What happened
LG Uplus revealed a suite of ‑optimization technologies aimed at reducing the operating costs of AI agents for business‑to‑business (B2B) customers. The company plans to launch the service in 2027 after internal tests showed significant cost savings.
LG Uplus said it will enter the B2B market in 2027 with a ‑optimization platform designed to lower operating costs. The platform combines three main techniques: (1) replacing widely used commercial models with an in‑house model for token‑heavy coding tasks, which the company says reduces costs by about 90%; (2) "model routing," which assesses task difficulty and directs simple queries to smaller, cheaper models while sending complex reasoning tasks to larger models, yielding internal test savings of over 25% for general questions and over 30% for coding agents; and (3) a "Chip‑Mix" hardware strategy that splits workloads between Nvidia and AMD GPUs and a Neural Processing Unit (NPU) from FuriosaAI, aiming to lower infrastructure expenses.
The company also reported technical improvements to its own on‑device AI model, shrinking its size from 2.4‑3 GB to roughly 800 MB while retaining more than 95% of the performance of the cloud counterpart. Additionally, LG Uplus claimed to have advanced , where a small model generates answers quickly and a larger model verifies them, more than doubling generation speed.
LG Uplus CTO Lee Sang‑yeop emphasized that the ‑optimization approach is intended to be a plug‑and‑play solution for enterprise services, positioning the company to lead the market in cost‑effective AI delivery once the B2B offering launches next year.
Source details: biz.chosun.com
Why it matters
The announcement matters because AI agents—especially coding assistants—consume large numbers of tokens, driving up inference costs for enterprises. By replacing commercial models with its own, routing tasks to appropriately sized models, and using a mixed‑chip hardware approach, LG Uplus claims to cut usage by up to 90% for coding tasks and overall infrastructure costs by more than 25‑30% for general queries. If the claims hold, enterprises could deploy AI agents at substantially lower expense, making large‑scale automation more financially viable and potentially accelerating AI adoption in sectors that have been cost‑constrained.
consumption is a primary driver of AI inference cost, especially for agents that perform repetitive searches or generate code. By cutting token usage dramatically, LG Uplus could make AI‑driven automation financially attractive for a broader range of enterprises, from software development firms to customer‑service operations.
The mixed‑chip strategy addresses a growing concern about hardware cost and supply chain dependence. Leveraging both GPUs and NPUs may provide a more resilient and cost‑effective compute stack, which could influence how other providers design their AI infrastructure.
If LG Uplus’s in‑house models match the quality of commercial alternatives while using fewer tokens, it could reduce reliance on external AI providers such as OpenAI or Google, reshaping competitive dynamics in the enterprise AI market.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Watch for LG Uplus’s rollout timeline, pricing details, and how enterprise customers integrate the ‑optimization stack with existing AI workflows. Also monitor whether other telecom or cloud providers adopt similar cost‑reduction architectures, and whether third‑party developers can access the in‑house models or must rely on LG‑managed services.
The exact pricing model and licensing terms for the ‑optimization suite, as these will determine adoption speed among cost‑sensitive enterprises.
Whether LG Uplus opens its in‑house models to external developers via APIs or keeps them proprietary, which will affect ecosystem integration.
Performance validation by independent third parties, especially regarding the claimed 90% reduction for coding tasks and the speed gains.
Potential partnerships or competition responses from major cloud AI providers who may introduce similar cost‑optimization features.