● The money 21.3 trillion won for AI and three megaprojects in 2027, nearly double this year, inside a record 820.9tn won budget.
● The machine Hangang - 614.7 petaflops, 8,496 NVIDIA GH200 chips, live from today in Daejeon.
● The deadline President Lee has asked parliament to pass the budget by 2 December.
One headline claim circulating today does not appear in Korean reporting of the announcement. More on that below.
South Korea made two AI announcements on the same day. One is a budget proposal that still needs parliamentary approval. The other is a machine that is already running.
The 2027 budget: 21.3 trillion won for AI
The proposed 2027 budget totals 820.9 trillion won, 12.8 percent larger than this year. Within it, 21.3 trillion won goes to AI and three megaprojects - semiconductors, physical AI and AI data centres - which President Lee Jae Myung described as nearly double the current year’s figure.
The breakdown published in Korean reporting:
| Allocation | Amount | What it covers |
|---|
| AI and three megaprojects | 21.3tn won | Semiconductors, physical AI, AI data centres |
| Frontier AI | 4.7tn won | Model development |
| R&D total | 39.5tn won | A record, including a 4tn won increase for the seven SEED areas |
| NEXT 10 strategic technologies | 13.8tn won | Growth engines intended to follow semiconductors |
| Semiconductor production | 2tn won | Early operation of fabs in greater Seoul and a southwestern cluster |
| Chip infrastructure | 2.1tn won | Power, water and industrial complexes |
| Physical AI | 500bn won | Field demonstrations across eight public and private sectors, including manufacturing, smart factories and defence |
Alongside the budget sits a 162 trillion won future-oriented fund, about $120.5 billion, backed by windfall tax revenue from the semiconductor boom. Lee described it as "a platform for strategic investment to boost the economy’s growth potential."
A programme called AI for All covers public AI agents intended to guide citizens from public-service information through to applications.
About that "world’s highest performance" headline
Several Western outlets are running this story with a line about Korea targeting the world’s highest performance sovereign AI model. That phrasing is worth handling carefully.
Korean reporting of the announcement does not use the word "sovereign." Seoul Economic Daily reports the government aiming to bring homegrown AI models into the top tier of global performance, and Lee’s own quoted language was that the government will "invest boldly in artificial intelligence and future growth engines." The stronger superlative appears in headline framing rather than in the budget documents as reported locally. The ambition is real; the specific claim is a paraphrase being sharpened in translation.
This matters because budget allocations are proposals. Parliament has until 2 December to pass the budget and the related legislation, and the figures above are what has been put forward rather than what has been enacted.
Hangang: the part that is already running
On the same day, the Ministry of Science and ICT and the state-run Korea Institute of Science and Technology Information brought Hangang, the sixth national supercomputer, into operation at KISTI headquarters in Daejeon. Pilot testing ran through late October and official full service begins on 16 December.
| Specification | Hangang |
|---|
| Cost | 382.5bn won, about $276 million |
| Theoretical peak | 614.7 petaflops |
| GPUs | 8,496, including NVIDIA GH200 Grace Hopper Superchips |
| CPUs | 9,936 |
| Versus Nurion (No. 5) | 24x theoretical compute, 10x storage, 4x network |
| Cooling | Direct cooling, to reduce running costs |
| Expected ranking | In or near the global top 10 in the upcoming Top500 |
Capacity is allocated 55 percent to R&D innovation, 30 percent to national strategic projects and 15 percent to industrial innovation. KISTI says it plans to meet about 30 percent of national demand for large-scale AI. Users include universities, government-funded research institutes and companies, including AI-focused small and medium-sized enterprises.
Its stated primary mission is powering national strategic research, including K-Moonshot, a government initiative applying advanced AI models to scientific and technological problems. KISTI plans to link it with a quantum computer next year.
Two numbers for the same machine, and why they differ
Coverage of Hangang carries two performance figures that look contradictory. They measure different things, and the distinction is worth keeping straight:
- 24x is the theoretical computing performance against Nurion, the machine it replaces. That is a specification comparison.
- 6.5x is what Hong Tae-young, director of KISTI’s National Supercomputing Infrastructure Center, expects on a CPU basis. It is a stated expectation rather than a published benchmark.
Neither is wrong. A headline that uses one without the other is incomplete, and the Top500 listing is the figure to wait for if you want an independently measured number.
Why this matters outside Korea
Three things worth taking from it.
It is NVIDIA hardware. A national sovereign-compute programme built on 8,496 GH200 Grace Hopper Superchips is a reminder that compute sovereignty and supplier independence are not the same objective. Korea is buying capability, not building an alternative stack.
The capacity split is unusually explicit. Publishing a 55/30/15 allocation between research, strategic projects and industry, and a target of meeting 30 percent of national large-scale AI demand, is more transparent than most national compute programmes manage. It also implies the other 70 percent is expected to come from commercial providers.
The money is proposed, the machine is running. That gap is the whole story. One of these is a parliamentary question with a 2 December deadline; the other is drawing power in Daejeon today. Our prompts for telling an announcement from a product are built for exactly this distinction.
FAQ
How much is South Korea spending on AI in 2027?
21.3 trillion won is allocated to AI and three megaprojects - semiconductors, physical AI and AI data centres - within a proposed 820.9 trillion won total budget. Frontier AI specifically receives 4.7 trillion won. The budget still requires parliamentary approval by 2 December.
What is the Hangang supercomputer?
South Korea’s sixth national supercomputer, operated by KISTI in Daejeon. It cost 382.5 billion won, about $276 million, and has a theoretical peak of 614.7 petaflops from 8,496 GPUs and 9,936 CPUs. It entered operation on 2 November 2026 with full service from 16 December.
Is Hangang in the global top 10?
The ministry expects it to place in or near the global top 10 in the upcoming Top500 rankings. That has not been confirmed yet - the ranking is the independent check.
Is Korea building a sovereign AI model?
The government has stated an ambition to bring homegrown AI models into the top tier of global performance, with 4.7 trillion won allocated to frontier AI. The specific phrase "world’s highest performance sovereign AI model" appears in headline framing rather than in Korean reporting of the budget documents.
What chips does Hangang use?
NVIDIA GH200 Grace Hopper Superchips, among its 8,496 GPUs.
Related reading
Sources
Figures on this page are taken from Korean outlets reporting the announcement directly, read on 2 November 2026, in preference to wire copy. Budget allocations are proposals pending parliamentary approval. Performance figures for Hangang are as stated by the ministry and KISTI and have not been independently benchmarked.