Nvidia orders data centre clients to reduce China exposure, FT reports - WHTC

Nvidia's move to push compliance risk onto its data center clients is a structural escalation, not an incremental tweak. Technical ripple: The CUDA-Mellanox-DGX closed stack means clients now face a dual-stack dilemma where training and inference clusters cannot be mixed. Upstream, TSMC (Taiwan, China) wafer allocation gets further bifurcated. Downstream, SaaS firms are forced to qualify alternatives, inadvertently opening order windows for AMD MI300 and Huawei Ascend. Compliance cost: Drawing on the 2018 precedent of multinationals carving out China subsidiaries, the compliance tax will consume 15-25% of client IT budgets. More critically, once dual-stack operations normalize, CUDA's developer lock-in erodes within 18 months—that is Nvidia's true moat. Market dynamics: AMD will weaponize zero compliance risk as MI300X's headline pitch. Ascend 910C accelerates from backup to primary, with CANN ecosystem subsidies likely doubling. Gaudi 3 may emerge as Europe's third option. 12-24 month trajectory: Global AI infrastructure formally splits into CUDA and non-CUDA camps. Nvidia's China revenue share could compress from roughly 25% to under 10%. The real battleground is not silicon—it is who completes the inference-side ecosystem migration by 2026. That is the trillion-dollar prize.