"Lower Total Token Costs"... LG Uplus Targets Corporate Customers with Token Optimization

Cutting AI Coding Costs by Up to 90%
Model Routing and Compression Improve Processing Efficiency
Commercialization Built on ixi-O and Internal AX

LG Uplus will target the business-to-business (B2B) AI market next year with its "token optimization" technology, which can reduce the operating costs of AI agents by up to more than 90%. Rather than competing to lower the price per token, the company aims to reduce the total cost of completing work with AI.

Lee Sangyeop, CTO and executive vice president of LG Uplus, explained "AI token optimization" on the 12th, saying it could reduce AI agent operating costs by up to more than 90%. Courtesy of LG Uplus.

At a press briefing on the 12th, LG Uplus said it plans to apply experience accumulated through the operation of actual services—including its AI call agent "ixi-O," its AI contact center (AICC), and its internal AI transformation—to its token optimization business. For "AI for Everyone," a service developed in collaboration with Kakao and scheduled to launch at the end of this year, it will manage AI model operations and strengthen its capabilities in operating large-scale services.

LG Uplus highlighted the AI model operation recipes it has developed through managing large volumes of traffic and adapting to customers' usage patterns and work requirements. For example, at launch, ixi-O used both an on-device model from LG AI Research and Google's Gemini, but practical tuning helped improve service quality and cost efficiency. The approach assigns tasks requiring high-performance AI and simpler tasks to different models.

For its internal AI transformation (AX), LG Uplus is also using its own AI models in AI coding tools used by developers, reducing costs while maintaining the existing work environment.

The company introduced "model routing" as a key technology for this approach. It automatically assigns an appropriate AI model based on a user's question: a lightweight model handles simple questions, while a high-performance model handles complex reasoning. In particular, in coding, it cut operating costs by 90% compared with other models. LG Uplus said its internal technical validation showed that it reduced the cost of handling general questions by 25% and the cost of coding agents by more than 30%.

The company has also made progress in model-lightening technology that makes AI models themselves smaller. It reduced the size of existing on-device AI models, which ranged from 2.4 to 3 GB, to about 800 MB while achieving performance more than 95% that of the cloud models used for comparison. Its "speculative decoding" technology, in which a small model quickly generates a response and a larger model verifies it, more than doubled token generation speed. Two papers based on these research results were accepted at NeurIPS, one of the world's most prestigious AI conferences.

At the infrastructure level, LG Uplus is pursuing a strategy of combining chips according to the characteristics of each task. Its architecture distributes tasks across graphics processing units (GPUs) from NVIDIA and AMD and neural processing units (NPUs) from FuriosaAI. Through partnership-based models with partners and the development of its own infrastructure, LG Uplus plans to launch a token business next year to help B2B corporate customers address their AI costs.

Lee Sangyeop, CTO of LG Uplus, said, "By combining globally competitive model-lightening and inference technologies with operational recipes that reflect the execution pathways of AI tasks, we will lead the market from next year, enabling corporate customers to achieve the best AI results at the most reasonable cost."

This content was produced with the assistance of AI translation services.