DeepSeek plans to train its V5 model on Huawei Ascend 950 chips

The Chinese AI lab reportedly intends to deploy at least 160,000 Huawei Ascend 950DT chips at a new data center in Ulanqab, Inner Mongolia. This represents a significant shift for DeepSeek, which has previously relied on Nvidia hardware for its most demanding AI training tasks.

The chips were initially meant for inference—the process of answering user prompts with pre-built models. However, DeepSeek reportedly plans to use them for the more computationally intensive pre-training phase, where models learn from vast amounts of data, starting as soon as late 2026.

Hardware Challenges

DeepSeek has previously struggled with stability and performance issues when training models on earlier-generation Ascend chips. To avoid repeating these problems, Huawei engineers have reportedly embedded with DeepSeek's team to customize the chips' design to the lab's specific model requirements.

Software and Timeline

The Ascend 950DT features 144 GB of HBM-style memory and approximately 4 TB/s of bandwidth. Huawei targets a commercial release in Q4 2026, which closely aligns with DeepSeek's reported pre-training timeline, leaving little room for delays.

DeepSeek has developed open-source libraries optimized for Ascend hardware, reportedly achieving over 95% of the chip's peak performance during tests. Despite these efforts, the lab acknowledges that using Huawei's chips requires more hardware, power, and data center space than Nvidia alternatives.

Strategic Shift

This pivot toward domestic hardware is largely driven by US export restrictions on advanced AI chips bound for China. DeepSeek's approach combines open-source software with high hardware utilization to reduce dependence on foreign technology.

The broader industry roadmap suggests mainstream adoption of Huawei's Ascend stack could begin as early as 2027, marking a significant development in the global AI chip race and potentially reshaping supply chains for artificial intelligence development.