China set to limit access to Nvidia’s H200 AI chips

China is preparing to restrict access to Nvidia’s H200 AI chips, even after US president Donald Trump indicated exports to China would be allowed.

Regulators in Beijing are considering an approval process that would require buyers to explain why domestic chips cannot meet their needs, according to sources familiar with the discussions.

No final decision has been made.

The H200 is Nvidia’s second-most advanced AI chip, and had previously been barred from export to China under US restrictions aimed at limiting military applications.

Trump recently said exports would be permitted to approved Chinese customers, but a new US Senate bill seeks to block such sales for 30 months.

China has used previous US chip bans to boost local chipmakers and encourage the use of domestic technology in data centers.

🔗 Source: Financial Times

🧠 Food for thought

Implications, context, and why it matters.

Domestic chip swaps need proven performance

China’s approval plan would push firms toward local AI accelerators like Huawei’s Ascend 910C, a data center chip that reaches about 60% of Nvidia’s H100 inference speed 1.
Training lags where Nvidia leads. Ascend 910C struggles when building models, with reliability worries from less mature hardware and software integration 1.
Memory limits also bite as the chip has 128GB HBM3 2, below H200’s 141GB HBM3e with 4.8 terabytes per second bandwidth 3. Large models need more memory or higher bandwidth, so bottlenecks can appear 4.
Teams should judge if large language model (LLM) inference or scientific computing can live with about a 40% gap vs H100 on some inference tasks, since blanket limits could slow Chinese AI progress 1.

Global clouds gain leverage as H200 supply shifts

Less access in China could redirect H200 supply to regions, which may ease supply that kept rental rates at $3.00 to $10.60 per GPU-hour 5.
Enterprise buyers and cloud operators outside China should track lead times plus local prices to time purchases, since extra supply could cut prices about 15% when Blackwell GPUs (Nvidia’s next-generation data center chips) launch 6.
Largest cloud providers offering H200 in 8-GPU bundles at $80 to $85 per hour may need to trim rates if demand shifts, which helps managed service providers in bulk talks 3.
This shift opens a window for buyers outside China to lock in better terms on multi-GPU servers priced at $350,000 to $500,000 7.