Nvidia reportedly to boost H200 AI chip supply for China

Nvidia is considering ramping up production of its H200 AI chips after receiving more orders from Chinese companies than it can currently fulfill, according to sources.

The H200, produced by Taiwan’s TSMC, is Nvidia’s second-fastest AI chip and is now permitted for export to China with a 25% fee after recent US government approval.

Major Chinese tech firms, including Alibaba and ByteDance, have expressed interest in large quantities, but the Chinese government has not yet approved these orders.

Very limited numbers of H200 chips are in production, as Nvidia is also focusing on its newer Blackwell and Rubin chip lines.

Chinese officials are discussing whether to allow H200 imports, with one proposal involving bundling each purchase with domestic chips.

The H200 is significantly more powerful than Nvidia’s downgraded H20 chip for China, and current local alternatives, which has driven strong demand from Chinese cloud and enterprise buyers.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

U.S. approvals allow case-by-case H200 exports to China

Approvals apply to approved customers with U.S. security vetting rather than automatic clearance 1 2. Commerce is still setting details, which slows China deliveries even if Nvidia ramps output 3.
Exports face a 25% fee to the U.S. government on H200 shipments to China, not a blanket tariff 3 4.
The U.S. Department of Justice seized over $50 million in Nvidia technologies plus cash and charged several people in a scheme tied to at least $160 million of export-controlled H100 and H200 GPUs, which makes clear that unauthorized exports carry heavy criminal risk 5.

Cloud providers and OEMs can earn more as China approvals stall

H200 rental prices range from $3.72 to $10.60 per GPU (graphics processing unit)-hour 6. Providers with inventory can charge premiums while lead times stretch six to nine months, and Chinese reviews add more delay 3 4.
AMD’s MI308 AI accelerator received export clearance to China, yet about 75% of chips training AI models in Chinese data centers run on Nvidia’s CUDA platform (Nvidia’s software ecosystem for programming GPUs) 7. System integrators (third-party firms that design and deploy hardware with software systems) can win work by selling AMD migration help to firms stuck in H200 queues, including for inference workloads (running trained models in production) 4.
Compared with H100, H200 delivers 40–80% better performance for memory-intensive tasks, yet costs only 15–20% more 6. Buyers outside China should prebook 2026 H200 allocations now to fix pricing before supply tightens.