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
Context & Ripple Effects
DeepSeek’s Huawei strategy has been uneven: reported Ascend training problems delayed R2 and led the company to use Nvidia for training while retaining Huawei for inference in 2025. Its later plan to use Ascend for smaller R2 variants marked a narrower re-entry into Huawei-based training. The earlier Ascend training difficulties are therefore the key benchmark for the reported larger commitment.
The reported chip delivery would arrive alongside DeepSeek’s planned deployment of more than 160,000 Ascend 950DT chips in Inner Mongolia, extending its Huawei relationship from a large inference cluster toward model development. It also sits beside DeepSeek’s reported work on its own inference chip, which points to a broader effort to diversify compute supply rather than rely on one vendor.
First-order effects
- DeepSeek’s reported commitment makes the forthcoming Huawei training hardware a direct test of whether Ascend can support its model-development workload after the earlier training setbacks.
- Huawei gains a high-profile prospective training customer alongside DeepSeek’s planned large Ascend 950DT deployment, tying the credibility of its new training chips to a demanding AI lab.
Second-order effects
- Nvidia retains an important role in DeepSeek’s reported strategy for the largest models until Huawei’s new chips demonstrate that they can handle training workloads at the required scale.
- DeepSeek’s infrastructure planning must accommodate two hardware paths—Huawei for the reported Ascend buildout and Nvidia for top-end training—raising the value of software and operations that can support both.
Third-order effects
- If DeepSeek can move substantial training onto Huawei hardware, China’s AI developers will have a more credible second-source compute option for advanced model work rather than only inference.
- The pattern points toward vertically diversified AI stacks, where model builders combine domestic accelerators, external suppliers, and in-house chip efforts to manage hardware constraints.
The trend: Chinese AI developers are pursuing second-source compute strategies that expand domestic chips from inference deployments into the harder training layer.