Nvidia Tells Data Centres To Reduce China Client Base - Silicon UK

Nvidia's directive to data center clients to shrink their China end-user share is not a sales tweak—it is preemptive compliance architecture. The company is drawing its own decoupling line before regulators force the blade. Technical cascade: When hyperscalers reconfigure GPU procurement, CoWoS advanced-packaging allocation in Taiwan, China bears the first shock. Ascend 910B and Cambricon 590 fill the hardware gap, but CUDA migration costs create a 12-month "hardware works, ecosystem fractures" window. Compliance reality: This internalizes BIS export-control reach into commercial contract language. Mirroring the 2020 Huawei supply-chain rupture, Nvidia is building a compliance firewall at the cost of roughly 15–20% of incremental revenue. Competitive positioning: AMD's MI300X and Intel Gaudi 3 slot precisely into this gap. More critically, once China's training clusters drop below the CUDA-dependence threshold, Nvidia's moat narrows irreversibly. 12–24-month trajectory: Global AI compute solidifies into a dual-track system—West locked to Nvidia + TSMC + HBM; East converging on domestic GPUs + SMIC N+2 + CXMT HBM. Nvidia's voluntary contraction confirms a hard truth: technological lead cannot offset institutional decoupling. The true long-tail effect? CUDA open-sourcing gets accelerated by coercion, not choice.