The U.S. has announced it will approve NVIDIA to deliver its H200 chips to qualified Chinese clients, with 25% of the chip sales revenue to be paid to the U.S. government. The U.S. Department of Commerce is finalizing details, and the same approach will apply to AMD, Intel, and other U.S. companies.
The H200, launched in 2024, is built on TSMC’s 4nm process and features six layers of HBM, delivering total processing performance (TPP) nearly 10 times above current export control thresholds. - **Inference Performance**: The H200 offers price-performance in memory bandwidth comparable to the B300. - **Training Performance**: Its FP8 price-performance reaches 70% of the B300. - The H-series remains highly competitive, as most cutting-edge models recently released (e.g., Grok3, Llama4) were trained on Hopper chips.
The approval of H200 exports is unlikely to hinder China’s long-term goal of domestic chip substitution. Instead, it will primarily expand China’s total computing power supply rather than replace local production.
NVIDIA management previously stated that if geopolitical tensions ease and approvals are granted, quarterly revenue from H20 sales to China could reach $2–5 billion. Assuming a 73% gross margin and deducting the 25% revenue share to the U.S. government, the net gross margin for H200 sales would be around 64%.
This development is positive for Chinese cloud service providers (CSPs) like TENCENT, Alibaba, and ByteDance, boosting their capital expenditure (Capex) investments and accelerating AI application adoption.
Comments