Disaggregated inference is an operations story, too Most conversations about AI inference start with tokens per second and cost per token. Those metrics matter. But as inference architectures become more heterogeneous, delivered performance depends on