EPA=Yonhap News

China has reportedly instructed ByteDance, the parent company of 'TikTok,' not to use NVIDIA chips.

However, China is said to be allowing the use of NVIDIA chips for training AI models.

Chinese regulators have blocked ByteDance from using NVIDIA chips in its new data centers, according to a report by The Information, an American technology news outlet, citing sources on the 26th (local time).

This directive applies not only to new orders of NVIDIA chips but also to chips the company already possesses.

This represents an intensified measure compared to China's previous instruction to local companies to refrain from new orders of NVIDIA's AI chips.

ByteDance is known to have been the largest purchaser of NVIDIA chips among Chinese companies this year.

This is interpreted as part of a strategy to foster China's indigenous technology, independent of the U.S. AI ecosystem.

Ultimately, the directive is to use products manufactured by domestic companies like Huawei and Cambricon instead of NVIDIA chips.

Brady Wang, an analyst at Counterpoint Research, explained that Chinese companies are moving quickly to reduce their reliance on NVIDIA chips, stating, "This is not something they want to do, but something they have to do."

However, sources indicate that China has only prohibited the use of NVIDIA chips for running and inference tasks of AI models, not completely blocking their purchase for model training.

This is because while the technical capabilities of Chinese-made chips are sufficient for performing AI model tasks, they are still inadequate for training AI models that require absorbing vast amounts of data and understanding the relationships within it.

China also lacks sufficient production capacity for AI chips. Chinese companies like Huawei and Alibaba used to produce chips through TSMC, the world's leading semiconductor foundry in Taiwan, but this has become impossible since the U.S. strengthened its controls.

The U.S. currently only permits the sale of NVIDIA's lower-performance version, the 'H20,' to China, but this product is considered insufficient for training and only useful for running and inference tasks.

However, there is potential for a turnaround if the 'H200,' which the U.S. is currently considering whether to permit for export to China, is actually exported.

The H200 boasts approximately twice the performance of the H20 and is also suitable for model training.

Yonhap News