Nvidia tells data centre clients to scale back Chinese relations - Digital Watch Observatory

Nvidia's directive to data centre clients to decouple from China is fundamentally a compliance cost-shifting play—transferring the burden of export-control exposure from the chipmaker to end users. The technical implication is severe: CUDA's deep coupling with CoWoS advanced packaging means any client switching stacks faces an 18-to-24-month rebuild cycle for inference clusters. That is where the moat actually breaks. Strategically, this mirrors the 2018 ZTE precedent: preemptive severance to neutralise secondary-sanctions risk before regulators act. The concrete cost? Chinese hyperscalers accelerate migration toward Huawei Ascend 910C and Cambricon Siyuan 590, while AMD's MI325X captures the interim window. The 12-to-24-month trajectory points to a hardened bifurcation: Nvidia's China revenue share likely compresses from roughly 15% to under 5%, while the domestic EDA-fab-packaging loop (SMIC 7nm plus JCET) achieves first-scale production. This is not market segmentation. It is a fork in the technology tree, and the branching point is now irreversible.