100 days into Korea's AI Framework Act: LG leads, Naver is out, and the chip policy ground has shifted

In 2023, the prevailing Western view was that sovereign foundation models were a vanity project — expensive duplication of what Anthropic, OpenAI, or Google already offered. Three years on, the question has shifted, and so has the Korean answer. It is not whether sovereign models make sense in the abstract, but which institutional and corpus conditions make them work in practice — and which Korean providers are still in contention to deliver them. As of early May 2026, both questions look different than they did at the start of the year.

Five became three: the January cut

On 15 January 2026, the Ministry of Science and ICT announced the first-stage evaluation results for the AI for Everyone sovereign-model program. LG AI Research ranked top, with Upstage and SK Telecom also advancing. Naver Cloud — the consortium most Western coverage assumed would lead the program — failed the 'sovereignty' criteria and was eliminated. NCSoft also did not advance. The startup-led Upstage consortium outperformed conglomerate-led teams for the first time in a Korean national R&D program of this scale. MSIT has signalled it will add a replacement fourth team to sustain competitive pressure; phase 2 runs through June 2026, with the final evaluation set for December 2026.

Three places the gap is real

The value of a sovereign model is almost never in raw benchmark scores against global frontier systems. For reasoning, long-context coding, and English-language synthesis, the gap to Anthropic or OpenAI remains material. The value is in what happens inside regulated, Korean-specific workflows that global models cannot legally or linguistically reach.

• Legal research — Korean case law, Hansik contract phrasing, and Financial Supervisory Service guidance that has no clean English corpus and that global API providers cannot guarantee residency on.
• Public-sector procurement — where the Personal Information Protection Act and the AI Framework Act's high-impact AI provisions create data-residency and human-oversight obligations global APIs do not currently satisfy.
• Clinical documentation — Korean EMR text, insurance claim codes, and NHIS review protocols where a 3-point F1 improvement on in-domain language translates directly into reimbursement decisions.

What the models actually look like — and what shipped in April

On 9 April 2026, LG AI Research launched EXAONE 4.5 — a 33-billion-parameter Vision-Language model that unifies a proprietary vision encoder with the LLM. It is the modality-expansion stepping-stone toward what LG describes internally as K-EXAONE. Naver Cloud's HyperCLOVA X SEED family, released through 2025, still includes the 8B Omni multimodal variant and a 32B Think reasoning model that scored 44 on the Artificial Analysis Intelligence Index — but the firm's exclusion from the MSIT program changes the commercial backdrop sharply. Upstage's Solar Pro 2 (31B parameters) remains a frontier-class system on Korean-language evaluations, and the firm's April 2026 Series C raised ₩180 billion at unicorn valuation, alongside a Tokyo subsidiary established with JETRO support and active discussions with AMD for roughly 10,000 MI355 GPUs.

The Framework Act, 100 days in

The Act took effect on 22 January 2026 and applies a risk-based regime. High-impact AI systems — deployed in healthcare, finance, public services, and critical infrastructure — generative AI, and high-performance AI defined by cumulative compute thresholds face enhanced obligations on safety, reliability, human oversight, and risk management. MSIT confirmed in its post-enforcement guidance that penalties (capped at ₩30 million) carry a one-year grace period while the regulator issues supplementary decrees clarifying meaningful-explanation duty, user protection plans, and generative AI labelling. Combined with PIPA and sector-specific guidance from the Financial Services Commission and the Ministry of Health and Welfare, the implementation signal for regulated sectors remains unambiguous: a domestic compute-and-corpus stack is the path of least regulatory friction.

The compute stack — and the export-control reset

A sovereign model without a sovereign compute stack is a rented strategy. Naver Cloud completed a 4,000-unit NVIDIA B200 GPU cluster at its Sejong data centre in March 2026 — its largest single GPU build to date. MSIT is running a second-round national GPU tender; NHN Cloud and Kakao reportedly opted out of the bidding due to data-centre capacity constraints, with Coupang re-entering the field as a contender. The harder constraint, however, is upstream. By March 2026 the Trump administration began deploying AI chip export licences as a trade lever. Korea lost its Biden-era Tier 1 status and now faces conditional export licences tied to US data-centre investment commitments — the same template applied to Saudi Arabia and the UAE.

“You cannot call a model sovereign if the inference stack is a phone call to Santa Clara away from being revoked. We assumed that constraint was theoretical. In March it stopped being theoretical.”

— Advisor to Korean Presidential AI Committee

What we advise boards to ask

The question is not whether to build a model. For most enterprises, buying global frontier capability and renting Korean-specialised endpoints for regulated surfaces is still the right architecture. The question is where the regulated surface area of the business touches the foundation-model layer, and whether the chosen provider can guarantee legal, linguistic, and compute residency on that surface. For a widening list of sectors — finance under FSC guidance, healthcare under the NHIS review regime, public-sector procurement under the PPS rules — only a domestic stack currently qualifies. And for the next twelve months at least, that domestic stack is being shaped less by benchmark leadership and more by which firms cleared the MSIT gate in January.