DeepSeek reportedly plans to train its V5 model on Huawei chips instead of NVIDIA's

What the report says

As TechNews reported, citing several Chinese tech outlets and supply-chain sources, DeepSeek is planning to hand training of its next large language model, V5, to Huawei. The work would run on Huawei's Atlas SuperPoD systems, which are large clusters of linked chips, built around Ascend 950DT processors.

If it happens, V5 would be DeepSeek's first large model trained from scratch on the Ascend platform. It would also be the first large model built from scratch on Chinese chips, which is the claim that remains unproven.

From running models to building them

The key detail is the type of work. DeepSeek has so far used Ascend chips mainly for inference, meaning running a finished model to answer requests. Pre-training, the first and most compute-hungry stage of building a model, is a much heavier job.

Sources say DeepSeek and Huawei already did considerable work on low-level algorithm tuning and customization during the V4 stage, which would have paved the way for this move.

The software groundwork

Reports also say DeepSeek open-sourced infrastructure components for Huawei's Ascend platform at the end of September. They include TileLang, a compiler tool for writing high-level code that runs on chips, plus a compute library and a distributed communication library. These mirror components it had earlier released for NVIDIA hardware.

Observers read this as an effort to build a fuller training and deployment ecosystem around Ascend. The broader backdrop is tightening US export limits on advanced AI chips.

What to treat as fact

This is a rumor. The reports we reviewed show no confirmation from DeepSeek, and no weights, license or technical report exist for V5. Whether Ascend chips can carry a training run at this scale is exactly what a release would need to demonstrate.