Sources: DeepSeek CEO Liang Wenfeng says training on Huawei chips is one of DeepSeek's biggest bets and Huawei is set to deliver training chips in Q4 or Q1 2027
DeepSeek CEO Liang Wenfeng told investors that a major priority for his company is to use more domestic chips to train its models …
Context & Ripple Effects
DeepSeek’s Huawei strategy has been staged rather than wholesale: 2025 reporting tied an R2 delay to Ascend training problems and a switch to Nvidia for training, while a later plan limited Huawei training to smaller R2 variants and kept Nvidia for the largest models. The reported new commitment follows DeepSeek’s planned large Ascend 950DT deployment in Inner Mongolia, raising the stakes from targeted workloads to training capability.
China’s advanced-chip constraints were already identified by Liang Wenfeng as a bottleneck in 2025. The reported delivery timetable makes Huawei’s next training hardware a test of whether DeepSeek can move beyond the earlier Ascend-related training setbacks without depending as heavily on Nvidia for its largest training jobs.
First-order effects
- The reported chip deliveries would position Huawei as a prospective supplier for more of DeepSeek’s training workloads, beyond the smaller-model and inference roles described in earlier coverage.
- DeepSeek would be committing model-development plans to Huawei’s training-chip roadmap, making training performance and delivery timing operational constraints for its next large models.
Second-order effects
- Huawei would gain a demanding deployment through which to demonstrate whether its training systems can support frontier-scale workloads after DeepSeek’s earlier Ascend difficulties.
- If DeepSeek shifts additional training onto Huawei hardware, Nvidia faces a narrower role in an account that had previously reserved Nvidia chips for the largest models.
Third-order effects
- The pattern points to a bifurcated AI-compute stack in China: developers retain Nvidia where available while qualifying domestic hardware for increasingly central training workloads.
- A successful transition would make hardware-software co-development, rather than access to a single chip supplier, a more important competitive capability for Chinese model builders.
The trend: Chinese AI developers are turning domestic accelerators from inference and secondary-training options into strategic alternatives for large-model training as advanced-chip access remains constrained.