Delivering AI-powered entertainment that captivates users for hours

Overview

Scatter Lab used Databricks' flexible GPU clusters to build a proprietary Korean-language large language model in three months on a limited budget, powering its AI companion chatbot platform Zeta, which reached 1.5 million users within nine months with over 12 hours of average weekly engagement.

The challenge

As a startup, Scatter Lab lacked the GPUs, experience and know-how to train large language models, had a budget of less than 1 billion KRW compared with typical LLM training costs, and could not secure the 1,024 A100 GPUs it needed from cloud providers due to increased global demand, while an external LLM API could not meet its conversation-quality, data-security and cost-effectiveness requirements.

The solution

Scatter Lab adopted Databricks, which offered GPU clusters flexibly on an hourly basis with A100 and H100 options, provided LLM training hyperparameters optimized for data size and model scale, and ran a 24/7 monitoring system with automatic recovery, enabling Scatter Lab to build proprietary LLMs of various sizes in just three months.

Reported business value

Scatter Lab's Zeta platform surpassed 1 million user-created characters and reached 1.5 million users in Korea plus 150,000 in Japan, with users engaging for about 12 hours a week or more.