SORT-AI: Interconnect Stability and Cost per Performance in Large-Scale AI Infrastructure—A Structural Analysis of Runtime Instability in Distributed Systems

The continued scaling of large-scale AI and HPC systems increasingly encounters limits that are not imposed by raw compute capacity, but by the dynamics of interconnects that bind distributed components into a single execution fabric. As system size, heterogeneity, and synchronization demands grow, performance degradation manifests in non-linear and often opaque ways, leading to a collapse of effective cost per performance despite sustained investment in additional hardware. Classical performance and network metrics, while necessary, fail to capture the structural origins of these effects and therefore provide limited guidance for architectural or economic decision-making. This article argues that interconnect-induced instability should not be understood as a collection of incidental faults or implementation bugs, but as an emergent structural property of tightly coupled, large-scale runtime systems. We analyze how latency drift, synchronization loss, and non-local coupling effects propagate through operator dependencies and give rise to hidden economic costs, including re-runs, over-provisioning, and diminished result usability. The contribution of this work is a structural problem analysis that reframes stability as a first-order economic variable rather than a secondary performance artifact. The methodology is deliberately conceptual and analytical, avoiding implementation details or prescriptive solutions. By isolating the structural mechanisms underlying cost-per-performance collapse, this analysis establishes a foundation for structure-oriented approaches to runtime stability control in AI and HPC infrastructures.