Strategic Shift Toward Domestic Hardware
DeepSeek is reportedly planning to pre-train its next-generation large language model, DeepSeek V5, using Huawei Ascend chips, according to industry supply chain sources cited by STAR Market Daily. This move marks a significant transition for the company, which has previously utilized Huawei’s Ascend hardware primarily for inference tasks rather than the computationally intensive pre-training phase.
The development follows a period of rapid integration between the two entities. According to Fang Xingdong, Executive Dean at Zhejiang University, Huawei leadership directed Huawei Cloud to provide urgent support to DeepSeek during the 2025 Lunar New Year. During that period, Huawei and DeepSeek reportedly completed the deployment and verification of over 3,000 Ascend chips to host DeepSeek-V3 and R1 models at the Huawei Cloud Guian Data Center.
Software Stack Alignment and Infrastructure
DeepSeek has moved beyond simple hardware utilization by systematically aligning its software stack with the Ascend architecture. On September 30, the company announced the open-source release of several infrastructure components specifically tailored for Huawei’s platform. These include the TileLang high-level language compilation tool, high-performance computing libraries such as DeepGEMM, FlashMLA, TileKernel, and DeepSelect, and the distributed communication library DeepEP.
These releases correspond to the open-source components previously developed for NVIDIA platforms. To support this integration, Huawei has provided the “Ascend SuperPoD Flex” architecture and a UBL128 networking solution, while also opening the Ascend C programming interface to support TileLang.
While large-scale deployment for inference has been established, the transition to training remains a complex technical hurdle. A project team involving the AI Training Platform at Shenzhen Hetao College, the Harbin Institute of Technology (Shenzhen), and the Shenzhen Big Data Research Institute recently utilized an Ascend 910C processor cluster to complete the full-parameter post-training of the 1.6 trillion-parameter DeepSeek-V4-Pro model. However, this specific achievement did not include the initial pre-training phase.
Market Implications and Performance Constraints
The shift toward domestic hardware carries significant economic implications for DeepSeek’s service offerings. Following the release of DeepSeek-V4 in April 2026, the company noted that service throughput for the V4-Pro model remains limited by the availability of high-end computing power. DeepSeek anticipates a reduction in the price of the Pro version once the Ascend 950 super-node series reaches mass-market availability, expected in the second half of the year.
Industry observers on platforms like Reddit have noted that while some developers have seen performance improvements on consumer-grade hardware—such as unverified reports of 20x faster agent session starts on standard gaming GPUs via custom patches—the broader industry remains focused on the scalability of the non-CUDA software stack. If DeepSeek successfully completes the pre-training of V5 on Ascend hardware, it would represent the first large-scale validation of China’s top-tier models being trained entirely on domestic computing infrastructure.