{
  "researchId": "JPC-26-00139",
  "researchUrl": "https://jipcet.org/r/JPC-26-00139",
  "articleUrl": "https://jipcet.org/articles/jipcet-2026-0139",
  "articleRef": "JIPCET-2026-0139",
  "externalDoi": null,
  "title": "Latency-Aware Cost Models for Serverless Inference at the Network Edge: Derivation and Multi-Platform Validation",
  "authors": [
    {
      "name": "Michelle Tan",
      "affiliation": "Universiti Sains Malaysia, Malaysia",
      "orcid": null
    },
    {
      "name": "Joseph Haddad",
      "affiliation": "University of Calgary, Canada",
      "orcid": null
    },
    {
      "name": "Peter Osei",
      "affiliation": "Drexel University, USA",
      "orcid": null
    },
    {
      "name": "Aisha Rahman",
      "affiliation": "Universiti Sains Malaysia, Malaysia",
      "orcid": null
    }
  ],
  "abstract": "Background. Cost models developed for centralised cloud inference assume warm capacity and amortised residency, assumptions that fail at the network edge where cold starts, short-lived instances and per-location traffic sparsity dominate spend. Objective. We derive a latency-aware analytical cost model for serverless inference at the edge, parameterised by arrival rate, batch window, residency policy and cold-start distribution, and evaluate its predictive accuracy against commercial billing. Methods. We formalise expected cost per completed request as a function of these four parameters under a Markov-modulated Poisson arrival process, and derive the batch window that minimises cost subject to a p95 latency constraint. We validated the model on three commercial serverless platforms across 14 edge locations over a four-week window, comprising 41.6 million requests and independently reconciled invoices, plus a controlled sweep of 1,320 configurations varying arrival rate, batch window and residency. Results. Predicted cost tracked billed cost within 6.0% mean absolute percentage error on all three platforms and within 9.1% in the sparsest quartile of locations. Modest batching — windows of 40–80 ms at our measured arrival rates — reduced per-token expense by 22% at a cost of 8% throughput and 31 ms added p95 latency. Cold starts accounted for 34% of total spend at locations below 4 requests per second, but under 6% above 40 requests per second, making residency policy the dominant lever only in sparse deployments. A closed-form approximation stayed within 11% of the full model and requires only three observable parameters. Conclusion. Edge inference cost is governed by residency and batching interactions that centralised models do not represent; the model presented here supports capacity and placement decisions using parameters operators already collect.",
  "keywords": [
    "serverless computing",
    "edge computing",
    "machine learning inference",
    "cost modelling",
    "queueing theory",
    "cold start latency"
  ],
  "area": "Computing & Software",
  "issue": "JIPCET 2026; 3(2)",
  "pageRange": "73–104",
  "published": "2026-07-01",
  "received": "Received 12 Dec 2025",
  "version": "1.0",
  "versions": [
    {
      "version": "1.0",
      "date": "2026-07-01",
      "note": "Version of record published after editorial decision."
    }
  ],
  "articleType": "Research Article",
  "reviewStatus": "Peer reviewed under named human editorial supervision",
  "peerReviewed": true,
  "handlingRecord": "Handled by a named JIPCET editor of record, who read the manuscript, oversaw review and approved the decision.",
  "methodology": "Research Article",
  "dataAvailability": "Available from the corresponding author on reasonable request, in line with participant and organisational agreements.",
  "codeAvailability": "Analysis procedures are documented in full in the methods section of the article.",
  "referenceCount": 6,
  "references": [
    "Haddad, J. (2025). Cold starts revisited: initialisation tails in inference workloads. Proceedings of SoCC, 311–324.",
    "Osei, P., Tan, M. & Rahman, A. (2024). Residency policies for serverless inference. ACM Transactions on Computer Systems 42(3), 1–27.",
    "Tan, M. (2026). Time-window batching under sparse arrivals. JIPCET 2(4), 91–118.",
    "Rahman, A. (2026). Evaluation cost models for multi-step inference workloads. JIPCET 1(1), 12–29.",
    "Marsh, S. (2025). Markov-modulated arrivals in edge deployments. Performance Evaluation 168, 102–121.",
    "Whitfield, B. & Russo, G. (2025). Reconciling platform billing with request-level accounting. Proceedings of ICPE, 74–86."
  ],
  "replications": "No independent replication recorded to date.",
  "corrections": "No corrections issued. Any correction is published as a new version with a dated entry in the version history.",
  "license": "CC BY 4.0",
  "fullTextHtml": "https://jipcet.org/articles/jipcet-2026-0139",
  "fullTextPdf": null,
  "publisher": {
    "name": "JIPCET Editorial Office",
    "url": "https://jipcet.org",
    "email": "editor@jipcet.org"
  },
  "journal": {
    "title": "Journal of Innovation, Product, Computing & Emerging Technologies",
    "short": "JIPCET"
  }
}