Huawei and DeepSeek are reportedly joining forces to improve Huawei’s Ascend AI processors and their supporting software, with the work focused on the bottlenecks that emerge when training large AI models. The effort could strengthen Huawei’s hardware-software stack as it seeks to compete more directly in a market led by Nvidia.
According to The Information, DeepSeek will provide Huawei with feedback on processing challenges encountered while running large language models. Huawei is expected to use that information in future Ascend chip designs.
The reported engineering priorities are memory bandwidth and communication between processors. Both are critical in large-scale AI systems, where thousands of chips may need to operate together; improvements could reduce model-training time and associated costs.
The collaboration also extends to software. Huawei plans to enhance CANN, its software platform positioned as an alternative to Nvidia’s CUDA ecosystem. The companies are also targeting an open-source ecosystem that would include a local AI model intended to help optimize hardware and software components.
DeepSeek’s experience building AI models designed to deliver strong performance with fewer resources could be relevant to that optimization effort. Huawei has already outlined its 2027 Ascend roadmap, with the Ascend 960DT scheduled for the first quarter and the Ascend 960PR expected in the third quarter. Whether the partnership can translate those plans into a meaningful performance and ecosystem challenge to Nvidia remains the central unanswered question.
