Nvidia pushes cloud customers to cut back on China business - the-decoder.com

Nvidia pressuring cloud customers to scale back China operations signals a critical escalation: export controls are migrating from physical chip delivery to the service layer. AWS, Azure, and Oracle Cloud are no longer distribution channels—they have been redefined as compliance chokepoints. Technically, the CUDA ecosystem is undergoing "soft bifurcation." Once China-region deployments shift to downgraded silicon or Huawei Ascend 910C, cross-platform adaptation costs for PyTorch and TensorRT will grow exponentially. This is the most underpriced engineering bottleneck of the next 18 months—more lethal than the silicon gap itself. On compliance, BIS entity-list restrictions intersect with China's data-security regulations, creating a dual-audit trap. No single data-center architecture can satisfy both regimes simultaneously. Physical isolation becomes the only viable path, structurally doubling capex and forcing a full rewrite of cloud TCO models. Competitively, AMD's MI300X and Intel Gaudi 3 will attempt to exploit the regulatory gray zone, but the true structural beneficiary is China's domestic GPU cohort. Policy certainty is the strongest commercial catalyst available; Cambricon and Huawei's ecosystem positions are being institutionally locked in. 12-24 month call: global AI compute completes a hard fork. China consolidates around Ascend + Cambricon as a sovereign stack; the West locks in a Nvidia-AMD duopoly. Cloud providers' China revenue shifts from a P&L growth line to a compliance cost center. Margin compression is structural, not cyclical.