Compute Carbon Budgeting & Efficiency Monitoring

Control objective

Compute Carbon Budgeting & Efficiency Monitoring checks that each AI workload runs within a defined compute and carbon budget and that its energy/efficiency is actively monitored, rather than letting large extraction, RAG, or chat workloads consume unbounded compute. This matters because AI inference and indexing carry real cost and environmental impact, and unmonitored growth undermines sustainability commitments and resource-efficiency obligations. There is no numeric metric or formula supplied; the control is operated against defined budgets. To implement it, set a per-workload compute budget and a carbon-per-task ceiling, instrument the serving and batch pipelines to track compute consumption and derive carbon-per-task, and emit alerts as consumption approaches the budget. The threshold has three parts: a per-workload compute budget is set, an alert fires at 80% of that budget, and carbon-per-task stays within its ceiling. Crossing 80% should trigger proactive review (right-sizing models, batching, caching, or scheduling), and exceeding the compute budget or the carbon-per-task ceiling should trigger remediation such as throttling, optimisation, or escalation before the workload is allowed to continue scaling.