SK Telecom deploys proprietary AI model in steel, auto parts manufacturing

Manufacturing-focused AI agents piloted with KG Steel, Konex

A.X K1-based demo developed; field validation set for second half

SK Telecom employees review the company's proprietary AI foundation model, A.X K1. [SK Telecom]

SK Telecom will apply its proprietary AI foundation model to manufacturing sites, expanding its reach from defense into steel production and automotive parts processing as it accelerates its push into industry-specific AI transformation markets.

SK Telecom said Wednesday it signed separate MOUs with steelmaker KG Steel and automotive parts manufacturer Konex to pilot manufacturing-focused AI agents built on its proprietary AI foundation model. The agreements mark the first time SK Telecom's in-house AI foundation model has been applied to the manufacturing sector.

Since April, SK Telecom has been collecting process error and incident analysis reports, equipment manuals and log data from KG Steel and Konex to develop a demo version of an AI agent tailored to factory environments. The underlying model is SK Telecom's large language model A.X K1, which has 519 billion parameters in total but activates only about 33 billion during inference, enabling efficient deployment in industrial settings.

In the second half of the year, the company will apply the demo to the cold-rolling line for coated steel sheets at KG Steel's Dangjin factory and to Konex's casting and machining processes for on-site validation. KG Steel and Konex will supply additional process data, while SK Telecom plans to use feedback gathered during the pilot to improve the AI agent's response accuracy, inference speed and overall capabilities.

A diagram showing the collaboration structure for the manufacturing-focused AI agent. [SK Telecom]

Manufacturing has long been considered a difficult sector for AI adoption. Process data tends to be scattered across departments and equipment, and the expertise of skilled workers often stays with individuals rather than being systematically documented. SK Telecom believes that organizing manufacturing data as digital assets and deploying AI agents trained on that data can shorten error response times and improve process efficiency.

Security is another key consideration. Manufacturing sites face significant risks of sensitive process and equipment data leaking externally, which has constrained the adoption of cloud-based AI. SK Telecom's proprietary AI foundation model supports not only cloud deployment but also on-premises installation on a company's internal servers, letting manufacturers use AI without sending data outside their facilities.

SK Telecom also plans to use data gathered during the pilot to train its next model, A.X K2, currently under development. Once the pilot concludes, the company will assess whether to commercialize the manufacturing-focused AI agent and proceed with full deployment. It is also considering upgrading to subsequent models in its proprietary AI foundation model series as needed.

SK Telecom recently signed an MOU with the Ministry of National Defense to apply its proprietary AI foundation model to defense applications. The company intends to expand from defense and manufacturing into finance, the public sector and healthcare — industries where security and specialized expertise are essential.

SK Telecom's Euljiro headquarters building. [SK Telecom]

"For manufacturing sites where security is paramount, a proprietary AI foundation model that can be used without sending data outside is an effective solution," said Jeong Seok-geun, head of SK Telecom's AI CIC. "Starting with our partnerships with KG Steel and Konex, we will accelerate AI transformation in manufacturing and expand the use cases for our proprietary AI foundation model."

Bae Seon-woo, head of KG Steel's technology research institute, said the collaboration "lays the groundwork for AI adoption grounded in on-site data." Lee Gwang-pyo, chief executive of Konex, said rapid response to recurring quality issues on the factory floor has long been a challenge for the manufacturing industry, adding that the company would use AI to improve operational efficiency at its production sites.

Meanwhile, A.X K1 recently achieved the highest score among domestically developed AI models in a mathematical reasoning evaluation conducted by a KAIST research team using problems from the Korea Mathematical Olympiad. In the assessment, which drew on 20 questions from the most recent first-round KMO exam, A.X K1 recorded a single-answer accuracy rate of 83.75 percent, outperforming Upstage's Solar and LG Corp's AI research institute model Exaone.

rim@heraldcorp.com