Serverless Solution Architecture for AI-Powered Cloud-Native Applications

Abstract

Serverless computing promises to run AI-powered, cloud-native applications with no idle cost and automatic, event-driven scaling, but the serverless model collides with the realities of modern AI workloads: large models, accelerator dependence, and multi-second cold starts. This paper presents a solution architecture for building AI applications on serverless primitives. We organize the system into four event-driven layers — application, orchestration, AI services, and data/state — in which loosely coupled functions and managed services are triggered by events and scale independently from zero. We analyze the central tension: cold-start latency, which for large models can reach tens of seconds to minutes, versus the scale-to-zero economics that make serverless attractive. We give a cost model that exposes the crossover beyond which reserved capacity is cheaper than pay-per-use, a cold-start latency model, and a survey of mitigation strategies (quantization, snapshotting/pre-loading, and provisioned concurrency) with their latency–cost trade-offs. Illustrative results, consistent with reported figures, show that an event-driven serverless design can serve bursty AI workloads at near-zero idle cost while bounding tail latency. Public sources are cited throughout.