Generative AI Application Deployment on Microsoft Azure: Performance, Security, and Cost Considerations

Abstract

The rapid adoption of generative artificial intelligence (AI) has pushed organisations to look beyond experimentation and toward production-grade deployment on public cloud infrastructure. Microsoft Azure has emerged as one of the leading platforms for hosting generative AI applications through services such as Azure OpenAI Service, Azure Machine Learning, Azure Kubernetes Service, and Azure Container Apps. However, moving a generative AI workload from prototype to production surfaces three interconnected challenges: sustaining acceptable inference performance under variable load, protecting sensitive data and model assets against an expanding attack surface, and controlling the often unpredictable cost of large-model inference. This paper synthesises the current literature on cloud computing, large language model (LLM) serving, cloud security, and cloud financial management to propose an integrated framework for evaluating generative AI deployments on Azure. We describe a reference architecture spanning identity, orchestration, model serving, data grounding, and cost-and-operations layers; compare the performance and scalability profiles of five common Azure deployment options; and summarise seven cost-optimisation levers available to architects. The analysis shows that performance, security, and cost decisions are not independent: provisioning choices made to reduce latency (for example, reserved throughput) directly increase fixed cost, while security controls such as private networking and content filtering introduce measurable latency overhead. We conclude that a mature Azure-based generative AI deployment requires continuous, cross-functional governance—combining MLOps, FinOps, and security operations—rather than a one-time architectural decision, and we outline directions for future empirical research.