South Korea's $3.5 Billion: Sovereignty Starts on the GPU Shelf
South Korea has put 4.7 trillion won (about $3.5 billion) into its 2027 budget proposal to launch a frontier-class national model programme. But $2.6 billion of that buys only 10,000 Nvidia Vera Rubin chips — the first step of sovereignty is paying an American chip company. Officials admit they cannot catch the US giants; the target is China's top open-source models.
- 01THE SHOCK — In April, Anthropic released Claude Mythos 5; the 'Mythos Shock' triggered South Korea's sense of crisis.
- 02THE PRICE TAG — At a July press conference, the minister quoted the price: 3.5 trillion won buys only 10,000 Vera Rubin chips.
- 03THE CONFIRMATION — On October 6, at a parliamentary audit, Bae Kyung-hoon confirmed the new frontier plan, with both tracks running in parallel.
WHAT HAPPENEDWhat Happened
At a parliamentary audit on October 6, 2026, Bae Kyung-hoon, Minister of Science and ICT (MSIT), confirmed in person: South Korea will launch an entirely new frontier-level AI model programme with a budget of 4.7 trillion won (about $3.5 billion), already included in the 2027 government budget proposal. Note: parliament has not approved it. This is still at the proposal stage, not an appropriation.
The minister had already quoted the spending plan at a July press conference: 'Developing a frontier model would require about 3.5 trillion won just to purchase 10,000 of Nvidia's new Vera Rubin chips.' Some 3.5 trillion won (about $2.6 billion) buys only 10,000 of Nvidia's new Vera Rubin chips, due in the second half of the year; the rest buys training data, and participating companies must match the funding with equal amounts of private money.
This is the 'second track' of South Korea's AI strategy. The existing 'sovereign AI foundation model' project is a knockout competition: five companies initially split 530 billion won (about $390 million); LG AI Research, SK Telecom, Upstage and others remain in a third round, with two winners to be chosen by year-end. The new plan starts from scratch — the sovereign models take the 'good enough for government use' route, while the frontier programme's stated goal is refreshingly realistic: it cannot catch the US giants, so it aims to arm-wrestle China's top open-source models.
The trigger was the 'Mythos Shock': after Anthropic released Claude Mythos 5 in April, Bae said model performance is improving exponentially and the gap keeps widening, making an investment of a different magnitude necessary. Behind the plan sits an even bigger puzzle: during the October 2025 APEC summit, South Korea signed a deal with Nvidia for 260,000 GPUs worth $9.8 billion, to be delivered by 2030, with roughly 50,000 allocated to the government and the national AI computing centre.
THE GPU BILL3.5 trillion won buys only chips: sovereignty's first invoice
Break the budget down and the most striking figure is the ratio: $2.6 billion of the $3.5 billion buys just 10,000 American chips. The first step of a so-called 'national custom model' is wiring money to Nvidia. Sovereignty starts on the GPU shelf — that is not sarcasm; it is the literal meaning of South Korea's budget proposal.
Earlier commitments have already paved the road: the Lee Jae-myung government has pledged 50,000+ GPUs for national AI data centres, and the APEC deal covers 260,000 GPUs worth $9.8 billion through 2030. The 4.7 trillion won plan is the 'model end' of this national compute puzzle — hardware first, models after.
But the arithmetic does not end there: 10,000 Vera Rubin chips are only the entry ticket. Training a frontier-class model puts power, data, talent and trial-and-error costs all behind it. The 3.5 trillion won chip purchase is just the first invoice; more invoices follow. This time South Korea has written 'you cannot buy sovereignty' into its budget: first admit the dependence, then talk about overtaking.
TWO TRACKSKnockout rounds plus a new track: an official admission it cannot catch the US
South Korea is being unusually honest this time. Officials openly admit the frontier programme's goal is not to catch the US giants but to compete with China's strongest open-source models. The sovereign model knockout continues on its 'good enough' route — usable by government, at controllable cost.
But the knockout itself has already been questioned by the industry. Korea JoongAng Daily cited industry criticism: 'many in the industry also criticized the government's attempt to pick AI models through competitive rounds as a failure.' The government picking models by competition was branded a failed attempt — and now it is launching a separate frontier track, which amounts to an official admission that the first road was too narrow.
Bae's remarks at the audit deserve a close read: 'The sovereign model will stay on target, but we judged that a frontier AI project is additionally needed to tackle humanity's unsolved problems and respond to cybersecurity threats.' In other words: the old track continues, the new one adds capacity — walking on two legs so that if one goes lame, the whole programme does not collapse.
AI FOR ALLAI for All: the open gambit of free, unlimited national AI
The supporting ambition is bigger than the model itself: 'AI for All', a free and unlimited national AI service launching by year-end, must run at least 50% on South Korean domestic models. The winners of the sovereign model knockout will become the foundation of this nationwide service.
It is a clever move: a national-scale traffic gateway underwrites domestic models. Whatever rank a model holds on international benchmarks, once the national service runs on it, the ecosystem, usage data and iteration flywheel start turning. It turns 'sovereignty' from a paper metric into national infrastructure.
But the risk sits in the same place: free and unlimited means the state underwrites the compute cost. A 260,000-GPU deal, a 4.7 trillion won budget, a 50% domestic-model requirement — this is a complete playbook of trading fiscal spending for an ecosystem. The bet: money burned today buys freedom from being choked tomorrow.
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISSovereignty starts on the GPU shelf
South Korea has, for the first time, written 'you cannot buy sovereignty' into a budget: $2.6 billion of $3.5 billion goes first to American chips, and participating companies must match with equal private funding. That is not weakness; it is bookkeeping — admit shelf dependence first, then talk about overtaking on models. The real starting line of sovereign AI is not in research papers. It is on Nvidia's shipping manifest.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
Korea JoongAng Daily: South Korea is splitting its AI strategy into two tracks — the sovereign model knockout continues, and a new frontier programme targets China's top open-source models; the 4.7 trillion won is in the 2027 budget proposal, not yet approved by parliament.
PRIMARY SOURCESPrimary Sources
- Korea JoongAng Daily: Korea splits AI push into two tracks amid confusion over sovereign model project (2026-10-06)
- Aju Press: Report on South Korea's frontier AI programme budget (2026-10-06)
- The Decoder: South Korea bets $3.49 billion on building a homegrown frontier AI model to rival China's best
- www.koreajoongangdaily.com: Original article
https://www.koreajoongangdaily.com/business/korea-splits-ai-push-into-two-tracks-amid-confusion-over-sovereign-model-project/12904222 - the-decoder.com: Original article
https://the-decoder.com/south-korea-bets-3-49-billion-on-building-a-homegrown-frontier-ai-model-to-rival-chinas-best/
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook