Governing AI data-centre resources before consumption: the zero framework

Abstract

Abstract AI data-centre sustainability is usually evaluated after electricity, water and materials have already been committed. This Perspective develops ZERO as an operational governance framework that moves the decision boundary upstream through hard admissibility constraints, workload and resource-state separation, federated comparison, local vetoes and execution-time revalidation. To examine two testable properties with observed data, the revised paper adds a retrospective secondary-data evaluation across six U.S. Google data-centre locations. Public FY2024 site water reports and 2024 Google Cloud regional grid-carbon intensities were harmonised with 52 weekly county-level U.S. Drought Monitor observations, yielding 312 location-week records. A carbon-only spatial comparator selected the lowest-intensity region; ZERO then applied predeclared drought-coverage hold conditions before selecting the lowest-carbon remaining site. At county-area thresholds of 50%, 75% and 90% in moderate-to-exceptional drought (D1-D4), the carbon-only choice was held in 13, 6 and 3 of 52 weeks, respectively (25.0%, 11.5% and 5.8%). Observed state changes also reopened admissibility during the year. In the retrospective replay, observed drought conditions altered the admissible set before carbon ranking and triggered revalidation under the declared rules. The results do not establish universal drought thresholds, measured workload savings or operational superiority; those claims require operator telemetry and controlled deployment trials.