DeepSeek is reportedly planning to pretrain its next major model, DeepSeek V5, on Huawei's Ascend chips. Several people in the chip supply chain told Star Market Daily (科创板日报). Until now, Ascend chips have mainly handled inference, the work of answering user requests with an already trained model. Using them for pretraining would put them to work building the model itself, the most expensive and demanding stage of the process.
The link between DeepSeek and Ascend is not new. DeepSeek's first large-scale deployment on the Ascend platform came around the 2025 Lunar New Year. According to The Rise of Ascend by Fang Xingdong, executive dean of the College of Media and International Culture at Zhejiang University, Huawei's senior leadership decided in late January 2025 to quickly support DeepSeek through Huawei Cloud. Between New Year's Eve and the second day of the Lunar New Year, Huawei and DeepSeek reportedly deployed and verified more than 3,000 Ascend chips, bringing the V3 and R1 models to Huawei Cloud's Guian data center. In April 2026, after the release of DeepSeek-V4, Huawei said its full Ascend supernode lineup now supports the V4 series. DeepSeek said V4-Pro's capacity remained limited by access to advanced compute, and that prices would fall once the Ascend 950 supernode reached volume availability.
In September 2026 DeepSeek also open-sourced components for the Ascend platform, including the TileLang compiler, high-performance computing libraries and a distributed communication library, matching those already released for Nvidia GPUs. It is not only a matter of hardware use: the underlying software has been aligned with Chinese chips.
An earlier episode matters. Last year DeepSeek tried to train its R2 model on Huawei hardware, but stability, chip communication and software problems pushed the company to finish training on Nvidia GPUs. TokenPost reports the same episode. The open question, then, is not whether Ascend chips can do the work, but whether the platform can sustain large-scale training.
With V5, DeepSeek would take a new step: pretraining on Ascend 950 would be the first native use of this chip generation to train a frontier model. TrendForce reports that the Ascend 960DT, expected in the first quarter of 2027, is now set three quarters ahead of its original schedule, citing Huawei rotating chairman David Wang. Huawei would reportedly prioritize DeepSeek for much of that supply.
Capacity figures remain the weak point. According to the sources, China's advanced chip output is constrained by SMIC's limits and US export restrictions. Huawei itself has acknowledged that its production capacity does not cover domestic demand. Sources also say HBM memory for these chips is scarce, and that Huawei reportedly asked some customers to source memory themselves, including through parallel channels, according to The Information as relayed by The Decoder.
If confirmed, V5 pretraining on Ascend would offer a first answer to a question open for some time: whether a leading Chinese model can be built entirely on domestic compute. For now, DeepSeek and Huawei have not publicly commented on the plan.
No comments yet. Be the first!