China's strategic posture ('hide strength') increasingly embraces acquisition and use of advanced U.S. AI chips despite export restrictions.

Subtle shifts are emerging in the AI technology competition between the United States and China, marked by signs that China is approving imports of NVIDIA's new AI graphics processing units (GPUs) after previously blocking them.

According to *The Information* on September 27th, China's Ministry of Industry and Information Technology recently requested companies such as Alibaba and ByteDance to report their willingness to purchase NVIDIA's RTX Pro 5500 and its intended applications. Some companies were informed of the ministry's intent to approve the purchase. While not a top-tier AI data center chip, the RTX Pro 5500 is capable of driving and training AI models.

Initial interpretations suggested that China was easing U.S.-China tensions stemming from AI competition, using the U.S.-China summit as an opportunity. China had previously restricted its companies from purchasing U.S. chips in response to American export bans on advanced technologies, including the latest AI processors.

However, analysts increasingly characterize this move as a form of "hiding capabilities and biding time" and "learning the enemy's strengths"—a pragmatic strategy to fill technological gaps using American technology, then ultimately surpass the U.S. based on that foundation.

To slow China's AI advancement, the U.S. has banned exports of NVIDIA's top-tier chips, such as the Blackwell series and above. While limited sales of lower-performing chips like the H20 and H200 were permitted, Chinese companies were required to pay 15–25% of revenue from these sales to the U.S. government.

China has responded by intensively fostering domestic companies and mandating the use of locally produced AI chips. According to Bernstein, NVIDIA's market share in the Chinese AI chip sector is expected to drop from 40% last year to 8% this year, while Huawei's share is projected to rise from 40% to 50%. Market research firm TrendForce anticipates that China's domestic production ratio of high-performance AI chips will approach 90% this year.

Chinese companies are now releasing high-performance AI models without relying on NVIDIA chips, keeping pace with the U.S. through a "volume strategy" that connects hundreds to thousands of chips to compensate for individual performance shortcomings. Huawei plans to unveil the Ascend 960DT in the first quarter of next year and continue releasing new AI chips annually. DeepSeek CEO Liang Wenfeng recently identified expanding use of Chinese-made chips for AI training as a core task, with forecasts that Huawei could begin supplying training chips to DeepSeek as early as the fourth quarter of this year.

China is gaining traction in AI models through price competitiveness. According to Artificial Analysis, Xiaomi's MIMO V2.6 Pro achieved an intelligence score of 46, matching xAI's Grok 4.7, yet cost 0.18 dollars per 1 million tokens—approximately one-seventh of Grok's 1.35 dollars. As of September 28th, DeepSeek's share of total text requests on OpenRouter was 22.6%, surpassing Google (18.6%) and OpenAI (15.8%).

Confronted with China's push driven by affordable models and domestic chips, U.S. AI companies are maintaining their lead by continuously releasing ultra-high-performance models. Although discussions on moderating AI development pace are emerging within the U.S., major companies are relentlessly improving top-tier performance. In cutting-edge closed models, NVIDIA GPUs, and the supporting CUDA ecosystem, the U.S. is widely regarded as holding the advantage.

To preserve this edge, the U.S. is tightening export control loopholes. Congress is discussing legislation to track actual usage of exported AI chips overseas and block pathways through which China accesses advanced chips via third countries. A high-ranking source in the AI industry stated: "As both countries' competitive strategies in the AI technology race become more sophisticated, U.S.-China technological conflicts could intensify in the mid- to long-term."