Central China: Domestic AI chip adoption rate rises to 64%; Kimi returns to the ranks of top-tier models

date
10:32 30/10/2026
avatar
GMT Eight
If the A-share AI computing power sector is to stage an independent rally, it may still face significant challenges in terms of market confidence.
Central China issued a research report, assigning an outperform investment rating to the industry. Currently, AI applications are accelerating their penetration into fields such as programming, film/television and gaming, mathematical research, and cybersecurity, while the pace of breakthroughs in physical AI is also accelerating. The autonomous execution of Agents means the frequency of model invocations is also growing substantially, keeping computing power supply persistently insufficient and driving rental prices for compute cards continued upward. Relatively speaking, constrained by the chip shortage factor, domestic computing power tightness is more severe. With the release of domestic chip capacity in 2027 and the launch of new cards, domestic computing power demand is expected to be better met. We continue to hold an optimistic outlook on the development of China's AI computing power industry, but given that overseas tech stocks are generally at high levels and may subsequently be affected by factors such as Anthropic's IPO and concerns over 2027 capital expenditure, the A-share AI computing power sector may still face significant confidence challenges if it is to stage an independent rally. Central China's main views are as follows: Domestic substitution In the first nine months of 2026, the unit prices of integrated circuit imports and exports continued to rise, while export unit prices surpassed import unit prices for four consecutive months, reflecting continuous improvement in the performance of domestic chip products. In the first half of 2026, the domestic substitution rate of AI chips further rose to 64%, and in 2026Q3 manufacturers' performance continued its high-growth momentum. DeepSeek and Ascend are becoming increasingly closely integrated, accelerating the domestic chip substitution process through open-source approaches. Alibaba's new-generation self-developed chip V900 will launch in 2027Q1, two quarters earlier than originally planned, and the Ascend 960, also slated for launch in 2027, is similarly planned to be moved up. Thanks to logic folding, at the same performance level, the Kirin 2026 saw NPU power consumption drop 66%, GPU drop 58%, and performance-core CPU drop 41% in benchmark tests. As of October 21, 2026, the number of HarmonyOS operating system terminals has exceeded 95 million, and Huawei expects to surpass 100 million users in 2026. AI Moonshot AI released Kimi K3.1 Pro in October, featuring 2.8 trillion parameters. Through enhanced post-training and reinforcement learning with data from more domains, it has returned to the ranks of top-tier models, while other domestic models mainly occupy the second tier. Based on the invocation pricing of Anthropic's Opus 5.5, the strongest U.S. model, its input and output prices are both approximately twice those of the comparable Kimi-K3.1 pro, and amid the impact of domestic models and rising user cost sensitivity, U.S. closed-source models have recently shown a clear trend of price cuts. According to OpenRouter's model invocation data, DeepSeek holds a 26% share and has ranked as the largest model vendor for 12 consecutive weeks. Meanwhile, Chinese model vendors' share reached 46%, temporarily leading U.S. model vendors (44% share). In October, Microsoft offered voluntary departure packages to approximately 60,000 employees (over 25% of its workforce) to cut costs and reallocate resources toward AI, as layoffs at major tech companies show signs of accelerating again. In September, U.S. enterprise model adoption rate reached 55.9%, with Anthropic once again surpassing OpenAI (40.6%) with a 43.8% share. Risk warnings: market shocks from IPOs of domestic and international tech giants; uncertainty in the international situation; market adjustment risks arising from changes in the competitive landscape of the overseas AI industry.