China Greenlights Limited Imports of NVIDIA H200 Chips, Easing a Key Bottleneck in AI Hardware Access

Chinese regulators have quietly reversed an earlier refusal and authorized a constrained consignment of NVIDIA H200 AI processors for selected domestic users, a decision industry sources link to recent high‑level contacts between NVIDIA executives and Chinese authorities. The authorization reportedly covers several hundred thousand H200 units and appears concentrated primarily among three unnamed internet companies, with other applicants placed in a waiting pool for possible future allocations. Washington had already approved H200 sales to pre‑cleared Chinese entities late last year, but Beijing’s prior reluctance stalled shipments until this new sign‑off. NVIDIA’s top‑end Blackwell B200 chips remain blocked under export controls, maintaining a ceiling on the most advanced acceleration capacity that can legally enter China. Separately, industry watchers note that over $1 billion worth of high‑end NVIDIA GPUs have previously reached China through unofficial channels, underscoring enforcement and diversion risks. The H200 clearance will relieve acute compute constraints for approved cloud and AI service providers and is likely to accelerate model training schedules and product rollouts among those firms. That effect will be layered on top of a parallel domestic push: several Chinese AI‑accelerator vendors have advanced toward public listings or larger procurement programs in recent months, and cloud operators are formalizing multi‑year roadmaps to integrate both imported and homegrown accelerators. Upstream bottlenecks — notably substrate supply, packaging and test throughput, wafer allocation (including DRAM and HBM prioritization) and firmware integration — will now determine how quickly approved H200 capacity can be converted into usable cluster deployments. Foundries and packaging partners have reported larger tool orders and are accelerating capex to meet hyperscaler demand, while memory makers are reprioritizing wafer allocation toward server‑grade DRAM and HBM for AI servers. Operational advantages in China, such as growing grid and data‑center infrastructure, make denser deployments more feasible, increasing the near‑term value of cleared imports. The decision reads as a calibrated, pragmatic adjustment rather than a policy reversal: it eases a short‑term commercial choke point without abandoning strategic export‑control limits on the most advanced silicon.