Nvidia tells data centre partners to curb business with Chinese clients, FT reports - Malay Mail

Nvidia's directive to data center partners to restrict Chinese client business is not a compliance footnote—it is a structural reallocation of export-control enforcement costs from the chip design layer to the systems integration layer. The technical ripple extends well beyond the GPU. OEM partners must now maintain dual BOMs and firmware stacks for identical AI servers. TSMC's (Taiwan, China) advanced-node capacity allocation fragments further, and cloud deployment cycles stretch by 3-6 months. The deeper threat targets CUDA's de facto standard status: artificially severing hardware supply chains for the first time systematically dilutes a software ecosystem's gravitational pull. On compliance economics, benchmarking against post-H100 restriction data, friction costs consume roughly 8-12% of transaction value. Gray-channel routing through third countries will force Nvidia into real-time global transaction tracking, driving operational complexity to exponential levels. In the competitive landscape, AMD's MI300X and Huawei's Ascend capture immediate upside, but ROCm immaturity caps the former while sub-7nm and HBM constraints bind the latter. Intel's Gaudi 3 is the quiet winner—precisely occupying the "compliant-enough, performant-enough" middle tier. 12-24 month outlook: global AI compute irreversibly bifurcates into "compliance-zone" and "sovereign-zone" parallel stacks. The "partner directive" model becomes the new regulatory paradigm—more flexible than bans, more opaque than licenses, adjustable in real time. Chip control has formally migrated from product-level to relationship-level governance.