Key facts
- China's AI compute capacity is expected to double by 2027 to 8.8 million H100-equivalents.
- US-designed chips will account for approximately 65% of China's AI compute capacity in 2027.
- Chinese firms are renting advanced Nvidia chips remotely via data centers in countries like Thailand, Malaysia, and Japan.
- The current US export-control regime focuses on physical chips, not remote cloud access.
- The Remote Access Security Act (RASA) aims to close this loophole but is pending in the Senate.
China's artificial intelligence computing power is projected to nearly double by 2027, reaching approximately 8.8 million H100-equivalents. However, a significant portion, around 65%, will still rely on chips designed in the United States, highlighting a persistent dependence on foreign hardware. Chinese AI firms are circumventing US export controls by accessing advanced Nvidia chips, such as the GB300, through remote cloud services hosted in data centers in countries like Thailand, Malaysia, and Japan. This loophole allows them to utilize powerful hardware without physically importing it, contributing to rapid advancements in Chinese AI models.