DeepSeek is preparing to launch its V4 model on Huawei's Ascend chips, marking a decisive shift toward domestic hardware independence for China's most prominent AI lab.

Nvidia Shut Out

According to The Information, DeepSeek has granted early access to V4 exclusively to domestic chipmakers including Huawei and Cambricon Technologies, while denying it to Nvidia and AMD. This breaks the industry norm of collaborating with multiple chip vendors before a major model release.

The company has spent months rewriting core parts of V4 to run efficiently on Huawei hardware, reflecting lessons learned from a painful setback: its previous reasoning model R2 failed during training on Huawei Ascend 910C chips due to a maturity gap between Huawei's CANN software stack and Nvidia's CUDA ecosystem.

Bulk Orders Signal Confidence

Chinese tech giants Alibaba, ByteDance, and Tencent have placed bulk orders totaling hundreds of thousands of Huawei's upcoming AI chips. The scale of these orders suggests strong industry confidence in Huawei's ability to support frontier AI workloads.

What's Coming

V4 is expected to feature a next-generation dynamic computation architecture with a reported 1 trillion parameters. DeepSeek is also developing two additional V4 variants optimized for different capabilities, all designed to run on Chinese-made hardware.

The move carries significant geopolitical weight. DeepSeek's earlier R1 model release triggered a single-day loss of $589 billion in Nvidia's market capitalization. If V4 demonstrates that frontier AI models can be built without American chips, it could accelerate the decoupling of the US and Chinese AI ecosystems.