Small batches of Nvidia H200 artificial intelligence processors have entered mainland China in recent weeks, with ByteDance and Tencent each receiving roughly 10,000 units in the first meaningful deliveries since Washington cleared the chips for export. The arrivals end more than eight months of paralysis in which U.S. case-by-case licenses were already in place while Beijing held orders back, and they still leave each buyer at barely 13% of a 75,000-unit legal ceiling. Every further purchase needs individual National Development and Reform Commission sign-off that forces firms to disclose volumes and justify why domestic alternatives will not suffice, giving Beijing ongoing visibility into who is training on Nvidia hardware. Companies may also ship H200s to Hong Kong, but limited data-center power and rack space there cannot absorb full licensed volumes. The chips matter most for frontier model training, where H200 memory bandwidth still outpaces Huawei Ascend parts that Chinese groups already use for inference, while Nvidia's newer Blackwell GPUs remain barred and domestic sourcing mandates continue to favor local suppliers.