Ascend AI chips surpass Nvidia in China as Huawei weans off American tech - Notebookcheck

The Ascend-Nvidia crossover in China is not a spec-sheet victory; it is a policy-engineered ecosystem migration. The CUDA moat is being dismantled not by technical parity but by compliance pressure forcing a parallel software stack. The binding constraint has shifted from silicon to software. CANN operator coverage and developer migration friction will dictate actual penetration over the next 18 months. Da Vinci is inference-viable but structurally weak in training workloads—memory bandwidth and interconnect topology remain hard constraints that no single-generation leap will close. Nvidia's counter will not be a competitive product. The H20/B30 compliance SKUs are cash-flow preservation plays, not technology leadership. AMD is filling the mid-tier gap but lacks domestic advanced-node manufacturing credibility in China, capping its ceiling sharply. The 12-to-24-month inflection: whether Alibaba Cloud, ByteDance, and peers shift 40% or more of inference workloads to Ascend. Cross that threshold and CUDA's position in China enters an irreversible decline channel. The global AI chip market will formally bifurcate into two incompatible technical ecosystems—a structural split far more consequential than the 2019 Entity List ever was.