Huawei Ascend and Cambricon power open-weight model deployments; CUDA alternatives emerging

AI Impact Summary

The article describes a rapid shift of AI compute toward China’s domestic chips (Ascend, Cambricon, Kunlun) for both inference and training of open-weight models (Qwen, DeepSeek, GLM, Kimi). This reduces reliance on NVIDIA GPUs and CUDA, driving architectural optimizations (MLA, GRPO) and a broader ecosystem of open-source tooling geared toward non-NVIDIA hardware. For engineering teams, this implies rethinking deployment targets, vendor risk, and cost models as models increasingly optimize for domestic silicon, with open-weight ecosystems gaining traction.