Nvidia was "stuck" with half a million chips. Then the solution was found

The global AI chip market is experiencing rapid growth, with China striving to strike a balance between fostering technological self-reliance and securing advanced computing resources. Recently, Chinese authorities granted approval for limited imports of H200 chips from American chipmaker Nvidia, according to a report in the Financial Times. This move aims to alleviate bottlenecks faced by leading Chinese technology firms competing against American AI powerhouses.

ByteDance, the company behind TikTok, and Tencent each obtained around 10,000 H200 processors under this new regulatory approval. More Chinese tech companies are anticipated to receive similar permits for shipments of this scale soon. However, these chips lag at least two generations behind Nvidia's most powerful processors, which are prohibited for purchase by Chinese customers due to stringent US export controls.

Chinese authorities imposed restrictions on chip imports, favoring the development of local semiconductor manufacturing. As a result, Nvidia currently possesses an inventory of approximately half a million H200 chips, originally intended for the Chinese market. Some of Nvidia's commercial partners, such as computing giant Lenovo, have informed Chinese customers that orders for products incorporating H200 processors can be reinstated.

Nonetheless, actual purchases would still necessitate explicit approval from China's National Development and Reform Commission.

The Chinese government's cautious approach toward chip imports reflects its strategy to bolster domestic industry, particularly led by Huawei, which aims to become a dominant force in the Chinese semiconductor market. Despite this, China still grapples with a shortage of advanced chip manufacturing equipment, constraining its ability to meet escalating domestic demand. Consequently, importing Nvidia chips remains crucial, particularly for the intricate training phases of AI models.

While Chinese research institutions are increasingly leveraging locally produced chips for inference tasks - generating real-time responses - most organizations still depend on Nvidia's processors for the far more resource-intensive training stage involved in building knowledge infrastructure and recognizing patterns within AI models.