DeepSeek delays launch of new AI model. Reason: problems with Huawei chips

DeepSeek has postponed the launch of its new R2 model due to technical problems with training on Huawei Ascend chips, Financial Times reports.

• According to the FT, during training, the R2 encountered constant problems with Ascend chips, forcing the company to use Nvidia for training the model and Ascend only for inference (generating responses).

• The report highlights the difficulties Chinese AI developers face in reducing their dependence on American technology, particularly Nvidia chips. Previously, China encouraged the use of Ascend, especially in light of restrictions on the export of American chips to the country.

• Even after the US allowed the sale of Nvidia H20 chips in China in July, the Chinese authorities pointed to the security risks of using American chips in critical AI projects.

Nvidia H20 remains the most popular AI chip in China, despite attempts by Huawei and the state to promote Ascend. The Ascend line has been blacklisted by the US due to the use of restricted American technologies in its development.

The previous DeepSeek model, R1, was developed primarily on Nvidia chips and became a breakthrough in the Chinese AI sector at the beginning of the year.

H20 is also used by other major AI developers in China, including ByteDance, Tencent, and Alibaba.

Read more: Nvidia has overtaken Microsoft to become the most valuable company in the world. Its value is now $3.76 trillion