Roughly 750,000 H200 GPUs that NVIDIA secured US export licenses to ship to Chinese AI firms โ Alibaba, Tencent, ByteDance, JD.com, and around ten firms total, each licensed for up to 75,000 units, with Lenovo and Foxconn authorized as distributors โ are sitting undelivered. Not a single H200 has shipped to China since Trump first authorized the sales in December 2025. The Trump-Xi summit came and went without semiconductor controls on the bilateral agenda. Jensen Huang was added to Trump's delegation at the last minute. Commerce Secretary Lutnick named the dynamic: Chinese firms are keeping investment focused on domestic suppliers including Huawei. NVIDIA's China revenue dropped from above 20% historically to roughly 5%, and current-quarter guidance assumes zero China revenue.
The deal is frozen because the two regulatory regimes are mutually exclusive. Washington's licensing condition: H200 chips must be used only in China. Beijing's directive to Chinese firms: limit Nvidia chips to overseas operations while supporting domestic manufacturing. There is no overlap. The chips are physically allocated, the licenses are firm-specific, the export paperwork is clear โ but the buyers cannot use them under their own government's constraints. 750K H200 units is real allocation: each card carries 141 GB HBM3e and roughly 4.8 TB/s memory bandwidth, making the H200 the workhorse generation for current production LLM inference. This is a chunk of global supply, not a marginal SKU.
Two ecosystem effects to track. First, supply: 750K H200s in allocation limbo represent a non-trivial fraction of NVIDIA's near-term global production. NVIDIA has not formally announced reallocation, but the chips physically exist. If even half redirect to non-China buyers, that is 300K+ additional H200 units entering the open market โ material for CoreWeave, Lambda, Nebius, and the cloud operators queued for H200/H300 deliveries, and material for sovereign AI buildouts in Europe and the Middle East. Second, structural: NVIDIA's China revenue dropped from above 20% to roughly 5%, with current-quarter guidance at zero. That is a structural shift in revenue mix, not a transitional blip. The China-domestic-silicon thesis (Huawei Ascend, SMIC nodes, Cambricon, MetaX) is now Beijing strategic priority, validated by Beijing's own directive blocking H200 deployment.
Monday: if you are capacity-planning H200 inference clusters, the 750K allocation question matters for your near-term cost curve. NVIDIA's next earnings call is where reallocation guidance, if any, will surface โ track that. If your production LLM workload is H200/H300 inference-heavy, the secondary supply story could improve availability and pricing through Q3 and Q4. For longer-term planning, the deal's structural failure validates that any China-route hedge in your hardware sourcing is fragile; plan on US, EU, and Asia-ex-China NVIDIA supply only, even when bilateral approvals look formally clean. The geopolitical premium on NVIDIA is now structural. On the China side, watch Huawei Ascend production volumes and SMIC's 5nm node yield numbers โ those, not H200 import licenses, are the real China AI hardware story of 2026.