Korea relaxed its chip mandate after two tenders selected no operator, then cut Naver over foreign model weights. Enforcement went where state power reached.
Between August 8 and 11, 200 South Korean citizens will spend four days with the four models still standing in Korea's Sovereign AI Foundation Model project. The National IT Industry Promotion Agency recruited them by open application and computerized random draw, weighted to mirror the resident registry's sex and age mix (recruitment notice). Their evaluations feed the project's second-phase review. Public input into a state procurement, not a popular vote.
The panel feeds a program with money behind it. Under AI for All, the Ministry of Science and ICT (MSIT) plans a free, unlimited chatbot for all citizens this year. Its two or three operators must run at least 50 percent of the service on domestic models that meet the sovereign-model criteria and at least 30 percent on domestic models from other developers, foreign models only where minimally necessary, an 80 percent domestic floor (MSIT). Operator selection is a separate procurement, with applications open through August 11; competition winners are not automatically suppliers.
Set the panel beside Seoul's two other instruments this year and a pattern emerges. The National AI Computing Center, which broke ground at Haenam on August 3, began with hard sovereignty conditions: a 50 percent domestic-chip mandate, put-option obligations guaranteeing returns on public capital, and 51 percent public equity. All three were relaxed after two tender rounds produced no selected applicant. The model project runs on a categorical provenance standard; in January MSIT enforced it against Naver, one of Korea's largest and most experienced AI developers. AI for All writes domestic-model floors into a state-funded service.
Three instruments, one pattern. Where sovereignty needed private partners to bid and build, the state traded its conditions away. Where it could be exercised by administrative decision, a disqualification here, a contract floor there, it held. A model rule needs a ministry willing to say no; an infrastructure rule needs bidders, capital, accelerators, and grid power.
Two calls for proposals ran on the original terms between January and June 2025; in the ministry's own words, "no applicants were selected" (revised implementation plan). The revision kept the building and shed the conditions. Public equity dropped below 30 percent, the put option was removed, and the chip mandate became "phased adoption and demonstration" of domestic silicon on operator-proposed terms. A Samsung SDS-led consortium bid: preferred bidder March 10, selected May 11, operating company formed June 10, groundbreaking August 3 (Samsung SDS). The operator puts project cost at ₩2.5 trillion, about $1.8 billion, and completion at 2028.
What the state holds is thinner. Public ownership sits below 30 percent; the center is state-anchored, not state-owned. Committed scale is narrower than the figures orbiting it: more than 15,000 GPUs by 2028 for the center itself; the 50,000-plus figure belongs to the broader public-private "AI Highway" initiative nationally.
In January, MSIT eliminated Naver Cloud from the competition, which Korean industry shorthand calls Dokpamo, after finding its HyperCLOVA X Seed 32B Sync model used the vision encoder from Alibaba's Qwen 2.5-VL with frozen weights. The standard is categorical, no percentage anywhere. Teams must reset weights and train independently, end to end, on an original architecture, the ministry's "minimum requirements for securing the sovereign nature" (MSIT). MSIT argued the strictness on security grounds: overseas AI models in defense, diplomacy, and national infrastructure "could pose threats to national security or increase the risk of leaks of confidential national information."
Naver had answered before the decision. In a January 7 statement, it called the Qwen component "a technical decision to enhance the overall completeness and stability of the model by using already standardized high-performance modules, not because of insufficient technological independence," adding that it is "possible to switch to our own encoder in the future as the model is further developed" (Asia Business Daily). MSIT cut the company anyway.
The chip mandate dissolved when two tender rounds ended with no applicant selected; the provenance standard was enforced against a company already inside the program, at the cost of losing it.
The four remaining teams shipped in late July, and the releases travel under one word: open. The licenses disagree. SK Telecom's A.X K2, a 688-billion-parameter mixture-of-experts model, is Apache 2.0. LG's K-EXAONE 2.0, 749 billion parameters (styled 750B), is Apache 2.0 as well. Upstage's Solar Open2, 250 billion parameters, ships under the custom Upstage Solar License: derivatives must carry the Solar name, a "Built with Solar" notice, and the license itself. Motif Technologies' Motif-3 Beta is a checkpoint restricted to noncommercial use; commercial use requires written permission. All four are open weight. Only the Apache pair permits commercial use without bespoke conditions.
Weight access is one layer of sovereign control, and only one. A ministry can hold Apache-licensed weights and still not hold the terms that decide a dispute: who owns the model, who authorizes the next release, what the license permits once interests diverge. SCN raised the same questions about Genesis-Science-1, the model Arcee AI is building with the US Department of Energy, whose ownership, license, and release-authority terms remain undisclosed. Korea settles provenance by state decision; the winners' license terms belong on the same watch list.
The largest numbers sit at the infrastructure layer, where one week can mix a signed procurement, conditional equity, and a nonbinding term sheet under one headline figure. Sorted by instrument:
Project | Stated scale | Instrument | What is firm | Primary dependency |
|---|---|---|---|---|
National AI Computing Center, Haenam | 15,000+ GPUs by 2028; ₩2.5T | Signed procurement; public-private operator | Operator selected; construction under way | Accelerator order; grid interconnection |
NAVER GAK Sejong AI factory | 55 MW DSX buildout announced June 2026, due online 2027, then 200 MW by 2028 | Announcement; conditional equity; nonbinding term sheet | The campus, operating since November 2023, designed for up to 270 MW | NVIDIA's $1B closing conditions; $9B+ committed financing |
SK Telecom AI cloud | "2-gigawatt-scale AI Cloud"; first AI factory planned for 2027, Vera Rubin with SK hynix HBM4 | Letters of intent | The letters; no site or financing disclosed | Contracts; gigawatt-scale generation and transmission |
The NAVER row is easiest to misread: the campus runs, while the AI factory inside it is not yet built, and the expansion's financing is partly contingent. NVIDIA's $1 billion equity is subject to customary closing conditions, including NAVER securing at least $9 billion in committed financing; Brookfield's up to $9 billion is a nonbinding term sheet (NVIDIA).
Grid position explains more of this map than strategy does. As of March 2026, CBRE Korea counted 522 first-stage power-system impact assessment applications in the Seoul metropolitan area, totaling 33,592 MW; ten had final supply approval, 1.9 percent, capital-region scope only (CBRE). The geography, Haenam in the far southwest, NAVER at Sejong, is consistent with developers following available power out of the capital region, an inference from siting, stated as one. SCN has argued that contracted power is the AI buildout's scarcest asset; Korea is the national-scale test.
Public money sits across every layer. On April 30, the National Growth Fund's deliberation committee approved the ₩400 billion capitalization of the Haenam operating company and a ₩560 billion package for Upstage, one of the four finalists; the fund's direct portion is ₩100 billion (Kyunghyang). Its first direct equity investment, in late March, was ₩250 billion into NPU designer Rebellions. One vehicle holds public exposure at all three layers: compute center, model competitor, domestic chipmaker.
Every strand ends in a document with a date. The AI for All provider contracts, due after applications close on August 11, will show how the floors are measured and which models carry the service. Final model licenses will show what sovereign means commercially. At Haenam: an accelerator order, an interconnection agreement, and an early-service plan, since MSIT's terms favored proposals launching service before 2027, a year before completion. For NAVER, the test is closed financing. For SK Telecom, anything harder than a letter of intent.
As of August 7, 2026, SCN had not located a published interconnection agreement for Haenam, a disclosed accelerator vendor, an operator plan for pre-completion service, or license terms binding a winning model into AI for All. The nearest date is already fixed: the citizen panel finishes on August 11.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
