Generative decision intelligence: leveraging large language models for autonomous multi-objective resource orchestration in dynamic cloud environments

With the expansion of hyperscale datacenters, optimal resource allocation (RA) requires simultaneously balancing conflicting objectives: resource utilization, cloud service provider (CSP) net profit, customer satisfaction (QoS), scheduling latency, and load balancing across physical machines. Traditional meta-heuristics encounter superlinear computational bottlenecks under high workload volumes, while numerical machine learning models lack semantic awareness to interpret high-level Service Level Agreements (SLAs) or adapt to runtime workload shifts. To bridge this gap, this paper introduces a Generative Decision Intelligence (GDI) framework based on a dual-layer neuro-symbolic architecture. The Strategic Reasoning Layer employs a Large Language Model (LLM) that translates natural-language SLAs and runtime telemetry into dynamic fuzzy logic boundaries and objective trade-off weights. These parameters configure the Tactical Execution Layer, which comprises high-speed machine learning models (Random Forest, Deep Neural Networks) and adaptive fuzzy inference systems that assign tasks in real time while tracking resource releases upon task completion. Extensive experiments on 1,000 physical machines processing 100,000 dynamic requests from Google Cluster Workload Traces show that GDI achieves a 97.1% request acceptance rate, 81.5% average infrastructure utilization, and reduces load standard deviation to 7.5. An ablation study demonstrates that LLM-guided rule distillation yields an independent 3.7% acceptance gain and 20.2% lower load variance over static expert rules. Model inference runs in O (1) time, yielding an overall per-request allocation complexity of O ( N ) across N physical machines.