Article Information

Abstract

Background: The deployment of Large Language Models (LLMs) introduces substantial operational costs due to token-based pricing and unpredictable usage patterns. Organizations face cost increases of 200-400% without systematic optimization strategies.

Methods: We propose and evaluate a comprehensive three-layer optimization framework spanning Model Selection, Intelligent Proxy, and Infrastructure layers. Using the HELM benchmark dataset (50,000 queries) and real-world Azure OpenAI pricing, we conducted simulated experiments comparing our framework against baseline approaches across cost, accuracy, and latency metrics.

Results: The integrated framework achieved 68.4% cost reduction while maintaining 96.2% accuracy compared to GPT-4-only baseline. Individual layers contributed: Model Cascade (52.3% reduction), Semantic Caching (81.7% reduction at 89.3% hit rate), and Infrastructure Optimization (34.2% reduction). The Risk-Based Economic Evaluation framework demonstrated that for error costs greater than $0.10, deploying premium models remains economically optimal.

Conclusions: Sustainable AI cost management requires integrated optimization across all infrastructure layers, guided by value-based FinOps principles rather than absolute cost minimization. Our framework provides actionable strategies for organizations to achieve 40-70% cost reductions while preserving service quality.

Keywords

Cite

Citation: Saleh Dasin (2026) Cost-Optimized AI Infrastructure: A Three-Layer Framework for Large Language Model Deployment. Epistora J. Artif. Intell. & Intell. Syst. 1(1), 1-10. Article EJAIS-102