DeepSeek and Huawei team up to develop their own AI chips and reduce reliance on Nvidia

A report published by The Information on Monday, 19th, revealed that Chinese startup DeepSeek is collaborating with Huawei to develop AI processors tailored to its models. The aim is to reduce dependence on Nvidia chips during AI training, given the restrictions on access to American components.

Huawei engineers work at DeepSeek’s offices, making adjustments to both calculations and communication between processors. According to the same report, the company plans to start using the Ascend 950DT chip for training as early as the end of 2026.

Why software-hardware integration is a strategic move

Large AI models distribute tasks across multiple chips. In addition to performing mathematical operations, these chips need to exchange data quickly. If a processor has to wait for information from another, overall performance drops, even if it still has processing capacity available.

DeepSeek has publicly released two tools that demonstrate this integration. DeepGEMM-Ascend combines calculation routines optimized for Huawei’s Ascend 950 chip family, while DeepEP-Ascend manages communication between processors during training and inference. DeepGEMM-Ascend was released on September 30th and has already been validated on the Ascend 950 platform.

Breaking free from the Nvidia ecosystem

Switching suppliers also requires software adaptation. Nvidia offers CUDA, a comprehensive set of compilers, libraries, and tools that simplify development on its chips. To ease the transition, DeepSeek created TileLang, a language that enables the creation of routines compatible with multiple processors, including those from Nvidia and Huawei. The TileLang repository has supported the Ascend 950 natively since September 30th, while also maintaining compatibility with other platforms. However, TileLang does not fully replace CUDA’s functionality, despite its usefulness.

Current status of the partnership and future challenges

The tools already show tangible progress, but they still have limitations. DeepEP-Ascend tests were conducted in a manually adjusted experimental setup, which may not reflect the performance of a full training model.

The collaboration comes at a time when China is seeking alternatives to American semiconductors and facing challenges in local production. For the partnership to become a viable replacement, technical integration must be combined with an industrial-scale hardware supply.

In conclusion, the partnership between DeepSeek and Huawei marks a significant step towards diversifying China’s AI chip ecosystem, but it is still far from completely eliminating dependence on Nvidia.