China blocks Nvidia H200 imports, sources say

Chinese customs authorities have told agents that Nvidia’s H200 AI chips are not allowed to enter the country, according to three people familiar with the matter.

Chinese officials also reportedly instructed domestic tech companies during meetings to avoid buying the chips unless necessary.

Sources told Reuters that the language used by officials amounts to a de facto ban, but it is unclear if the restriction is temporary or permanent.

No official reason for the directive has been given, and it’s not clear if the move affects existing orders or only new ones.

There are discussions about possible exemptions for university research and development.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

China’s H200 freeze has limited AI impact

China shipped 507,000 Huawei Ascend chips in 2024, with 2025 at 805,000 units and 653,000 as 910C 1. Reuters says the 910C delivers 60% of Nvidia’s H100 inference performance (inference means running trained AI models) 2.
Samsung exported 11.4 million HBM stacks before controls, enough for 1.6 million Ascend packages 1. Semiconductor Manufacturing International Corporation (SMIC) is China’s largest chip foundry and targets 80,000 advanced-node wafers per month by 2027, enabling millions of Ascend dies 1. Huawei CloudMatrix 384 with 910C hits 300 petaflops BF16 vs Nvidia GB200 NVL72 at 180; its power draw is almost four times 2.

CUDA migration gains ground in China

Huawei plans to open-source its Compute Architecture for Neural Networks (CANN) toolkit and is embedding engineers in Chinese tech firms for code migration help 3 4. It keeps a forked PyTorch using the PrivateUse1 mechanism (a PyTorch namespace for custom devices) and supports ONNX Runtime (a cross-platform runtime for Open Neural Network Exchange models) 4.
Tool makers using Intel’s CUDA-to-SYCL tools and SYCL researchers (a C++-based open standard for heterogeneous computing) can win work as firms adopt dual stacks that pair Nvidia with domestic gear 5 6. Training and consulting on stability, documentation gaps, and debugging on Ascend chips bring revenue, with developers estimating CANN maturity around 2027 4. Rebuilding what Nvidia built over two decades remains hard, and China will keep funding domestic AI stacks for self-sufficiency, even without hardware parity 4 3.