Can Chinese silicon replace Nvidia? Here are 5 AI models trained on local chips - South China Morning Post

As China's artificial intelligence sector rapidly advances, domestic chips are increasingly used in model inference, but pre-training—still largely dependent on foreign hardware like NVIDIA's GPUs—rem... Read original →

U.S. GPU export controls are forcing China’s AI sector to rebuild its tech stack from the ground up: migrating training workloads to domestic chips like Ascend and Cambricon sacrifices short-term speed but accelerates vertical integration of compilers, communication libraries, and frameworks. While compliance costs surge, supply chain fragility declines—especially critical amid restrictions on HBM and advanced packaging. NVIDIA may cling to Chinese revenue via downgraded H20 chips, but its CUDA moat is eroding; once local silicon proves stable at trillion-parameter scale, developer lock-in could shift irreversibly. Over the next 18 months, Chinese chipmakers must solve software maturity and large-scale cluster reliability. Success could split the global AI infrastructure into two parallel ecosystems: one CUDA-based, the other built on homegrown ISAs and interconnect protocols.