The assertion that China will refuse to purchase NVIDIA’s H200 chips appears to be FUD imo.
Economic imperatives and technological realities strongly suggest otherwise.Recent policy shifts under the Trump administration have relaxed restrictions, permitting H200 exports to pre-approved Chinese customers subject to a 25% surcharge earmarked for United States purposes (Reuters).
For hyperscalers such as Tencent, ByteDance, and Alibaba, which have been constrained by the limited performance of the sanctioned H20 or by reliance on smuggled inventory, the H200’s 989 BF16 TFLOPS represent a roughly sevenfold leap over the H20’s 148 TFLOPS.Domestic alternatives, principally Huawei’s Ascend 910C, remain markedly inferior. Independent benchmarks place the 910C at approximately 76% of H200 performance per chip (15,840 TPP vs 12,032 TPP) and significantly lower efficiency at cluster level, with power consumption up to 2.3 times higher for equivalent throughput.
Production is further hampered by SMIC’s 7 nm yields, reportedly below 30%, creating chronic supply shortages. Beijing’s reported requirements—mandatory justification for foreign purchases and restrictions on public-sector use—are largely procedural and political signalling rather than outright prohibition.
President Xi’s recent “positive” remarks on United States-China technology co-operation reinforce pragmatic acceptance. In the absence of a credible domestic substitute capable of training frontier-scale models, Chinese enterprises are expected to acquire H200s in volume, potentially generating 10s of billions USD in additional revenue for NVIDIA in 2026–2027.