Sinolink: The investment in computing power by major companies will enter a white-hot stage, and the prospects for domestic alternatives in the artificial intelligence sector are expected to accelerate.

Sinolink's research report states that behind the AI entrance battle is a battle for traffic, directly targeting the advertising and e-commerce business foundations of large internet companies. By 2026, the investment in computing power for models and AI application products by various companies will enter a heated stage. The expansion of the intelligent computing centers continues, with the possibility of accelerating domestic substitution. The scale of China's intelligent computing power is expected to maintain a compound growth rate of 57% from 2020 to 2028. In terms of supply and demand, downstream inference demand, along with the entry of AI phones and other terminal AI devices, is approaching the critical point of certainty enhancement; the Big Model Scaling Law is still effective, and the continuous iteration of models accelerates the increase in training demand with certainty. Sinolink's main points are as follows: The battle for AI entrances sees increased investment by major companies Continuous iteration of large models has effectively changed the competition landscape dominated by North America. Within the global Top 10 camp, GLM-4.7, DeepSeek V3.2, and Kimi K2 Thinking have already secured three spots. Inference demand has surged: the emergence of type o1 inference models has unlocked about 10 times the potential computational power in the inference stage compared to traditional models, with the demand for computing power shifting from a single "training-driven" approach to a "training+inference dual-drive" approach. Reshaping traffic entry: On January 15th, the Qianwen App announced full integration with Alibaba's ecosystem businesses including Taobao, Alipay, Taobao Flash Buy, Fliggy, and Amap, enabling AI shopping functions such as ordering food, shopping, and booking flights, leading the AI industry from "chat dialogue" into the "business era." On January 16th, OpenAI announced the launch of a $8 monthly subscription "ChatGPTGo" package and plans to test advertisements among free users and Go users in the United States in the coming weeks. Sinolink believes that the battle for entrances is ultimately a battle for traffic, directly affecting the advertising and e-commerce business foundations of large internet companies. By 2026, the investment in computing power for models and AI application products by various companies will enter a heated stage. Continued breakthroughs in domestic cards, leading to an outbreak in 2026 The expansion of intelligent computing centers continues, accelerating domestic substitution: It is projected that the scale of China's intelligent computing power will maintain a 57% compound growth rate from 2020 to 2028. Domestic general-purpose GPUs are transitioning from "available" to "useful": Domestic computing chips have continued to make breakthroughs in process and architecture, narrowing the performance gap with international giants in handling large models, long texts, complex operator fusion, and other areas. Local wafer fabs, represented by Semiconductor Manufacturing International Corporation, have been operating at high capacity/utilization rates, providing a solid guarantee for domestic AI chip production capacity. CSP vendors are accelerating adaptation, supporting domestic chip ecosystem development: Tencent Cloud announced full adaptation to mainstream domestic chips, while companies like Baidu and Alibaba are ramping up adaptation to domestic chips, promoting the formation of a "chip-model-application" closed loop. High demand and supply will create a steep slope for domestic computing power Demand side: Inference demand downstream, along with the widespread application of AI phones and other terminal AI devices, is approaching the critical point of certainty enhancement; the Big Model Scaling Law remains valid, and the continuous iteration of models accelerates the increase in training demand with certainty. Policy-driven demand surge: In January, the Ministry of Industry and Information Technology introduced restrictions on the PUE for new intelligent computing centers, promoting hardware computing efficiency and energy efficiency upgrades, and for the first time, including the "National Integrated Computing Power Network" in the national-level infrastructure system in the "Fifteenth Five-Year Plan." Supply side: Domestic GPU performance continues to improve; domestic CSP vendors are accelerating adaptation to domestic GPUs. Overseas cloud computing power control is expected to push domestic AI companies to accelerate the domestication substitution process. Risk warning Risks of intensified industry competition; risks of slower-than-expected technological research and development progress; risks of periodic fluctuations in downstream capital expenditures in specific industries.