"As whether a nation holds frontier AI capabilities will determine its future, we must keep taking bold challenges rather than slowing down, even amid the recent debate over slowing AI," said Deputy Prime Minister and Science Minister Bae Kyung-hoon at the project's launch meeting in Seoul.
Bae's remarks brushed aside a slowdown debate that flared after Anthropic CEO Dario Amodei urged a deliberate pause in the pace of frontier model development, a call some U.S. tech leaders backed and Nvidia CEO Jensen Huang rejected.
The consortium plans a model of about 700 billion parameters built on Naver's HyperCLOVA X and LG's Exaone, trained on about 830 terabytes of real-world security data.
Participants include more than 30 companies, universities and public institutions including LG CNS and KAIST.
The ministry will provide 256 Nvidia B200 graphics processing units for up to 10 months, with the second half subject to a midterm review. Naver Cloud has already put 4,000 of its own B200 chips into pretraining the model, the company said.
Bae pledged support spanning infrastructure, competitiveness and deployment.
The model will be field-tested in seven sectors, including power, finance, semiconductors and defense, and engineered to run on closed networks sealed off from the internet. The consortium plans to release it as open source available for commercial use.
AJP Takeaways
- A Naver Cloud-led consortium formally launched development of a cybersecurity-specialized AI model for South Korea's critical infrastructure on Sept. 22, 2026, at a meeting hosted by the Ministry of Science and ICT.
- The planned model of about 700 billion parameters combines Naver's HyperCLOVA X and LG's Exaone, backed by 256 government-supplied Nvidia B200 GPUs and 4,000 B200 chips from Naver Cloud.
- Deputy Prime Minister Bae Kyung-hoon said on Sept. 22, 2026, that South Korea would press ahead with frontier AI despite the global debate over slowing AI development.
Copyright ⓒ Aju Press All rights reserved.