DeepSeek delays V4 to optimize for Huawei chips

DeepSeek delayed its V4 model to tune its software for Huawei’s Ascend chips, in a move that fits Beijing’s push to build a domestic AI supply chain as access to foreign chips tightens.

DeepSeek spent months reworking its software stack for hardware-specific tuning rather than basic compatibility, and the company released preview versions of a new flagship model in late April.

Wider use of Chinese semiconductors and more capital support are helping companies train and run large language models on local compute, although top global platforms still perform better.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

V4 performance and price drive interest in domestic chips

DeepSeek says V4-Pro matches top models from OpenAI and Anthropic, two US AI companies, on major benchmarks 1.
DeepSeek pairs that with low pricing. Its V4-Pro application programming interface (API) costs US$1.74 per million input tokens, far below Western rivals 1.
Efficiency supports that push. In a 1 million-token context, V4-Pro uses 27% of the computing power required by V3.2, according to DeepSeek 1.
The company has done this before. DeepSeek says R1, released in January 2025, cost less than US$6 million to build, which cuts against the idea that AI leadership needs huge spending 2.

DeepSeek helps chipmakers while pressuring rivals

The launch hurt domestic rivals. Shares in MiniMax and Knowledge Atlas Technology, also known as Zhipu, each fell about 8% 2.
Chipmakers moved higher. Shares in SMIC rose 9% 2. Hua Hong Semiconductor climbed 15% 2. SMIC is China’s largest contract chip manufacturer. Hua Hong is another contract chipmaker.
That fits Beijing’s goal of an “independent and controllable” technology ecosystem. Xi Jinping stressed that theme in an April 2025 Politburo study session on AI, a formal meeting of the Communist Party’s top leadership 3.
DeepSeek still seems partly reliant on Nvidia. Its technical report says Chinese chips handle inference, which means running the model after training. Only part of V4 training appears adapted for Chinese chips, and the report does not say whether Nvidia chips did most of the training 1.