September 2, and Google released Gemini 3.8 on the same day. Flash, three leading manufacturers completed a new round of flagship or main model updates in a very short period of time. Judging from the common trends among the three models, actual production efficiency has become an important criterion for measuring the value of cutting-edge models, and scientific research capabilities are one of the most focused subdivisions of the flagship model. At the same time, the product lines of different manufacturers have further differentiated: OpenAI strengthens computer operation capabilities, Anthropic continues to develop agents, and Google relies on the Flash series to strengthen the balance between performance, speed and cost. ASIC’s massive volume is reshaping the computing power landscape. Broadcom's latest financial report further confirms that AI computing power is accelerating the reconstruction, and customized chips have become an important part of the expansion of AI computing power. In FY2026 Q3, the company's revenue hit a record high, and it gave AI semiconductor revenue guidance for the next two fiscal years. Customized ASICs have entered the large-scale deployment stage, and have gradually become an important choice for leading model manufacturers and cloud vendors to optimize inference costs, energy consumption, and system efficiency. As customers such as Google, Anthropic, OpenAI, and Meta advance multi-generation XPU projects, the AI ​​chip market is forming a pattern of parallel development of GPUs and custom ASICs.