DeepSeek and Huawei work on AI chip designs and tools

The Information reported that DeepSeek is working closely with Huawei Technologies to shape future AI chip designs, citing two people familiar with the matter.[1] Reuters separately described a partnership to develop chip-programming tools.[4] DeepSeek said it would open-source software customized for Huawei AI chips, including a Huawei-compatible version of TileLang.[3] The Information said DeepSeek has also used TileLang, a high-level programming language, for Nvidia chips.[3]

Techmeme reported that DeepSeek plans to start using Huawei’s Ascend 950DT for pre-training as soon as the end of 2026.[2] Bloomberg reported a separate deployment plan involving at least 160,000 Huawei accelerators at a data center DeepSeek is building in Inner Mongolia.[6] Bloomberg said the planned deployment could create one of the largest known clusters of Huawei AI chips.[6] These are reported plans, not evidence that the chips are already deployed or that the stated schedule has been met.[2][6]

The Information said DeepSeek has decided to use Huawei AI chips to train some of its models, citing three people with knowledge of the effort.[5] It described the decision as a sign of reduced reliance on Nvidia chips.[5] The Information said the broader collaboration could help China build frontier AI models without Nvidia technology.[1] Reuters characterized the partnership as a sign that Chinese technology firms are deepening ties as they seek alternatives to Nvidia’s ecosystem.[4] The reports establish a direction of travel, but do not quantify how much of DeepSeek’s training will shift to Huawei hardware.