LG U+ Reduced LLM Vendor Dependency and Cut Costs 36% with Databricks Model Serving

Overview

LG U+, one of Korea's leading telecom carriers, migrated its U+one AI Search service to Databricks Model Serving to reduce dependence on a single external LLM vendor. Databricks hosts Gemini 2.5 Flash in a dedicated availability zone through a partnership with Google, letting LG U+ keep its existing LangChain interfaces and model while removing its direct reliance on Google's own infrastructure. The company also rebuilt its search architecture with an AI Gateway, Query Orchestration, a Semantic Elastic Cache, Hybrid Reasoning Retrieval, and Agentic Self-Correction Loops running within an LLMOps framework. In controlled tests, Gemini calls routed through Databricks showed 48% higher throughput, 22% faster average response times, and a 90% lower failure rate versus direct calls; in production, operating costs fell 36% and response speed improved 64% even as system complexity grew 107%. LG U+ is extending the same infrastructure to its ixi-O Voice AI services.

The challenge

LG U+ had become heavily dependent on external large language models. With nearly every request riding on an external LLM call, token costs, response speed, output quality and uptime all sat in the hands of outside vendors, and its primary model, Gemini, still ran on a single vendor's infrastructure, so any latency spike or outage there landed straight on service quality.

The solution