PolicyLaw, regulation & governance
South Korea proposes 4.7 trillion won for a 'world's highest performance' sovereign AI model in record 2027 budget
South Korea's proposed 820.9 trillion won 2027 budget sets aside 4.7 trillion won (about $3.5 billion) for a sovereign AI model meant to lead the world, though the government's latest evaluation shows Korean contenders still trailing US frontier models.
From the CNBC · Technology article of November 2, 2026 and cited sources: President Lee Jae Myung's 2027 budget sets aside 4.7 trillion won for a sovereign AI model he says will have the world's highest performance. [1] [3]
SOURCES[1] CNBC: South Korea targets 'world's highest performance'… cnbc.com[3] FN News: President Lee delivers budget address en.fnnews.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- President Lee Jae Myung's 2027 budget sets aside 4.7 trillion won for a sovereign AI model he says will have the world's highest performance. [1] [3]
- Original source
- CNBC · Technology article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 2, 2026 article published
- Covered
- Nov 2 edition
- Evidence
- 11 sources across 9 sites
- 1CNBC: South Korea targets 'world's highest performance'…cnbc.com
- 2Money Today: Second-round evaluationmt.co.kr
- 3FN News: President Lee delivers budget addressen.fnnews.com
- 4Korea Times: 5 consortia selectedkoreatimes.co.kr
- 5Gulf News: South Korea to triple AI spendinggulfnews.com
- 6Korea Times: Naver Cloud, NC AI cutkoreatimes.co.kr
StrongHow it’s measured
YouTube video · 115.2K views · 14.6K likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
President Lee Jae Myung's 2027 budget sets aside 4.7 trillion won for a sovereign AI model he says will have the world's highest performance. It also puts 21.3 trillion won toward chips, physical AI and AI data centers, nearly double the 2026 level. The National Assembly still has to pass the budget, and its statutory deadline is December 2. [1] [3]
The target is well ahead of current results. In the August 2026 government evaluation, the best Korean model scored about 47 on a global intelligence index, while leading US models scored 61 to 63. Three teams remain: LG AI Research (K-EXAONE 2.0), SK Telecom (A.X K2) and Upstage (Solar Open 2). Each gets about 1,000 B200 GPUs. Two finalists are due by year-end, and the Science Ministry is reportedly considering backing a single model. [2] [8]
The program defines 'sovereign' strictly. Models must be trained from scratch, and any reused open-source parts must have their weights fully reset and retrained. Naver Cloud was cut for not meeting this rule. Scoring also rewards real-world use, not just benchmarks: Motif was eliminated despite topping the global benchmark index because its usability fell short. [6] [2] [11]
Korea's best entry scored about 47 on a global index, against 61–63 for leading US models. Implication: it is unproven that new funding will close that gap, and firms building on Korean models should expect one or two state-backed options.
Executive brief
South Korea plans to spend 4.7 trillion won (about $3.5 billion) on a sovereign AI model that President Lee Jae Myung says will have the "world's highest performance." That is the biggest single AI-model commitment in the record 820.9 trillion won 2027 budget (CNBC). The target is far ahead of the evidence. In the government's own August 2026 evaluation, the best Korean contender scored about 47 on a global intelligence index, while leading US frontier models scored 61–63 (Money Today). The National Assembly still has to pass the budget, and the statutory deadline is December 2 (FN News).
What changed and event timeline
Five teams picked for the national model race
The Science Ministry (MSIT) chose Naver Cloud, Upstage, SK Telecom, NC AI and LG AI Research for its Independent AI Foundation Model Project, with a first evaluation set for December ().
The 2026 budget puts 10.1 trillion won into AI
Lee framed the spending as a push to make Korea one of the world's top three AI powers. The total 2026 budget was 728 trillion won ().
Naver Cloud and NC AI are cut
Naver did not meet the originality rule because its open-source encoders were not fully reset and retrained. LG led overall with 90.2 points ().
Motif fills the open slot
A consortium led by Motif Technologies entered the second round as the fourth team ().
Motif is eliminated and three teams remain
K-EXAONE 2.0, A.X K2 and Solar Open 2 advance. Each team gets about 1,000 B200 GPUs (;).
The record 2027 budget is proposed
It includes 4.7 trillion won for a sovereign model and 21.3 trillion won for chips, physical AI and AI data centers. The chips/AI figure is nearly double the 2026 level ().
Capabilities and access
- The three remaining models: LG AI Research's K-EXAONE 2.0, SK Telecom's A.X K2 and Upstage's Solar Open 2 (Let's Data Science).
- What the money buys: The budget reporting does not name a model, parameter count, release date or licence for the 4.7 trillion won model.
- Compute: Each team gets about 1,000 B200s, roughly 40 billion won in rental support per team (Money Today).
Read the full section
- The three remaining models: LG AI Research's K-EXAONE 2.0, SK Telecom's A.X K2 and Upstage's Solar Open 2 (Let's Data Science).
- What the money buys: The budget reporting does not name a model, parameter count, release date or licence for the 4.7 trillion won model. Linking that money to the existing foundation-model program is an inference, not something the budget reporting states.
- Compute: Each team gets about 1,000 B200s, roughly 40 billion won in rental support per team (Money Today). That is up from about 768 B200-equivalent units in the first half of 2026 (Let's Data Science).
Technical analysis for researchers and developers
- What counts as "sovereign": Models must be trained end-to-end from scratch, with their own architecture, data and training pipeline, and no outside licensing constraints.
- How teams are scored: benchmarks 40 points, expert review 35, user evaluation 25 (Money Today).
- Reproducibility: The reviewed sources do not describe full test sets, prompts or weights, so outside researchers cannot rerun the scoring.
Read the full section
- What counts as "sovereign": Models must be trained end-to-end from scratch, with their own architecture, data and training pipeline, and no outside licensing constraints. Reusing open-source components is only acceptable if their weights are fully reset and retrained (Korea Times).
- How teams are scored: benchmarks 40 points, expert review 35, user evaluation 25 (Money Today). In round 1, benchmarks combined NIA tests, global common benchmarks and team-specific targets (Asia Economy).
- Reproducibility: The reviewed sources do not describe full test sets, prompts or weights, so outside researchers cannot rerun the scoring.
Claims and evidence
- "World's highest performance": This is a political target (CNBC). No independent evaluation in the reviewed sources supports the target.
- "Virtually no net increase" in government debt: This is Lee's claim, based on using new bonds to refinance existing debt (CNBC).
- Exports: October exports rose 93.7% year on year to $115.3 billion, according to Korea Customs Service data (CNBC).
Read the full section
- "World's highest performance": This is a political target (CNBC). The latest reported evaluation shows Korean entries behind US frontier models such as Claude Opus and GPT-5.6 (Money Today). No independent evaluation in the reviewed sources supports the target.
- "Virtually no net increase" in government debt: This is Lee's claim, based on using new bonds to refinance existing debt (CNBC). The opposition People Power Party calls the budget's Future Response Fund a "future debt mountain fund" (JoongAng Daily).
- Exports: October exports rose 93.7% year on year to $115.3 billion, according to Korea Customs Service data (CNBC).
Context and prior work
- Why Korea can afford this: The AI build-out depends on chip export earnings.
- Earlier commitments: The 2027 plan follows the 10.1 trillion won AI allocation in the 2026 budget (Gulf News) and two elimination rounds in the model program (Korea Times).
- Defense: Lee also tied AI to the military. Defense spending rises 8.2% to 73.3 trillion won, aimed at combined manned and unmanned systems (CNBC).
Read the full section
- Why Korea can afford this: The AI build-out depends on chip export earnings. Samsung and SK Hynix supply memory used in AI data centers (Gulf News).
- Earlier commitments: The 2027 plan follows the 10.1 trillion won AI allocation in the 2026 budget (Gulf News) and two elimination rounds in the model program (Korea Times).
- Defense: Lee also tied AI to the military. Defense spending rises 8.2% to 73.3 trillion won, aimed at combined manned and unmanned systems (CNBC).
Limitations, safety and contested findings
- Usability can beat raw scores: Motif was eliminated even though it had the top score on the global benchmark index, because it struggled on usability (Money Today).
- One model or several: MSIT is reportedly considering putting all support behind a single model, given how far the teams trail global frontier models (Money Today).
- Disputed originality rules: Industry reporting described the originality bar as hard to meet (Asia Economy).
Read the full section
- Usability can beat raw scores: Motif was eliminated even though it had the top score on the global benchmark index, because it struggled on usability (Money Today).
- One model or several: MSIT is reportedly considering putting all support behind a single model, given how far the teams trail global frontier models (Money Today).
- Disputed originality rules: Industry reporting described the originality bar as hard to meet (Asia Economy).
- Safety: None of the reviewed sources discuss safety evaluations or model cards.
Business and practitioner implications
- Expect fewer winners: Two finalists are due by year-end (Let's Data Science).
- Suppliers: The 21.3 trillion won infrastructure line should bring procurement in data centers, physical AI and chip-design training, including 2,000 new specialists (CNBC).
- Timing risk: None of this is final until the Assembly passes the budget; the deadline is December 2 (FN News).
Read the full section
- Expect fewer winners: Two finalists are due by year-end (Let's Data Science). Firms building on Korean models should plan for one or two state-backed options, likely from LG, SKT or Upstage.
- Suppliers: The 21.3 trillion won infrastructure line should bring procurement in data centers, physical AI and chip-design training, including 2,000 new specialists (CNBC).
- Timing risk: None of this is final until the Assembly passes the budget; the deadline is December 2 (FN News).