The rapid advancement of AI has created a significant blind spot: while individual advanced chips are regulated, governments currently lack visibility into how these chips are combined into powerful clusters capable of training dangerous AI models. This gap makes it difficult to monitor and control potential risks from uncontrolled AI development.

A Two-Part Solution

One approach to address this could involve:

1. Chip Tracking: Creating a registry of all high-performance chips above a certain capability threshold, recording their ownership, location, and specifications. This would build upon existing supply chain tracking used by manufacturers.
2. Cluster Licensing: Requiring authorization to assemble large clusters of these chips, with "large" defined by aggregate compute capacity. The licensing process could provide visibility into frontier compute accumulation and enable risk assessment.

Implementation Strategy

A phased approach might work:

• Start with voluntary reporting from major chip manufacturers
• Develop prototype tracking systems
• Gradually implement mandatory reporting and licensing
• Focus initially on cloud computing environments where monitoring infrastructure exists

This could begin unilaterally in chip-producing nations, then expand internationally through diplomatic coordination.

Balancing Needs

The system would need to carefully balance oversight with supporting legitimate research. Some ways this might be achieved include:

• Exempting small-scale research clusters
• Using measurable performance metrics that can adapt as technology advances
• Minimizing unnecessary burdens on approved AI development

While not a complete solution, this approach could provide governments and safety organizations with crucial visibility into potential sources of uncontrolled AI development while creating accountability for high-power compute clusters.