DeepSeek and Huawei are working on joint development of AI chips — China is trying to reduce its dependence on Nvidia

Chinese artificial intelligence startup DeepSeek and technology company Huawei Technologies are collaborating on creating chips customized for AI models. According to The Information, the partnership will give China the ability to develop advanced AI systems without relying on American company Nvidia.

As part of the collaboration, Huawei engineers are working at DeepSeek's offices, using feedback received from the company to improve chip design. Attention is being paid to technical issues such as how AI models perform computations and how data is transferred between processors.

DeepSeek plans to use Huawei's Ascend 950DT chips for model training by the end of 2026. In addition, Huawei intends to allocate most of its next-generation Ascend 960DT chips, expected in the first quarter of 2027, to DeepSeek.

Technological independence is part of China's strategy

The partnership serves Beijing's broader goal — developing an AI industry in China that is less dependent on American technology. Despite US export restrictions, Nvidia still plays a significant role in training China's advanced AI models.

However, the technical capabilities of Huawei's chips still remain a significant challenge. According to estimates by research institute Epoch AI, the raw arithmetic performance of Nvidia's B300 chips exceeds that of Huawei's Ascend 950 processors by approximately sevenfold.

As a result, training an advanced AI model entirely on Chinese chips could require more than 10,000 Huawei processors operating simultaneously for several weeks. At this scale, both chip manufacturing capacity and the stability and reliability of the entire system would be critical.

DeepSeek is also becoming less dependent on Nvidia's software ecosystem

DeepSeek is also using the open-source programming language TileLang, which helps developers adapt AI software to various chip architectures. This approach could reduce dependence on Nvidia's CUDA ecosystem. Last month, DeepSeek released a version of TileLang compatible with Huawei's Ascend 950 chips.

Nevertheless, DeepSeek is not fully distancing itself from Nvidia and continues to seek access to its advanced processors, according to The Information.

If the collaboration enables DeepSeek to train an advanced AI model entirely on Chinese chips, this could become an example for other Chinese AI companies to follow. Moreover, such an outcome would represent a significant step in China's effort to reduce its dependence on US technology.