Kakao, serving 93% of South Korea's population on its messaging platform, faced scalability issues with its existing infrastructure. The engineering team strategically transitioned to Google Cloud TPUs using the JAX framework to optimize for cost and efficiency. This move enabled the development of their proprietary Kanana model family, with several models now open-sourced. The team detailed their technical journey, including adapting JAX, MaxText for custom data pipelines, and Mixture-of-Experts (MoE) model training. They customized MaxText to blend multiple data sources dynamically and modified token processing for efficiency. Validation experiments confirmed TPU performance matched their GPU-based pipeline. Further experiments involved upcycling a dense model into an MoE architecture on v5e TPUs, demonstrating the JAX stack's suitability for advanced research. Kakao observed a significant throughput increase and improved cost-performance with Trillium TPUs. Key advantages of the JAX stack include performance, scalability, customizability, and rapid feature adoption. The modular design of the JAX ecosystem, including MaxText, Flax, Optax, and Orbax, facilitates both production pipelines and advanced research.