AI Non-Proliferation & Compute Verification Lab

Evaluate verifiable arms-control pacts for frontier AI. Model detection thresholds, clandestine cluster breakouts, hardware telemetry, power surveillance, and treaty stability between rival superpowers.

Treaty Architecture

Protocol v2.4

Preset Diplomatic Framework

Compute Ceiling (FLOPs) 10²⁶ FLOP Frontier threshold requiring international registry and continuous tracking.

Rival Clandestine Training Goal 3.2 × 10²⁶ Target scale of secret training run attempted by defecting party.

Verification Regimes Implemented

Hardware Roots of Trust (On-Die HSM)

Cryptographic signing of cluster interconnects & firmware-level throttle locks.

High-Voltage Grid & Satellite Thermal

Detects >50MW continuous unlogged electricity consumption and IR signatures.

Challenge Inspections (Short-Notice)

Bilateral team access to suspect high-density datacenters within 72 hours.

Photolithography Chokepoint (EUV Ledger)

Bilateral auditing of EUV mirror wear, laser pulses, and bare-die serial batches.

Geopolitical Trust & Whistleblower Access 45% Likelihood of insider leaks or intelligence sharing before run completes.

Ready to evaluate treaty stability.

Verification Telemetry & Detection Horizon

Treaty Viable

Breakout Detection Prob. 91.4% Before completion

Mean Time to Detection 32 Days Run length: 120 Days

False Alarm Hazard 4.2% / yr Risk of false treaty breach

Breakout Compute Gap +0.50 OOM Excess over legal threshold

Clandestine Training Run vs. Surveillance Mesh

DAY 120 / FINISHED

Clandestine Compute Accumulation

Legal Treaty Threshold

Multi-Vector Sensor Cross-Check

First Breach Interception

Surveillance Vector Contribution

Surveillance Vector	Status	Evasion Difficulty	Sovereignty Cost	Marginal Signal

Strategic Assessment & Policy Viability

Under this configuration, clandestine compute aggregation above 10²⁶ FLOP is caught with high confidence before the training run reaches completion. The combination of hardware-level telemetry and power monitoring forces a defector to disperse training across thousands of micro-datacenters, suffering a 4.8× latency penalty that makes frontier model emergence impractical within treaty response windows.

Why AI Treaties Differ from Nuclear Pacts

Unlike enriched uranium (HEU) or plutonium-239, which emit characteristic gamma/neutron signatures and require gargantuan centrifuge cascades, AI weights are intangible bitstrings that can be duplicated across optical fiber in seconds.

Verification cannot occur at the software layer without violating privacy and commercial IP. Consequently, viable pacts focus exclusively on the physical compute stack: wafer fabs, power substations, and high-bandwidth interconnects (NVLink/optical transceivers).

Key Chokepoints in the Compute Supply Chain

Lithography Hardware: Only one firm (ASML) manufactures High-NA Extreme Ultraviolet (EUV) scanners. Every machine contains cryptographically signed mirrors and maintenance logs.
Mega-Datacenter Power: A 10²⁷ FLOP cluster requires 80–150 MW of continuous baseload electricity, creating distinct substation heat blooms visible to commercial infrared satellites.
High-Bandwidth Memory (HBM): Secret distributed clusters face insurmountable latency bottlenecks unless interconnected with dedicated fiber trunks.

Resolving the Verification Dilemma

The Economist's core dilemma highlights that neither Beijing nor Washington will accept intrusive on-site code inspection. However, zero-knowledge hardware telemetry allows servers to verify compliance mathematically without revealing weights, dataset composition, or proprietary architectures.

Use this simulator to identify the minimum viable verification regime that preserves national sovereignty while making clandestine breakouts mathematically irrational.