According to The Information, Chinese officials have notified leading AI companies such as Alibaba, ByteDance, and DeepSeek that they may be granted a certain number of NVIDIA H200 chip purchase permits for training and running AI models; companies must first declare the quantity and purpose of the chips needed to obtain approval, and it is not an unlimited open policy. The maximum number of H200 chips approved for purchase in China this time is likely less than 200,000.
Citing two informed sources, The Information reports that Chinese AI companies applying to purchase NVIDIA H200 must first declare the quantity and purpose of the chips to gain approval. It is not open to everyone, nor unlimited.
The U.S. Department of Commerce approved about 10 Chinese companies to purchase H200 chips at the end of last year, with a cap of approximately 75,000 units per company. However, the initial application volume from Chinese companies was much higher than this ceiling. Beijing currently seems inclined to approve a total amount less than half of the applications, making this release more symbolic relaxation than a genuine supply boost.
According to reports, Beijing has historically been highly cautious about importing U.S.-designed AI chips, fearing that a flood of American processors would slow China’s long-term domestic chip development plans and raise security concerns, which caused H200 to be stuck at customs for over half a year.
However, the surge in computing power demand ultimately overrode these concerns: Chinese AI labs are suffering from insufficient computing resources, as the computational requirements for training next-generation large models continue to rise. Domestic chip capacity and performance cannot meet the short-term gap, and the urgency of the computing power shortage has overshadowed long-term goals of supporting local supply chains.
According to reports, H200 belongs to NVIDIA’s Hopper architecture and is the most powerful AI chip on the market before the release of Blackwell (expected at the end of 2024). NVIDIA is currently developing more advanced Rubin product lines, expected to launch in the second half of this year.
Due to national security concerns, Washington continues to restrict the export of advanced chips like Blackwell and Rubin to China. H200 is almost the ceiling of what China can currently purchase and NVIDIA is willing to sell. The approved purchase limit of less than 200,000 units for H200 contrasts sharply with a single data center’s consumption of over 400,000 Blackwell chips, illustrating that this release remains very limited in scale.
According to reports, NVIDIA CFO Colette Kress stated in May 2026, “We are not sure if these chips will actually be allowed for import,” and noted that H200 has not yet contributed any revenue in China.
This indicates that although the Trump administration approved NVIDIA’s sale of H200 to China in December 2025, the real obstacle has always been on China’s side. The number of applications from Chinese companies far exceeds the U.S. set limit, and Beijing’s approval decision had been delayed until now.
According to The Information, Chinese AI companies must first declare the quantity and purpose of the H200 chips needed to obtain purchase approval. The U.S. Department of Commerce previously approved a cap of about 75,000 units per company, and Beijing’s likely approved total may be less than half of the applications, with the purchase limit this time possibly under 200,000 units.
Reports indicate that Beijing previously hesitated to approve H200 imports due to concerns that U.S. chips would impact domestic chip industry development and security vulnerabilities. However, the rapid increase in computing power demand and the inability of domestic chip capacity to meet the shortfall ultimately made the urgency of the shortage outweigh the long-term protection of local supply chains.
According to reports, NVIDIA’s Blackwell and Rubin series chips are still prohibited from export to China by the U.S. government. H200 (Hopper architecture) is currently the latest generation of chips that China can purchase and NVIDIA is willing to sell.
Related News
TSMC plans to raise wafer prices by 5-10%, while new Samsung customers face a 15% increase.
OpenAI Receives U.S. Approval to Broadly Release GPT-5.6 Model Family
Trump administration lifts restrictions on OpenAI GPT-5.6, Commerce Department approves wide release
Zhipu considers self-developed GLM model to optimize AI chips, stock rises over 10%