A research team that includes Huawei has completed full-parameter post-training of DeepSeek's 1.6-trillion-parameter V4-Pro model on a cluster of at least 1,000 Huawei Ascend 910C accelerators, according to Tom's Hardware. The work is one of the clearest demonstrations yet that a frontier-scale model can be trained end-to-end on Chinese silicon rather than Nvidia GPUs.
The significance is geopolitical as much as technical. US export controls are premised on the idea that restricting access to leading-edge Nvidia chips slows China's AI progress. A 1.6T-parameter model post-trained on domestic accelerators -- backed by Huawei's CANN software stack with fused operators tuned for the hardware -- suggests China is assembling a credible full-stack alternative, even if Ascend still trails Nvidia on raw efficiency.
“US export controls are premised on the idea that restricting access to leading-edge Nvidia chips slows China's AI progress.”
DeepSeek has built its reputation on doing more with less, and this milestone extends that story from inference into heavy training workloads. For Western labs, the pressure is twofold: a capable open-weight competitor that keeps driving prices down, and a reminder that the compute chokehold the US has counted on is loosening.