Almost 600,000 Nvidia H200s stranded in Hong Kong as Chinese gov’t stalls on import approvals – report - Data Center Dynamics

600,000 H200s sitting in a Hong Kong, China warehouse is not a logistics hiccup—it is the physical manifestation of China's AI compute sovereignty shifting from policy language to supply-chain reality. Technical cascade: The H200's 141GB HBM3e means these units have locked in HBM production slots at SK Hynix and Samsung, plus CoWoS advanced-packaging capacity at TSMC (Taiwan, China). A slip beyond two quarters converts upstream capacity reservations into sunk costs and pushes downstream model-training roadmaps back six to twelve months. Compliance exposure: High-end AI silicon now faces a dual-gate bottleneck—U.S. export authorization and Chinese import clearance. The 2023 rare-earth export controls already stress-tested this supply-chain-as-leverage playbook. Nvidia's carrying and opportunity costs on this inventory likely exceed two billion dollars. Competitive reshuffle: Huawei's Ascend 910C gains its strongest guaranteed-delivery narrative right now. AMD's MI300X penetration in China accelerates. Domestic GPU players—Moore Threads, Biren—receive policy tailwinds and procurement quotas, creating a market-for-time acceleration window. Twelve-to-twenty-four-month outlook: China's AI compute market will hard-fork into an import-dependent track and a self-sufficient track. If Nvidia cannot ship a compliance-tailored SKU within twelve months, its China data-center share slides below thirty percent. Chip sovereignty is no longer a phrase; it is the weight of six hundred thousand chips in a warehouse.