'AI for All Project' to Balance Security and Speed: KT Forms Public-Private TF and Builds Its Own GPUs

[Financial News] KT plans to establish a task force with the Personal Information Protection Commission (PIPC) to ensure information security during the development of the government's 'AI for All Project.' To prepare for service delays caused by a surge in users, it has also secured additional graphics processing units (GPUs) of its own, separate from those supported by the government.
Lee Jin-hyung, head of KT's AX Business Division, held an online briefing on the Ministry of Science and ICT's 'AI for All Project' on the 11th. He emphasized, "We plan to establish and operate a task force with the PIPC to ensure information security during the development of the AI for All Project. We have also established a separate AI security organization within the AI Business Division and created an internal process for sharing security systems with companies participating in the consortium."
In preparation for a surge in users, KT has secured additional GPUs of its own in addition to those allocated by the government. The government plans to distribute KRW 170 billion in GPU support funds differentially among SK Telecom, KT, and Kakao based on their service performance. Lee said, "It is difficult to disclose the scale of the additional GPUs," but added, "We have a fallback mechanism that switches to GPUs owned by KT if a problem occurs."
Lee also unveiled KT's 'Ieum Inside' strategy during the briefing. The company plans to secure more than 60 million initial service touchpoints by combining a dedicated AI for All Project app with everyday services such as search, education, finance, shopping, real estate, and Internet Protocol television (IPTV).
KT's consortium includes Upstage, Saltlux, Daum, MUSINSA, Zigbang, Educational Broadcasting System (EBS), Mathpresso, and BC Card. Ieum Inside is an open ecosystem that embeds AI into each company's services and connects them through a single general-purpose AI agent.
KT is building a separate AI for All Project app and will also apply Ieum Inside to the services of Daum and other consortium participants, as well as Genie TV. Lee said, "If an Ieum Inside button is added to services such as Danawa, MUSINSA, and Zigbang, users will be able to access it directly from those services." He added, "We plan to provide a software development kit (SDK) that will allow Ieum Inside to be installed by companies that are not participating in the consortium as well."
The services of consortium participants have more than 60 million touchpoints in total. KT is also consulting with companies that have requested to join. Lee explained, "Having many partners participate in the ecosystem is a key objective of the Ieum Inside strategy. We will keep the door open and continue to welcome new partners."
The general-purpose service will combine search-based access to the latest information with multiple Korean AI models. It will also offer personalization features that remember search and inquiry histories. For example, on Zigbang, AI could analyze the actual transaction prices and listing information for apartments a user has viewed and identify listings priced below market value.
In the public sector, KT plans to recommend benefits based on age and place of residence, provide guidance on administrative procedures, and assist with issuing and applying for documents. However, government agencies must open their application programming interfaces (APIs) for AI to process applications through to completion. Lee said, "The agent can handle services within the scope that has been opened, but users must move to the relevant government app for services that have not been opened." He added, "We are discussing with the government how to open services that can be completed through the final application stage."
KT's AI platform, 'Token Factory,' will be used to operate the service. It will integrate Korean AI models, including Upstage Solar, Motif Technologies Motif, Saltlux LUXIA, NC AI VARCO, and KT Mi:dm, and connect each task to the model best suited for it.
Rebellions' neural processing units (NPUs), along with infrastructure optimization technologies from Lablup and VESSL AI, will also be applied. GPUs will be used for 'prefill,' which processes input prompts, while NPUs will be used for 'decode,' which generates responses. Lee added, "This can increase the number of tokens processed per minute by about 31% and reduce input tokens by up to 80%."
[email protected] Choi Hye-rim Reporter