A Provably Cost-Efficient Approach to Deploying MoE Inference Models at the Network Edge

Foundation model inference is rapidly turning into a pay-per-request utility, yet today's infrastructure still deploys experts as dedicated, always-on services. In Mixture-of-Experts models, sparse activation makes per-request execution paths highly dynamic, which amplifies provisioning uncertainty while creating opportunities for Inner-Service Resource Sharing (ISRS) across requests. This paper studies cost-efficient provisioning of MoE inference as a service over geo-distributed edge infrastructure, where a provider must satisfy the Service Level Agreement (SLA) constraints while minimizing the joint cost of compute, VRAM, and bandwidth. We formalize the Mixture-of-Experts Deployment (MoED) problem by explicitly modeling ISRS under dynamic routing and multicast communication. To solve MoED, we design an Expert Cost Loss (ECL) metric and a Sparsity Aware Augmented Graph (SAG) that together guide ISRS-aware expert placement and bandwidth reservation, and we propose the ECL-based Expert Placement in SAG (ECL-EPS) algorithm. We prove that ECL-EPS achieves a logarithmic approximation ratio under practical assumptions. Extensive simulations show that ECL-EPS consistently reduces deployment cost and improves resource-sharing efficiency compared with the benchmarks.