Free AI for a nation · Part 5 of 7
Sovereign models on leased American chips
The rule that makes the service sovereign counts the output tokens each model writes, and says nothing about the chips. Of the eight listed models, two were trained from scratch by their makers' account, three grew from their makers' earlier models, and for three the start isn't published.
AI for All is a sovereignty project as much as a public service, and its rules are about models. At least 50% of use must come from domestic models that meet the standard of the government’s sovereign foundation-model programme, and at least 30% from other Korean companies’ models. Foreign models are allowed only for the few functions domestic ones can’t manage, such as hard reasoning, and the state doesn’t pay for them.[22, 27]
The standard is about weights. A team must design and pre-train its model itself, from fresh weights: open code and architectures may be reused, pretrained weights may not.[29] In January 2026 the ministry dropped Naver Cloud from the programme for building on outside encoders whose pretrained weights it had frozen.[29]
“Korean” means several things: who developed the model, where its weights came from, under what licence, who controls the deployed version, where it is hosted, whose hardware it runs on. The rule tests the second. On that, the eight listed models’ cards and reports say this.[18] Two, SK Telecom’s A.X K2 and Motif 3, were trained from scratch, their makers say. Three, Upstage’s Solar Open 2 and LG’s K-EXAONE 2.0 and EXAONE, grew out of their makers’ earlier models. For three, Kakao’s Kanana-2, KT’s Mi:dm K 2.5 Pro and Upstage’s Solar Pro 4, the base isn’t named or the start isn’t published. No listed model is known to come from foreign weights; for three it isn’t known either way, and which versions will be deployed isn’t published. The blueprints are another matter: A.X K2 has the shape of DeepSeek-V3, a Chinese model, and Kanana-2 uses its model code, which the standard allows.[42, 55] A year earlier SK Telecom had built A.X 4.0 on Alibaba’s Qwen2.5-72B; for the sovereign programme it trained A.X K1 from scratch instead.[58] The standard pushed the teams to train their own.
The hardware has no such rule. Every GPU in the service is an Nvidia B200, and the 2027 proposal sets aside 3.85 trillion won ($2.86 billion) for Nvidia’s next platform.[10] A teardown of one B200 found SK hynix’s memory, about 48% of a B200’s build cost by one estimate; that is one unit, not the fleet.[49, 59] Korean chips are in the bids at the edges: FuriosaAI in Kakao’s consortium, Rebellions in KT’s, whose token factory is to route requests across GPUs and Korean NPUs.[12] In the reports I read on the 2027 proposal I found no line for domestic AI chips; that is a gap in my reading, not proof that there is none.
The operating rule is the part that will bite. The tender’s FAQ measures the 50% and 30% shares in output tokens, hidden reasoning included.[22] It decides whose model writes the answers and says nothing about whose chip computes them. Counting tokens also works against a Korean tokenizer, which writes the same answer in fewer of them, so the FAQ scores tokenizer efficiency separately.[22] The models, their origins and the other hardware lines of the 2027 proposal are in the appendix.