Who won, and over whom
South Korea’s Ministry of Science and ICT has selected a consortium led by Naver Cloud to build the country’s sovereign cybersecurity AI models, choosing it over a competing group assembled by SK Telecom. The ministry announced the decision on 3 September, citing partnership breadth, technology and strategy. A contract is expected to be signed later in September.
The consortium runs to 32 members. Alongside Naver Cloud it includes LG CNS, LG Uplus, KEPCO KDN, Korea Hydro & Nuclear Power and the Financial Security Institute — which is to say the companies that run the country’s electricity grid, its nuclear fleet and its financial plumbing.
Two models, and one of them attacks
The programme will produce two complementary models. One is built on Naver’s HyperCLOVA X and specialises in defence. The other is built on LG AI Research’s EXAONE and emphasises offensive capability — described in the ministry’s framing as cyberattack detection, analysis and response.
That pairing is the notable part. Frontier labs have spent 2026 shipping cyber-capable models behind trusted-access programmes and capability thresholds, on the argument that offensive capability is too dangerous to distribute. South Korea’s answer is to build one deliberately, inside a state programme, with named operators of critical infrastructure in the room.
The compute, and where it comes from
The government is providing up to 256 Nvidia B200 GPUs over ten months. That is the state’s contribution; the consortium’s own hardware dwarfs it. Naver Cloud is deploying 4,000 B200s, and LG is adding 256 H200s.
The ratio is worth sitting with. In a programme described as sovereign, the sovereign is supplying about 6% of the accelerators. What the state actually brings is convening power — the ability to put a nuclear operator, a grid operator and a financial regulator’s institute into the same data-sharing agreement.
The data is the real asset
Roughly 830 terabytes of training data will come from Naver Cloud, LG CNS and organisations running critical national infrastructure and industrial systems. Logs from an operating power grid or a nuclear plant’s control network are not scrapeable from anywhere. No amount of GPU purchasing substitutes for them.
Seven sectors will pilot the resulting models: energy, finance, science and technology, telecommunications, semiconductors, defence and aerospace.
The pattern this fits
A growing number of governments have concluded that a general-purpose model licensed from a foreign vendor is not an acceptable foundation for national security work, and have started funding domestic alternatives. What distinguishes this one is specificity. It is not a national general model in the abstract; it is two models for one job, with the customer sector and the data source named in the announcement.
What to watch
The contract terms later this month, and specifically what governs the offensive model: who may run it, against what, and under what authorisation. A state-funded offensive cyber model with critical infrastructure data behind it is a capability whose controls matter more than its benchmark scores.
