Nvidia Tells Cloud Clients to Take on Less China Business - The Information

Nvidia's move is not a withdrawal—it is a risk quarantine of its cloud ecosystem, surgically decoupling China revenue exposure from core accounts. The technical cascade is already underway: CUDA's regional network effects erode, training clusters fragment, and inference-layer dependencies on locally deployed pre-trained models face supply disruption. Global AI infrastructure is bifurcating into two incompatible stacks. On compliance, GPU architecture's scale economics dictate that shrinking deployment geography inflates per-unit R&D amortization and dual-compliance overhead, forcing cloud capex reallocation. Competition-wise, AMD's MI300 and Huawei's Ascend 910C will accelerate into the vacuum, but the real battleground is not raw FLOPS parity—it is ecosystem lock-in. Once domestic cloud providers standardize on non-Nvidia silicon, CANN and CUDA fork permanently into divergent training stacks. The 12-to-24-month long-tail: inference chips, not training chips, become the next flashpoint. Eighty percent of commercial AI value sits in inference, where absolute compute requirements are lower and the domestic substitution window is widest. This is where the AI Splinternet gets sealed.