Research Summary: Could Underwater Data Centers Pose a Risk to AI Treaty Verification?

Research examines the feasibility and detectability of underwater data centers (UDCs) as a potential method to evade international AI development restrictions. While power and cooling for UDCs appear tractable, significant obstacles exist regarding interconnectivity and the extensive hands-on maintenance required for large-scale AI training runs. These challenges suggest UDCs, at present, are not a practical evasion vector for frontier AI development, unless a highly advanced, automated capability exists to overcome these issues.

Why it matters

This analysis is critical for understanding potential vulnerabilities in future international agreements concerning frontier AI development, particularly regarding verification. It evaluates the technical viability of specific clandestine infrastructure, informing policy development around monitoring and compliance mechanisms for advanced AI compute capabilities.

Key insights

• International agreements limiting frontier AI development depend on effective verification mechanisms.
• Underwater data centers (UDCs) have been proposed as a potential evasion vector for undeclared compute facilities.
• Current UDC deployments were examined, alongside their construction and maintenance complexity compared to land-based facilities.
• A hypothetical 100,000 H100-equivalent training run in a UDC was analyzed for feasibility.
• Power delivery and cooling for such a UDC are considered tractable.
• Interconnect capabilities and the need for hands-on maintenance pose severe, potentially insurmountable, obstacles.
• Evasion via UDCs at a frontier scale is only surmountable by entities possessing extremely advanced and automated operational capabilities.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.18824