Chinese AI firms hit by Nvidia H200 shortage after border delays

Chinese AI firms face a shortage of Nvidia’s H200 chips, as customs authorities currently hold the shipments at the border despite US approval for export, according to sources familiar with the matter.

The hold-up has led some buyers to consider sourcing chips from the black market at significantly higher prices, with a bundled server of eight H200 GPUs selling for around 2.3 million yuan (US$330,400), about 50% above the official price.

Nvidia’s CEO Jensen Huang expressed optimism about strong demand in China and plans to visit the country this month to support market reopening efforts.

The delay is believed to be related to Beijing’s focus on supporting domestic AI chip development and reducing reliance on foreign suppliers, particularly amid ongoing US-China technology tensions.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Crucial details on performance and software maturity should be assessed

The performance trade-off between Nvidia H200 chips and domestic options such as Huawei Technologies must be spelled out for training large AI models, since inference usually needs less compute.
Huawei’s Ascend 910C has been reported at about 60% of Nvidia H100 performance, while other reports put it near 80%, which changes how training limits are judged 1.
The full cost of moving off Nvidia’s CUDA ecosystem (Nvidia’s widely used software platform for AI development on its GPUs) must be checked, including rewrite work, missing tools, plus throughput loss during the move.
Huawei’s CANN (Compute Architecture for Neural Networks) has been called difficult and unstable by developers, with hands-on help often needed from Huawei to get code running 2.
Supply reliability should be weighed since Ascend 910C yield has been reported near 40%, below a 60% target Huawei is said to be pursuing as an industry benchmark 1.

An opportunity exists for third parties to bridge the CUDA-to-Ascend gap

Software operators and service providers can build a growing business that helps teams move from Nvidia CUDA to platforms like Huawei Ascend.
Chinese AI developers run into heavy drag when they port code, which means rewriting plus tuning to fit Huawei’s CANN stack 2.
Vendors can sell porting utilities, compatibility layers, or tuning services that hide chip differences and reduce rework.
Huawei is putting effort into PyTorch support through torch_npu (a connector that lets the PyTorch AI framework run on Huawei’s neural processing units), so firms can specialize in that stack and help labs lower supply-chain risk without rebuilding models from scratch 3.