Risk-Weighted Compute Permit Markets under Imperfect Monitoring

Joel Christoph

Abstract: This paper studies the design of a tradable permit market for frontier AI compute under imperfect monitoring. A regulator sets an aggregate cap on risk-weighted compute; developers trade permits and must retire them in proportion to metered compute usage multiplied by evaluation-contingent risk multipliers. Because metering and reporting are noisy and strategically manipulable, the regulator conducts stochastic audits and imposes convex penalties for detected shortfalls. We characterize incentive-compatible audit and penalty pairs under enforcement budget constraints, establishing conditions under which truthful reporting constitutes a Bayesian Nash equilibrium. Under these conditions, competitive permit trading implements a second-best efficient allocation. We derive the optimal audit intensity and penalty schedule as functions of the permit price and detection technology, and compare equilibrium welfare with command-and-control alternatives, identifying parameter regimes where the permit market yields strictly higher surplus. An extension introduces insurance priced on verifiable model evaluations, showing that bundling insurance with permits improves compliance and allocative efficiency when auditing capacity is limited. The analysis connects cap-and-trade design, costly state verification, and the economics of AI governance.