DeepSeek plans to pre-train next-generation model on Huawei Ascend chips

A shift in hardware strategy

DeepSeek is reportedly planning to use Huawei’s Ascend AI chips to pre-train its next-generation model, DeepSeek V5. This move represents a significant change in how the company handles its most demanding computing tasks. Previously, Ascend chips were primarily used for inference, which is the process of running a trained model to generate responses.

Pre-training involves teaching a model from scratch by processing massive amounts of data. This process requires much more computing power and stability than inference. If DeepSeek moves forward with this plan, it would mark a major transition for the company's infrastructure. As Geopolitechs reported, this plan has not yet been independently confirmed by the companies involved.

The partnership between the two entities dates back to early 2025. The relationship has grown steadily as DeepSeek seeks to optimize its software for domestic hardware. This shift is intended to create a more self-sufficient pipeline for developing frontier large language models in China.

Building the software bridge

To make this hardware transition possible, DeepSeek has been working on deep software integration. In September, the company open-sourced several infrastructure components designed specifically for the Huawei Ascend platform. These tools include compiler tools, high-performance computing libraries, and distributed communication software.

These components act as the bridge between the AI code and the physical chips. By creating these tools, DeepSeek is aligning its software stack with Huawei’s specific hardware architecture. This effort suggests that the company is moving beyond simple usage toward a deep technical partnership with Huawei.

The goal of this alignment is to ensure that the software can communicate efficiently with the chips. High-performance computing libraries are essential for managing the complex math required by large models. Without these specific optimizations, using domestic hardware for massive training runs would be significantly more difficult.

Testing the limits of Ascend

The move to pre-train V5 comes after some successful smaller-scale tests. A research consortium in Shenzhen recently finished full-parameter post-training on a 1.6-trillion-parameter model called DeepSeek V4-Pro. This specific project used a cluster of Huawei Ascend 910C chips to refine the model after its initial training.

While post-training is a complex task, it is still less demanding than pre-training a model from scratch. Pre-training requires maintaining stability across thousands of chips for long periods. The reported plan for V5 will serve as a critical test of whether Chinese accelerators can handle these frontier workloads.

The success of this transition depends on several technical factors. Developers will need to monitor training stability, cluster performance, and the overall maturity of the software. If the hardware can handle the heat and the data flow, it will prove the viability of the Ascend platform for the highest level of AI development.

Impact on the AI ecosystem

If DeepSeek successfully pre-trains V5 on Huawei hardware, it could change the landscape for Chinese AI development. It would provide a clear path for other developers to move away from Nvidia hardware. This would strengthen Huawei’s position as a primary domestic alternative for advanced computing power.

The broader significance lies in the reduction of dependence on foreign technology. For Chinese companies, finding reliable hardware that can support frontier models is a major hurdle. A successful V5 launch would signal that domestic chips are capable of competing with global standards for large-scale training.

This development could lead to a more isolated but self-sustaining AI ecosystem within China. It would allow for faster iteration on models without the risk of supply chain disruptions from external sources. The outcome will likely be watched closely by both domestic and international tech observers.

What comes next for V5

The next phase for DeepSeek involves moving from planning to execution. The company must now prove that the hardware-software optimization can handle the sheer scale of a next-generation model. The technical viability will be judged by the final capabilities and efficiency of the V5 model once it is released.

Observers will be looking for signs of training stability during the pre-training phase. Any significant bottlenecks in the Ascend 910C clusters could delay the timeline or force a return to other hardware solutions. The success of this project is tied directly to how well the distributed communication software handles the workload.

Ultimately, the results of this collaboration will set a precedent for the entire industry. It will determine if Chinese AI accelerators can move from being secondary tools to primary engines for frontier AI. The results of this test will define the next era of Chinese domestic AI infrastructure.