Beijing to let Chinese AI companies buy Nvidia H200 chips - The Japan Times

Recent reports from The Information indicate that China is planning to allow select leading artificial intelligence companies to purchase a limited number of NVIDIA H200 chips, marking a potential shi... Read original →

Beijing’s limited approval for H200 imports reflects a tactical retreat driven by acute AI compute shortages. Technically, H200’s HBM3e and NVLink will force Chinese firms to optimize data orchestration in their AI stacks, delaying validation cycles for domestic 3nm GPUs. Compliance-wise, U.S. export conditions mandate end-use audits, raising hidden operational costs by over 15% for firms like Alibaba and ByteDance. Competitively, Huawei’s Ascend 910B and Cambricon’s MLU590 will likely accelerate software ecosystem lock-in to counter NVIDIA’s transient performance edge. Over the next 18 months, such ‘licensed imports’ won’t scale; instead, they’ll catalyze China’s push into chiplet-based heterogeneous integration and optical interconnects, fostering small-scale, NVIDIA-free compute clusters. This marks a new phase of ‘managed dependency’—neither decoupling nor convergence—but a strategic race to reclaim technical sovereignty on AI infrastructure.