DeepSeek Plans to Train V5 Model Using Huawei Ascend 950 Chips
nashnova research
Supply-chain sources say DeepSeek plans to use Huawei's Ascend 950 chip to train its next-generation V5 model as early as 2027 — its first full training run on a domestic chip, not just inference, testing whether China's top AI lab can break free from Nvidia.
Why is this different from before?
Previously DeepSeek used Huawei chips only for inference — running a trained model, not building one. Training still relied on Nvidia.
This means → if V5 completes end-to-end training on Ascend 950, it will be the first time a top Chinese model maker proves a domestic chip can handle the most compute-intensive stage.
The project is still in hardware selection; no final decision has been made. But a joint engineering team has already completed extensive low-level operator adaptation work in Hangzhou.
How did they solve the software problem?
Core difficulty: DeepSeek's existing workflow was built on Nvidia's CUDA platform. It cannot run on Ascend directly.
The fix: the team migrated operators — the basic computational instructions a chip executes — to an open-source Python ecosystem, enabling cross-architecture compatibility. In plain terms = they don't have to rewrite all the low-level code to switch chips.
Huawei built custom operators for DeepSeek during V4 development. DeepSeek also open-sourced an Ascend-950-compatible TileLang toolkit on September 30. The software foundation was laid in advance.
How deep is Huawei's commitment?
Huawei engineers are now stationed at DeepSeek's office, feeding model-training requirements directly into future chip designs.
Huawei also plans to prioritize DeepSeek for the bulk of Ascend 960DT supply when that chip launches in Q1 2027.
This reflects a relationship that has moved well beyond buyer-seller — into joint definition of chip design and model requirements.
What is the link between the data center and V5 training?
DeepSeek plans to build a large AI compute center equipped with Ascend 950 chips in Ulanqab, Inner Mongolia, expected to be operational by 2027.
But construction is still at an early stage, and the location for V5's initial training cluster has not been disclosed. In plain terms = the data center is a long-term investment; early V5 training may start elsewhere.
A full training run may require over 10,000 chips running continuously for weeks. Large-scale pre-training on the Ascend 950DT could begin as early as late 2026.
Why Huawei and not other domestic chip makers?
DeepSeek currently works primarily with Huawei on training hardware. It has not adopted chips from other Chinese suppliers such as Moore Threads.
This means → in China's domestic AI-chip race, Huawei Ascend is currently the only homegrown option in the running for top-tier model training.
Sources described the choice as a commercial response to restricted access to advanced compute: U.S. export controls block Nvidia's latest high-end AI chips from reaching China, and previously imported H-series chips can no longer meet V4/V5 training demands. DeepSeek is still seeking to supplement with Nvidia chips, but supply is constrained.
What is the nature of the relationship?
Huawei has no equity stake in DeepSeek. The relationship is limited to product supply, compute resources, and technical adaptation.
In plain terms = this is not an investment tie-up — it is a "you build the chips, we build the models" technical partnership.
Whether V5 can complete end-to-end training on Ascend 950 will be the definitive test of whether China's leading model developers can fully break free from Nvidia dependency.
市场有风险，内容仅供研究参考，不构成投资建议。