On May 31, 2026, Deputy Prime Minister and ICT Minister Bae Kyung-hoon did something regulators rarely do: he reopened a rule his own ministry had already enforced. Calling for Korea to become a "full-stack" frontier-AI nation, Bae welcomed a fresh debate over whether foreign open-source code and pre-trained weights — Alibaba's Qwen most pointedly — should be permitted under the Sovereign AI Foundation Model project's strict "from-scratch" independence rule. "This is exactly the bright future of South Korea's AI," he said, with second-phase results due in the second half of 2026. It was a notable softening, because that same rule had already cost Korea one of its strongest contenders.
What the rule did
The Sovereign AI Foundation Model project — backed by roughly KRW 213.6 billion in public funding over up to three years — selected five elite teams in 2025: Naver Cloud, Upstage, SK Telecom, NC AI, and LG AI Research. In the first evaluation round announced on January 15, 2026, only three advanced: LG AI Research (which topped the field at 90.2 points), SK Telecom, and Upstage. Naver Cloud and NC AI were cut.
Naver Cloud's elimination was the telling one. Per the Korea Herald, the company was disqualified not for weak performance — its scores were strong — but for using pre-trained vision encoders and weights from Alibaba's open-source Qwen model. Second Vice Minister Ryu Je-myung set out the standard plainly: "Even if open-source use is common in the global ecosystem, retraining from scratch is the minimum condition for independence." The ministry went further, characterizing certain open-source reuse as "free-riding." To qualify as sovereign, teams had to reset weights entirely and rebuild them on independently secured data.
The case for the rule, fairly stated
The strongest argument for from-scratch independence is not nativism — it is supply-chain risk. A model built atop foreign weights inherits whatever is baked into them: licensing terms that can change, latent biases, and security exposure no audit fully reaches. That concern is not theoretical. In May 2026, Ars Technica documented a hacker group poisoning open-source packages "at an unprecedented scale," turning a once-rare supply-chain attack into a near-weekly event. For models destined for defense, healthcare, and public administration, a government insisting it can trace every layer to a domestic source is being prudent, not paranoid. Sovereignty, in this framing, means never having a foreign vendor or attacker sitting silently inside the national stack.
Why the rule still gets the trade-off wrong
The problem is that "from-scratch" mistakes provenance for control. Modern frontier models are assembled, not conjured. Reusing an open-source encoder and then fine-tuning, distilling, or retraining around it is the global standard precisely because it lets small teams stand on shared progress instead of re-deriving solved problems at enormous cost. Naver Cloud's defense — that using open source is simply how competitive AI is built — describes the actual practice of every leading lab, including American ones that build on Llama, Mistral, and Qwen derivatives.
The penalty for ignoring that reality is steep, and Bae has effectively admitted it. By his own account, the Korean government's entire AI budget is "around the level of investment made by a single US Big Tech company." When you are that far behind on capital, forbidding teams from reusing freely available weights does not buy independence — it buys a slower, more expensive path to a weaker model. A rule meant to reduce foreign dependence can entrench it, because a Korean model that underperforms is a Korean model nobody uses, leaving the market to ChatGPT, Gemini, and Claude by default.
It also drew the line in the wrong place technically. A vision encoder converts images into machine-readable signals; using a well-tested open one is closer to using a standard compression library than to outsourcing the model's reasoning. Banning that while permitting other imported tooling is a provenance test dressed up as a security test — and it cost the project two of its five teams, with Naver and Kakao, Korea's two largest internet platforms, declining to rebid at all.
A better version of sovereignty
Real AI sovereignty is about control and capability, not purity of origin. Korea has the inputs: over 260,000 NVIDIA GPUs deployed nationwide, including up to 50,000 earmarked for sovereign AI, and roughly USD 1.1 billion committed to domestic AI computing per OECD records. The proportionate rule is not "no foreign weights ever" but "foreign components must be auditable, license-clean, security-reviewed, and replaceable" — graded transparency rather than a binary ban. That preserves the genuine security interest behind the steelman while letting teams reuse what the rest of the world reuses.
Bae's willingness to reopen the question, mid-project, is exactly the right instinct. A government that can revise a rule when the rule starts working against its own goal is more credible than one that treats a January definition as permanent. The second-phase results in H2 2026 are the real test: if Korea wants a full-stack frontier model on a fraction of Big Tech's budget, it should be measuring whether its models are good, controllable, and secure — not whether their encoders were born onshore.