- DeepSeek V4 is the first major Chinese AI model to be optimized for Huawei’s Ascend chips, marking a departure from Nvidia dominance.
- In a strategic pivot, DeepSeek V4 delays launch for deep integration with domestic chips, amid surging demand from tech giants.
DeepSeek’s upcoming V4 model has been fully optimized for Huawei’s Ascend chips, reversing the trend of prioritizing Nvidia’s GPUs for AI model development in China’s AI industry.
According to The Information, the decision marks a significant step towards reducing reliance on foreign technology in China’s AI landscape, as multiple tech giants, including Alibaba, ByteDance, and Tencent, rush to secure these domestically produced chips.
The V4, which was originally slated for an earlier release, saw a delay after extensive collaboration between DeepSeek, Huawei, and AI chipmaker Cambricon.
An source close to DeepSeek told The Information that engineers at the Hangzhou-based unicorn had worked tirelessly to rewrite the model’s underlying code, ensuring seamless compatibility with Ascend chips.
A key challenge was achieving precision alignment, ensuring the same model produced consistent results on both Nvidia and Ascend platforms, the source said.
The V4 model boasts a Mixture of Experts (MoE) architecture with an astonishing 1 trillion parameters, of which 37 billion are activated per inference.
It supports multimodal inputs, including text, images, and code. Huawei’s Ascend 910C and higher-end Atlas 350 graphic processing units (GPU), which provide foundational and high-end processing capabilities, are integral to its operation.
The Ascend 910C chip, launched in March 2026, is expected to compete with Nvidia’s H100, offering 800 TFLOPS of performance at FP16 precision and 80% of the H100’s inference efficiency.
With a shipment of 812,000 units in 2025, Huawei’s Ascend series represented nearly half of China’s market for AI accelerator cards, an IDC survey shows.
As demand for the new DeepSeek model surged, Huawei’s chips have seen a price increase of around 20%, media reported.
DeepSeek also rebuffed Nvidia’s request for early access to V4, providing exclusive access instead to Huawei and select Chinese chip manufacturers.
The move signals China’s growing push for technological self-reliance, as the country steadily weans itself off access to critical imports from companies like Nvidia, especially in the area of AI and semiconductor production.
