How to Reach an AI Arms Control Deal With China

Council on Foreign Relations | Chris McGuire

United States AI laboratories currently maintain a six-to-twelve-month lead over Chinese competitors, but growing risks of uncontrolled frontier models are driving calls for a bilateral artificial intelligence safety agreement. Negotiating arms control while two unequal powers actively compete for strategic dominance remains historically unprecedented. Historical precedents like the 1987 Intermediate-Range Nuclear Forces Treaty demonstrate that bilateral technology agreements only succeed after the trailing state concludes it cannot win the race.

China currently produces just 1 to 4 percent of its required AI computing power natively, relying on smuggled and remotely accessed American microchips to stay competitive. Beijing must realize it cannot win. Consequently, Washington must aggressively close semiconductor export control loopholes to expand its operational lead prior to enacting domestic safety regulations. Maintaining a decisive technological advantage will ultimately enable unilateral technical verification of Chinese compliance via advanced national technical capabilities without compromising sensitive intelligence infrastructure.

Comment

Artificial intelligence capability scaling relies entirely on physical semiconductor manufacturing throughput and advanced packaging facilities like TSMC CoWoS lines. Chinese domestic foundry capabilities, exemplified by Semiconductor Manufacturing International Corporation, remain restricted to legacy deep ultraviolet lithography systems. This physical hardware bottleneck limits domestic datacenter expansion regardless of software optimization.

Advanced frontier model training demands high-bandwidth memory chips produced predominantly by SK Hynix and Samsung. Without these components, indigenous hardware architectures cannot achieve the interconnect bandwidth required for massive cluster scaling. Consequently, Chinese compute capacity stalls at the advanced packaging stage rather than at the architectural design phase, reinforcing reliance on smuggled Nvidia H100 units.

Strategic Question for Discussion

If American export restrictions successfully restrict Chinese access to Nvidia H100 accelerators and high-bandwidth memory, will domestic fabrication at SMIC adapt through architectural workarounds or succumb to physical hardware bottlenecks?

The available evidence points toward persistent hardware bottlenecks preventing Chinese datacenters from matching Western cluster scaling over the medium term. While algorithmic optimization and low-bit quantization offer temporary performance gains, they cannot fully offset hardware yield deficits at SMIC fabrication facilities. My assessment is that hardware scaling constraints will maintain a structural capability gap in frontier AI training regardless of software innovation.

Share your assessment in the comments below.

To read in details click here