Chinese tech firms purchase more than $35 billion of Nvidia chips: Reuters - Semafor

The $35B figure signals a closing window, not a procurement milestone. These chips are almost certainly H100/H200-class, meaning China's AI training clusters will lock into the CUDA ecosystem with exponentially rising migration costs. Upstream, CoWoS advanced packaging capacity in Taiwan, China and HBM3E supply are further tethered to Nvidia's BOM. Compliance risk is the hidden bomb. The 2022 BIS restrictions proved policy can sever a supply chain within a single quarter. If these purchases transited through third-country channels, audit-trace exposure is severe. The 2018 Huawei decoupling remains the template: rebuilding a production line takes 18+ months, and the capacity gap is unbridgeable. Competitively, this is Nvidia's final China dividend. Huawei's Ascend 910C and Cambricon's Siyuan 590 are entering volume ramp. The real battleground is the software stack—operator library parity and compiler equivalence will determine the next procurement cycle. Over the next 12-24 months, three trajectories: Nvidia's China revenue cliff; HBM allocation power supplanting GPU design as the core bargaining chip; and China's AI infrastructure accelerating toward a dual-stack architecture with an ~18-month transition.