Naver Cloud and LG CNS consortium has preemptively begun pre-training its model by deploying 4,000 of its own graphics processing units (GPUs), ahead of the final contractor selection for the South Korean government’s cybersecurity-specialized artificial intelligence (AI) foundation model development project. The strategy is to secure a head start in training time by self-procuring computing resources approximately 16 times the scale the government plans to provide to the ultimately selected consortium.
According to industry sources on the 1st, the Naver Cloud–LG CNS consortium has initiated pre-training using 4,000 Nvidia B200 GPUs as training resources, even before being confirmed as the final contractor in the cybersecurity segment of the “Specialized AI Foundation Model Development Project” promoted by South Korea’s Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA). This is approximately 16 times the 256 B200 GPUs the government will provide to the selected consortium over a 10-month period.
The consortium’s decision to activate its own computing infrastructure without waiting for the official project start date—when government-supported GPUs would be supplied—stems from the need to complete a global-standard security-specialized model within the constrained project timeline. Deploying large-scale proprietary GPUs before final contractor confirmation is considered an unusual move in the industry.
Once government-supported resources are added, the consortium plans to build on the pre-training already underway to continue security domain-specific fine-tuning and performance optimization. Rather than starting from scratch with foundational training, the approach layers security-specialized processes on top of the model already in training, concentrating the head start in training time on performance enhancement.
The consortium has set internationally recognized benchmarks—not self-assessments—as quantitative performance targets, and has established a policy of developing a frontier-class security-specialized model based on these standards. The developed model is slated for use in AI transformation across South Korea’s security industry, including cyber threat analysis, vulnerability detection, and incident response.
The bidding has effectively narrowed to a two-way contest. Following the call for proposals closing on the 26th of last month, two consortia submitted proposals: one led by Naver Cloud–LG CNS and the other by SK Telecom–Upstage. Both sides have completed submission of evaluation materials and face on-site presentation evaluations on the 2nd. Final selection results are expected to be announced in September.
The SK Telecom consortium includes Upstage, SK Shieldus, SecuI, AhnLab, Igloo Corporation, RaonSecure, Genians, the Korea Information Security Industry Association (KISIA), and industry-academia cooperation foundations at Korea University and Soongsil University. They proposed “development of an all-encompassing cybersecurity-specialized AI foundation model suite and demonstration of agentic security services based on South Korea’s cyber threat environment” as their primary objective.
The Naver Cloud consortium has confirmed participation from LG CNS, the LG Management Development Institute’s AI Research Center, LG Uplus, Logpresso, MarkAny, S2W, East Security, Jiransecurity, Korea Hydro & Nuclear Power, KEPCO KDN, the Financial Security Institute, the Korea Institute of Science and Technology Information, and industry-academia cooperation foundations at Korea University, Seoul National University, Chung-Ang University, and Pohang University of Science and Technology. They have put forward “development of an AI foundation model for internalizing national cybersecurity capabilities” as their primary objective, highlighting model capabilities including LG’s “ExaOne” and the training data accumulated by participating security companies as key strengths.
An industry insider commented, “There are only a handful of operators in South Korea who can attach a 4,000-unit B200 GPU cluster to a national project right now. Having already begun pre-training means the remaining period can be devoted entirely to performance optimization and demonstration, which will work as an advantage.”
The Ministry of Science and ICT and NIPA stated in the project guidance, “Cybersecurity is core infrastructure that safeguards the stability of South Korea’s key industries—including communications, transportation, and finance—and the foundation of the national economy. This is the time to establish a South Korean strategic defense system capable of safely protecting against evolving cyber attacks and hacking threats amid a rapidly changing technological environment, including AI exploitation.”
The ultimately selected consortium will receive 256 B200 GPUs for 10 months, with the project scheduled to run through July next year following a mid-term evaluation in February.