A research team that includes Huawei says it successfully used Huawei's Ascend 910C chips for DeepSeek V4 Pro model's post-training, amid increased US sanctions
While Chinese chipmakers have found success in supporting AI inference, they are struggling with the far more complex process of training
Context & Ripple Effects
Earlier coverage documented a gap between Huawei’s role in inference and its difficulty supporting model training: DeepSeek’s R2 reportedly encountered Ascend training problems and retained Nvidia hardware for its largest models. DeepSeek subsequently worked with Huawei and Cambricon on V4 optimization, while Huawei positioned its newer Ascend 950 supernode for V4 support.
This reported post-training result is therefore a narrower but meaningful milestone in Huawei’s effort to move Ascend from deployment workloads into more of the model-development pipeline under tighter sanctions pressure.
First-order effects
- Huawei gains a concrete reference workload for Ascend 910C in DeepSeek V4 Pro’s post-training stage, rather than only claims around inference or future platform support.
- DeepSeek has evidence that at least one late-stage training workflow can run on Huawei hardware, potentially expanding its hardware options for that part of V4 Pro development.
Second-order effects
- Chinese AI-chip vendors will face greater pressure to demonstrate reproducible performance across distinct training stages, not merely inference or optimization support.
- Model developers may increasingly split workloads by capability—using domestic chips where they are validated and retaining other hardware for the most demanding training tasks—rather than making an all-or-nothing platform choice.
Third-order effects
- If similar results extend beyond post-training, China’s AI stack could become less dependent on a single foreign hardware supplier across model development; this report alone does not establish parity in full-scale pretraining.
- Sanctions pressure is likely to make software-hardware co-optimization between model labs and domestic chipmakers a central competitive lever, with practical ecosystem support becoming as important as chip specifications.
The trend: This is one data point in the push to turn Chinese AI accelerators from inference alternatives into viable components of the training workflow through close model-and-hardware co-development.