Reasoning Replacement Cycle Begins, Focusing on the Chinese Nationalization Wave

Morgan Stanley released a special in-depth report on China's AI accelerator industry, clearly identifying the two core forces of AI inference demand explosion and export controls, which will drive China's AI chip market to reach $67 billion by 2030. The localization rate will rise from 33% in 2024 to 86%, and the domestic chip market size will surpass imported chips by 2027. The report covers three major local manufacturers for the first time, using a dual framework of "inference economy × execution capability" to deconstruct the logic of industry winners, giving Cambricon and TianShu Zhixin an overweight rating, and Muxix a neutral rating.

Commercialization landing, localization enters a long cycle

China's AI GPU market has shifted from policy-driven to commercial landing, and the core contradiction of the market has changed from "whether domestic chips are usable" to "which manufacturers can take substantial shares in the large-scale deployment of inference". Two structural forces are reshaping the industry landscape: First, large models are shifting from training to large-scale deployment, and inference computing power has become the core of demand. The daily token usage of domestic mainstream large models has exceeded 100 trillion, and the hardware selection standards have shifted from peak computing power to single-token costs, utilization rates, and cost efficiency; Second, the ongoing export controls make the self-sufficiency of the supply chain a must, and NVIDIA's high-end chip sales to China are restricted, with advanced process foundry facing strict constraints. Localization has shifted from a short-term policy response to a long-term industry trend.

Recent industry data further confirm the acceleration of substitution: The supply of NVIDIA GPUs in China continues to tighten, and incremental demand shifts to domestic manufacturers; the spot prices of domestic RTX 5090, token pricing, and GPU rental prices continue to rise, confirming the strong demand for downstream inference. The only negative signal is that some manufacturers have started price reductions to seize market share, and the industry's price competition has arrived earlier than expected.

System-level capability becomes the core watershed

The report's core judgment points out that the competition for China's AI chips has bid farewell to a single-chip parameter competition. Although local products are still about two generations behind the United States in chip level, through multi-chip architecture, advanced packaging, rack-level system design, optical interconnection, and software and hardware co-optimization, the effective performance gap is narrowing rapidly.

In the market dominated by inference, the optimal token benefits under the acceptable software migration cost are the core to winning orders. Local manufacturers have formed a clear competitive advantage: in terms of performance, top products such as Huawei Ascend 950PR and Cambricon MLU690 have 50%-150% higher TPS performance than NVIDIA H20; in terms of cost, the total lifecycle TCO of local chips is 30%-60% lower than NVIDIA's products available in China, and the single-token cost of top products has been on par with NVIDIA A100/H20, with some scenarios achieving a lead; in terms of energy efficiency, domestic chips have basically caught up with NVIDIA A100 and H20, with a significant performance advantage per dollar.

Based on this, the report builds a two-dimensional evaluation framework of "inference economy × execution capability". The former includes quantifiable indicators such as TCO, single-token cost, and TPS performance, while the latter includes qualitative indicators such as the accessibility of advanced process capacity, the maturity of the software ecosystem, and the depth of cooperation with cloud manufacturers. Only manufacturers with advantages in both dimensions can stand firm in the industry integration.

Core target ratings and investment logic

The report gives different ratings to three core manufacturers based on the evaluation framework:

1. Cambricon (688256.SH): Overweight rating, target price of 1588 yuan. As the domestic leader in AI inference chips, its core advantage lies in deep binding with top cloud manufacturers and mature software and hardware co-capabilities. The MLU series of chips has been deployed on a large scale in companies such as ByteDance, and the new generation of MLU690 is expected to be mass-produced in the fourth quarter of 2026. The report expects the company's revenue compound annual growth rate from 2025 to 2028 to be 90%, with revenue of 20.944 billion yuan in 2026. As the first domestic AI chip company to achieve full-year profitability, its business model has been verified. The core risk is the constraint of advanced process yield and high customer concentration.