China bans foreign AI chips from state-funded data centers (www.tomshardware.com)

China has quietly instructed state-funded data centers to stop using foreign AI accelerators and to replace them in projects still under about 30% completion, according to Reuters sources. The retroactive ban would bar chips from Nvidia, AMD and Intel — including Nvidia parts such as H20 (designed to comply with U.S. export rules) and higher-end H200/B200 models — and mandate Chinese silicon from suppliers like Huawei, Cambricon and Enflame. The directive hasn’t been formally published but follows recent port crackdowns on Nvidia gear and comes amid more than $100 billion of public AI infrastructure spending that gives Beijing enforcement leverage. For the AI/ML community this accelerates supply-chain decoupling and will materially reshape training and inference deployments in China. Practically, projects may face hardware swaps, delays and higher operating costs as domestic accelerators generally trail Nvidia/AMD in performance density and mature software tooling (notably CUDA ecosystems). That gap risks lower utilization, higher power draw and additional engineering effort to port stacks to Ascend/Cambricon environments, while also curbing gray-market workarounds. Long term, the mandate will accelerate Chinese chip and software development but likely slow short-term progress on large-scale model training and raise geopolitical friction over AI infrastructure.