DeepSeek Joins Huawei's AI Chip Design Effort to Cut Nvidia Dependence

DeepSeek is collaborating closely with Huawei on the design of future AI chips, according to reports. The partnership goes beyond a typical supplier-customer relationship as the two Chinese tech firms aim to reduce reliance on Nvidia chips.

As part of the collaboration, Huawei engineers are being stationed at DeepSeek offices in phases to feed real-world large model training requirements into chip design decisions. Feedback on how computing tasks are distributed and how processors transfer data will be incorporated into a chip design planned for early 2027.

The next step is shifting training workloads. DeepSeek plans to start pre-training with Huawei's Ascend 950DT as early as late 2026, with tests still underway. Large-scale pre-training typically requires clusters of thousands to tens of thousands of AI chips running for weeks, placing high demands on compute, communication and stability. Huawei also plans to allocate most of the production capacity of its Ascend 960DT, slated for Q1 2027, to DeepSeek for both training and inference. DeepSeek founder Liang Wenfeng has told investors that using Huawei and other domestic chips to train next-generation models is one of the company's top priorities.

On the software side, DeepSeek has open-sourced the TileLang compiler for the Ascend platform, along with matching compute and communication components. TileLang runs across chips, and every operator used in DeepSeek's training has been implemented with high performance on Ascend. According to IDC, Nvidia's share of China's AI chip market fell to 34% in the first half of 2026, with Huawei at 28%.