Root Cause Analysis (RCA) for service systems is critical for ensuring their reliability, while its application remains challenging because of the large number of metrics and the complex causal relationships. Classical RCA methods typically rely on statistical or rule-based approaches, making them difficult to generalize to unseen systems. The introduction of Large Language Models (LLMs) has partly addressed these challenges with their understanding of the domain-specific semantics of metrics and reasoning capabilities. However, they still struggle with incomplete or shallow reasoning when facing a large amount of metrics. To address these limitations, we present FoundRoot, a reinforcement learning (RL)-enhanced LLM foundation model for zero-shot RCA. FoundRoot features a novel structured deep thinking paradigm, which breaks down the RCA reasoning into several goal-oriented substeps, enhancing the completeness and reasoning capability of the causal relationship. We collect and curate diverse open-sourced RCA datasets across different systems and introduce a data augmentation technique to ensure data scalability. We design a two-stage training pipeline that includes supervised fine-tuning (SFT) and RL to align the LLM with the structured deep thinking paradigm, which significantly improves its reasoning quality. Extensive experiments on four datasets show that FoundRoot outperforms both classical and LLM-based methods in RCA accuracy on unseen systems, achieving 2.9%-42% mean reciprocal rank (MRR) improvements.