AI Semiconductor Transitioning to 'Managed Trade'
The U.S. Department of Commerce adjusted policy from completely blocking advanced AI semiconductor exports to China to selective permission with conditions attached. From December 2025, advanced AI accelerators such as NVIDIA's H200 and AMD's MI325X become available for shipment only to 'verified customers' in China. End-user and purpose verification to prevent military and surveillance-only diversion, and re-export controls, remain in place.
This measure is closer to refinement rather than relaxation of export controls. The U.S. has been uniformly blocking advanced AI chips, but is interpreted as having judged that the method of also blocking commercial and civilian data center demand has revealed limitations in terms of market and technology standards. The approach is to manage 'who uses it, where, and for what purpose' rather than the chip itself as the control target.
Technical and industrial realities were also reflected. Generative AI and inference workloads are rapidly increasing even within China, and complete blocking could accelerate China's independent ecosystem while weakening U.S. companies' standard leadership and revenue. The H200 and MI325X are products that have established themselves as de facto standard accelerators of global AI infrastructure, and the U.S. government also appears to have judged it advantageous to maintain influence through controlled supply.
National security logic is also background to this decision. The United States, which has designated semiconductors as core infrastructure of military, AI, and data sovereignty, adopted a dual control method of blocking high-risk uses while permitting low-risk, verifiable uses. This is evaluated as a method transition to increase enforcement efficiency rather than a retreat from comprehensive prohibition to selective permission.
The key is the scope of 'verified customers.' Specific criteria have not been disclosed, but military and intelligence agency linkage, ultra-large-scale model training purposes, and surveillance and social control utilization possibilities are likely to be the core criteria. This is read as a signal to approach even within China by distinguishing between civilian company data centers and national strategy organizations.
This decision sends a clear message to the global industry as a whole. U.S. companies have regained some China revenue but have strengthened the role of quasi-public technology suppliers under government control, and China has retained access to cutting-edge AI chips but has acquired the structural constraint of always being conditional. For Korean, Taiwanese, and EU companies as well, the signal is conveyed that "technology can be sold, but only within American rules."
Ultimately, with this H200 and MI325X shipment approval, the U.S.-China technology relationship has entered a phase of 'managed connection' rather than comprehensive decoupling. AI semiconductors are no longer purely commercial goods but have become controlled trade assets whose rules are designed by nations. This measure is evaluated as an inflection point demonstrating that this order has actually begun to operate.
