Location Verification for AI Chips

Advanced, general-purpose AI models hold immense potential for economic growth and scientific progress, with the capacity to fundamentally alter many fields and aspects of daily life in unexpected ways. These models are powered by cutting-edge AI chips produced by a select few companies that are subject to export control measures due to concerns about their misuse, yet current regulatory frameworks lack reliable mechanisms to track the deployment locations of these chips.

This paper is meant to serve as an initial introduction to location verification use-cases for AI chips, comparing different methods that could be utilized for location verification and briefly discussing both obstacles and requirements for creating a secure version of this solution.

• Adding location verification features to AI chips could unlock new governance mechanisms for regulators, help enforce existing and future export controls by deterring and catching smuggling attempts, and enable post-sale verification of chip locations.

• Delay-based schemes appear to be the most promising approach to location verification. Combining such a scheme with additional governance mechanisms, such as a centralized chip registry, would provide regulators with more effective tools to verify that chips are not deployed in restricted locations.

• Our main finding is that it seems both feasible and relatively cheap to implement pure-software delay-based solutions on chips in the near future. These solutions are likely to aid current AI chip export control enforcement efforts. A solution of this kind would likely cost less than $1,000,000 to set up and maintain for several years. Chip design companies should consider taking the initiative by experimenting with this technology on chips that are at a particularly high risk of diversion. More proactive use of technology to prevent the diversion of chips could bolster international trade by reducing the need for broad country-level export bans or license requirements.