LG Uplus has secured technology to reduce the operating costs of AI coding tools by up to 90% and is set to enter the corporate AI cost-reduction business starting next year. The company aims to lower the burden of AI adoption for enterprises by reducing the size of AI models and utilizing different models and semiconductors depending on the task.
On the 12th, LG Uplus held a reporter study session related to “AI Token Optimization,” disclosing its own technology verification results and plans for commercialization in 2027. A token is the basic unit in which AI processes text, and service operating costs increase as usage grows.
According to LG Uplus’s internal verification, utilizing its proprietary AI model in the coding domain can reduce operating costs by approximately 90% compared to commonly used models. The company also confirmed results showing that improving the computational processing method of AI models increased processing efficiency by up to four times on the same infrastructure.
The company has also developed “model routing” technology, which selects the model suitable for a task among various AI models. In internal tests, this reduced general question processing costs by more than 25% and coding agent costs by more than 30%. The method involves using high-performance models for complex tasks and assigning simpler tasks to relatively lighter models.
LG Uplus is applying these optimization technologies to its AI voice agent “Exio,” AI Contact Center (AICC), and internal business automation. In the case of Exio, the company initially operated LG AI Research’s on-device model alongside Google Gemini, but later adjusted the usage ratio of its proprietary model and Exaone to improve cost efficiency. LG Uplus is also responsible for operating AI models in “Mudou-ui AI,” a collaborative service with Kakao that is being commercialized by year-end.
Research to lighten the AI models themselves is also underway. The company reduced the size of its proprietary on-device AI model from the previous 2.4–3GB to approximately 800MB while securing performance exceeding 95% compared to benchmark cloud models. Through technology where a small model generates answers and a large model verifies them, token generation speed was also increased by more than two times. Two related research papers were accepted at the global AI academic conference “NeurIPS.”
AI semiconductors will also be diversified. The company is verifying options to utilize not only GPUs (Graphics Processing Units) from NVIDIA and AMD, but also NPUs (Neural Processing Units) from FuriosaAI. This is to distribute computations across different semiconductors according to task characteristics in order to lower infrastructure costs.
Starting next year, LG Uplus plans to promote a corporate AI cost-reduction business through partnerships with collaborators and the construction of its own infrastructure. The goal is not merely to lower the price per token, but to reduce the total cost required for AI to complete a single task.
Lee Sang-yeop, CTO (Chief Technology Officer) of LG Uplus, explained, “The hegemony in the AI market has shifted beyond who can build larger models to how precisely and economically one can achieve the business results customers desire.” He added, “As AI agents spread, even if the unit price of models is cheap, repeating unnecessary searches and analyses causes total costs to snowball.”
in TECH
