Most governance focuses on training compute, but inference-time scaling can unlock big capability jumps from existing models. Is inference scaling an underappreciated overhang that breaks compute-based governance?