Infrastructure as Code (IaC) has significantly transformed infrastructure management by enabling automation, reproducibility, and scalability through machine-readable configuration. However, the manual authoring of IaC remains a complex and expertise-driven task, often limiting its accessibility to non-expert practitioners. This study investigates the potential of Large Language Models (LLMs) to automate the generation of Terraform configurations from natural language intent, thereby lowering the entry barrier to IaC adoption. We conduct an exploratory empirical evaluation using both open- and closed-source LLMs, applying various prompting strategies—zero-shot, few-shot and fine-tuning, to assess their effectiveness in generating syntactically correct and functionally valid infrastructure configurations. Our findings show that large closed-source LLMs demonstrate strong performance in zero-shot settings, while smaller open-source models can achieve comparable results when guided by few-shot examples or domain specific fine-tuning. These results highlight the adaptability of LLMs across different sizes and interaction paradigms.

Importantly, our study finds that LLMs not only assist in code generation but also substantially simplify deployment workflows translating high-level user intents into deployable configurations with minimal manual effort. This contributes to a broader shift in software engineering, where individuals without deep infrastructure expertise can take part in building and managing complex systems. Such accessibility paves the way for more inclusive and collaborative software development, aligned with the growing need to reimagine engineering practices around emerging AI capabilities. By enabling infrastructure creation that is faster, more intuitive, and less dependent on specialized knowledge, LLMs support the vision of a more democratized software development process. Further research is needed to enhance their reliability, integrate them effectively into deployment pipelines, and establish trusted human-in-the-loop validation practices to ensure robust, scalable adoption.All experimental materials, including code, prompts, datasets and outputs are publicly available to promote reproducibility: \url{https://github.com/llmiac-2024/llm-iac}.