CLOUDADV: Decision-Aligned Instance Sizing with Zero-Shot Foundation Models under Drift

Cloud virtual machines (VMs) are frequently overprovisioned, leading to significant financial waste and operational inefficiency. Traditional right-sizing methods often rely on static, percentile-based heuristics that fail to account for workload drift or future demand. This paper investigates whether zero-shot time-series foundation models can provide reliable, decision-aligned provisioning recommendations without the operational overhead of per-tenant model retraining and redeployment.

The authors introduce CLOUDADV, an interactive advisory system that integrates zero-shot forecasting with a bounded recommendation layer. The system constructs a structured decision context—incorporating historical utilization, forecast summaries, VM metadata, pricing, and explicit sizing heuristics—which is then processed by an LLM to generate actionable recommendations. The researchers evaluate the system using a deployment-aware pipeline, comparing a high-capacity reference model (Claude Opus) with a smaller production model (Qwen3.5-35B) to assess alignment under realistic latency and cost constraints.

In a case study of seven production VMs, CLOUDADV reduced simulated monthly costs from approximately $1,503 to$708, a 52.9% reduction. The system maintained high safety standards, with the highest observed ex-post exceedance rate among downgraded cases reaching only 1.5%. The zero-shot foundation models (Chronos-2 and TimesFM 2.5) performed competitively against supervised baselines, proving sufficient for decision-aligned provisioning. Furthermore, the smaller production model showed high agreement with the reference model, suggesting that deployable, lower-cost models can effectively mirror the decision-making quality of stronger, offline models.

This work shifts the focus of cloud capacity planning from raw forecast error minimization to decision alignment. By demonstrating that foundation models can operate effectively without task-specific retraining, the authors provide a scalable solution for managing non-stationary cloud workloads. This approach reduces the operational burden on engineering teams while enabling more dynamic and cost-efficient infrastructure management.