Telecom

LGU+ Cuts AI Coding Model Operating Costs by 90% ... to Launch B2B Business Next Year

IT DAILY ·

LGU+ Cuts AI Coding Model Operating Costs by 90% ... to Launch B2B Business Next Year
Lee Sang-yeop, CTO and executive vice president of LG Uplus, is presenting technologies for reducing AI operating costs and the commercialization plan at the "AI Token Optimization Study" held on the 12th. [Photo: LG Uplus]

✦ AI Summary

LG Uplus said it has verified cost-cutting effects through AI model operation technologies accumulated from actual service operations.

In coding tasks, using its own AI model lowered operating costs by about 90%, and applying model routing cut general question handling costs by more than 25% and coding agent costs by more than 30%.

LG Uplus plans to pursue a B2B AI cost optimization business in 2027 and to roll out a business for corporate clients by bundling AI model operation, inference, and infrastructure optimization technologies.

BizCrushWant to follow Korean talks and live streams in real time?Try CapCrush on the web

LG Uplus said it has verified cost-cutting effects based on AI model operation technologies accumulated through actual service operations. The basis for the cost reduction was model operation know-how accumulated through real services such as call AI and AICC. On the 15th, LG Uplus announced that it had disclosed its AI model operating cost reduction technology and commercialization plan.

At an 'AI Token Optimization Study' for reporters on the 12th, LG Uplus CTO Lee Sang-yeop unveiled the company's AI operating cost reduction technology and commercialization plan. LG Uplus said its own verification found that the cost reduction effect varied depending on the applied technology and the type of work.

Costs fell as the company used its own models suited to the nature of the work and applied model routing. In coding tasks, using the company's own AI model lowered operating costs by about 90% compared with commonly used models. LG Uplus secured a case in which coding task operating costs were cut by about 90% by using its own AI model. In internal tests, it applied 'model routing' technology that selects the AI model best suited to the task. As a result, the cost of handling general questions fell by more than 25%, and the cost of coding agents fell by more than 30%.

Based on this, LG Uplus plans to pursue a B2B AI cost optimization business in 2027.

The company sees reducing the total cost involved in completing a task as more important than the usage fee of an individual AI model in AI operations. It believes that if AI agents repeatedly conduct unnecessary searches and analyses, total costs can rise even when inexpensive models are used. Accordingly, it is developing technologies that deploy different models depending on the difficulty and characteristics of each task and reduce duplicate searches and verification.

This approach was shaped by experience operating the AI calling agent ixi-O, AICC, and internal AI services for work. At the beginning of the ixi-O launch, LG AI Research's on-device model and Google Gemini were used in parallel. Later, the company adjusted the share of its own AI model and EXAONE, saying that this improved service quality and cost efficiency.

At the same time, the company also secured technologies to lighten models and improve inference efficiency. LG Uplus said it reduced the size of its own on-device AI model from 2.4-3 GB to about 800 MB. It also said it secured performance of more than 95% compared with the benchmark cloud model.

LG Uplus said it applied an inference technology that separates answer generation from verification to improve performance. In this approach, a smaller model generates a preliminary answer and a larger model verifies it. LG Uplus presented its own verification results showing that, by combining model operation, inference, and infrastructure optimization technologies and using this approach, token generation speed more than doubled and computational processing efficiency improved by up to four times on the same infrastructure. As related research results, two papers were accepted at NeurIPS.

LG Uplus is also pursuing cost reductions by combining and operating AI semiconductors according to task characteristics and by classifying them by task type. The target semiconductors are NVIDIA GPU, AMD GPU, and FuriosaAI NPU. Tasks are divided into those requiring large-scale computation and those in which memory bandwidth is important, and the company plans to reduce infrastructure costs by distributing processing to the semiconductors best suited to each task. In this regard, it is conducting verification of FuriosaAI NPU, and it has been working with FuriosaAI for more than two years.

LG Uplus plans to bundle these optimization technologies and pursue an AI cost optimization business for corporate clients starting in 2027. As business methods, it plans to use partnership-based models with allies and build its own infrastructure.

Lee Sang-yeop, CTO of LG Uplus, said the standard for leading the AI market is changing, and that competition to build larger models is weakening. He explained that future advantages will depend on how precisely and economically companies can achieve the business outcomes customers want.

Source: IT DAILY · Kim Byeong-ju
Original: https://www.itdaily.kr/news/articleView.html?idxno=242924

References

This article was produced with the help of an automated content generation algorithm.


Source: IT DAILY

View original

This article was summarized and organized by BizCrush based on the original article from IT DAILY. For exact quotations and full details, please refer to the original article.

Keep reading

BizCrush
BizCrush showing live captions over an online meeting

BizCrush app

Live captions and summaries for meetings and calls

BizCrush Inc. · CEO Taemin Kwak

HQ 100 Middlefield Rd, Menlo Park, CA 94025, USA

Korea office Z112, Room 602, 6F, 602 Yeongdong-daero, Gangnam-gu, Seoul, Republic of Korea

Support help@bizcrush.ai · Partnerships business@bizcrush.ai

AI Notice

Some features use generative AI.

Those features carry their own pre-use notice and mark AI-generated results. AI output is for reference only — always check the original source and consult an expert before making an important decision.