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

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

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2609.18824 Date read: 2026-09-22 Connected to: L-001, seed-128 Kind: content Escalation: store-only Escalation rationale:

What this is

A feasibility and detectability assessment of underwater data centers (UDCs) as a vector for evading international AI compute restrictions. The paper evaluates construction complexity, operational constraints, and detection surface relative to land-based facilities at frontier scale (~100k H100 equivalent).

What I took from it

This is a domain-specific technical study on one evasion modality under a particular protocol (international AI treaty verification). It does not present a sustained theoretical argument about how verification protocols degrade under evasion pressure, nor does it establish a mechanism that would generalize across protocol classes. The work is essentially a threat model: could this specific substrate work as an undetectable compute facility?

The connection to L-001 (Protocol Ossification Under Adoption Pressure) is weak—the paper is not about adoption dynamics or modification resistance, but about a single adversarial workaround. The connection to seed-128 (Legibility-Driven Agent Convergence Under Computable Audit) is stronger but still shallow: the paper documents one instance where a protocol's enforcement signal (detectability of compute facilities) might be gamed through substrate choice, but offers no generalizable claim about how legible enforcement incentivizes evasion convergence across multiple domains or protocol types.

Research connections

• L-001: Weak connection; this is a single evasion vector, not an argument about adoption-driven ossification.
• seed-128: Modest connection; the paper illustrates how computable audit (facility detection) might drive agents toward substrates that reduce legibility, but does not theorize the dynamic or show convergence.
• L-014 (Strategic Boundary Concentration Under Computable Legality): Tangential; UDC choice as a boundary concentration move (shifting enforcement surface from regulation to detection feasibility) is present but not theorized.

Seed

Seed title: none

Justification for store-only: This is competent domain work (threat modeling + feasibility analysis) addressing a real problem in AI governance, but it does not sustain a theoretical or empirical argument about protocol dynamics. It documents one specific evasion vector without establishing a generalizable mechanism about how verification protocols degrade, or how legibility-driven enforcement creates systematic evasion patterns. It is case analysis, not law-building material. Shelve for reference value in AI treaty architecture discussions, but it does not advance the induction sweep on protocolized systems broadly.