Korea's sovereign AI stack: who gets the GPUs, models and chips

South Korea's sovereign AI effort has three layers: shared computing power, home-grown models and home-grown chips. Nvidia's own release of 30 October 2025 lists over 260,000 GPUs, split among the government, Samsung, SK Group, Hyundai Motor Group and Naver Cloud. Three teams remain in the national model contest, and Korean chip startups aim to supply the inference hardware.

Our South Korea infrastructure overview covers power and policy. This post goes one level down: it sets out the company by company GPU split from Nvidia's release, the model programme and its 2026 evaluations, and the chip startups, and flags what is still unclear. It is a factual summary with sourced figures, not investment advice.

Key numbers at a glance

• Over 260,000: Nvidia GPUs across Korea's sovereign clouds and AI factories (Nvidia press release, 30 October 2025).
• Up to 50,000: latest Nvidia GPUs for the Ministry of Science and ICT (MSIT), starting with an initial 13,000 Blackwell and other GPUs deployed by Naver Cloud, NHN Cloud and Kakao (same release).
• Over 50,000 each: Samsung's semiconductor AI factory and SK Group's AI factory (same release).
• 50,000: Blackwell GPUs for Hyundai Motor Group, with an investment of about 3 billion US dollars in Korea's physical AI landscape (same release).
• Over 60,000: GPUs for Naver Cloud for sovereign and physical AI (same release).
• About 45,000: Korea's existing GPU count as estimated by local analysts before the deal (Data Center Knowledge, 3 November 2025). This is an estimate, not an official count.
• 5: teams named at the start of MSIT's Sovereign AI Foundation Models project: LG AI Research, Naver Cloud, NC AI, SK Telecom and Upstage (Nvidia release, 30 October 2025).
• 3: teams that cleared the second evaluation, SK Telecom, LG AI Research and Upstage (MSIT result reported by Digital Today, 18 August 2026).
• 768 to 1,000: B200 GPUs supplied by the government to the remaining teams, rising from the first half to the second half of 2026 (Digital Today, 18 August 2026).
• 10.1 trillion won: proposed 2026 AI budget, against 3.3 trillion won in 2025 (The Korea Times, 29 August 2025, reporting the cabinet's proposal).
• 100 trillion won: growth funds the government planned to mobilise over five years, including 1 trillion won in 2026 (same report).
• About 100 trillion won: private investment targeted over five years in AI data centres and power grids (Financial News, 5 April 2026).
• 22 January 2026: date the AI Basic Act took effect (Cooley, 27 January 2026).
• Up to 30 million won: maximum administrative fine under the Act, with a one year grace period indicated (Cooley).

What "sovereign AI" means here

The phrase can mean several things. In Korea's programme it means that the country wants to own or control three things: the computing power used to train and run models, the models themselves, and, over time, the chips. The aim is to reduce dependence on foreign cloud services and foreign models for work that matters nationally.

That does not mean Korea avoids foreign technology. The largest line in the computing layer is a purchase of Nvidia GPUs, and home-grown chips still depend on memory that only a few companies make, as the Rebellions details below show. Sovereign here means choice and control, not isolation.

Layer 1: computing power and who gets it

Nvidia's press release of 30 October 2025, timed with the APEC summit in Korea, is the most direct public source on the split. It says Korea's sovereign clouds and AI factories will have over 260,000 Nvidia GPUs. The breakdown it gives:

• Government (MSIT): up to 50,000 of the latest GPUs. The release says it starts with an initial 13,000 Blackwell and other GPUs deployed by Naver Cloud, NHN Cloud and Kakao Corp.
• Samsung: a semiconductor AI factory with over 50,000 GPUs.
• SK Group: an AI factory with over 50,000 GPUs, plus an industrial AI cloud using RTX PRO 6000 Blackwell Server Edition GPUs.
• Hyundai Motor Group: 50,000 Blackwell GPUs.
• Naver Cloud: over 60,000 GPUs for sovereign and physical AI.

Adding the headline numbers gives 50,000 plus 50,000 plus 50,000 plus 50,000 plus 60,000, which is 260,000. That sum is simple arithmetic on the release, and it matches the release's "over 260,000", but Nvidia does not say that the five lines are the whole of the total, so do not treat the sum as an audited figure.

Two details are easy to miss. First, "up to 50,000" for the government is a ceiling, and the initial tranche in the release is 13,000. Second, the Data Center Knowledge report of 3 November 2025 describes the government's share a little differently: it says MSIT would secure 50,000 units for sovereign models and deploy them at the National AI Computing Center and at Naver Cloud, NHN Cloud and Kakao facilities. The two accounts agree on the number and differ on the sites, and the planned National AI Computing Center is not named in Nvidia's text. Treat the deployment location as unsettled.

What is a GPU deal, really? It is a supply agreement and a statement of intent. Neither document says how many units are installed and running. Delivery depends on Nvidia production, on finished buildings and on power at the site, which is the subject of our post on Korea's nuclear fleet and grid. Data Center Knowledge cites local analysts estimating about 45,000 GPUs in Korea beforehand, so the deal would be about five times that, and notes that neither MSIT nor Nvidia disclosed financial terms. The Chosun Daily's estimate of 14.8 trillion won, repeated there, rests on an assumed price per unit and is not a disclosed contract value.

The release also describes uses beyond language models. Samsung is using Nvidia software to develop home robots, SK Telecom's industrial cloud supports digital twin and robotics projects, and Hyundai plans an Nvidia AI Technology Center and a Physical AI Application Center. Physical AI means AI that controls machines in the real world, such as robots and vehicles. The release also lists work on AI-RAN and 6G with Samsung, SK Telecom, ETRI, KT, LG Uplus and Yonsei University, and a role for the state research institute KISTI around its HANGANG supercomputer. These are announced collaborations, and the release gives no completion dates.

Layer 2: the national foundation models

A foundation model is a large model trained on broad data that can be adapted to many tasks, such as the models behind modern chat assistants. MSIT's Sovereign AI Foundation Models project funds Korean teams to build such models. According to Nvidia's release, the participants are LG AI Research, Naver Cloud, NC AI, SK Telecom and Upstage.

The project is staged, with teams evaluated and the field narrowed, and 2026 brought two rounds of results.

First evaluation, January 2026. Inven Global reported on 15 January 2026 that MSIT had announced the first stage results. LG AI Research ranked first overall, and LG AI Research, Upstage and SK Telecom advanced. Naver Cloud was cut for failing the project's sovereignty requirement: its technical report said it had used the weights of an existing open model as is for its video and audio encoders, and MSIT said that did not meet the rule that a team designs and trains its own model. Naver said afterwards that it has its own encoders and the external share was small, but the ministry did not revisit the result. NC AI was not among the teams that advanced. MSIT opened an extra call for one more team, offering the same government supported GPUs, data and other resources, and said support would continue through 2027 with two teams expected to remain by the end of 2026. It chose not to publish scores or rankings beyond LG's.

Second evaluation, August 2026. Digital Today reported on 18 August 2026 that SK Telecom, LG AI Research and Upstage cleared the second evaluation, and that Motif Technologies, which entered through the extra call, did not advance. The models named are SK Telecom's A.X K2 at 688 billion parameters, LG AI Research's K-EXAONE 2.0 at 750 billion parameters and Upstage's Solar Open 2 at 250 billion parameters. Scores combined benchmarks (40 points, including the Artificial Analysis Intelligence Index and an NIA benchmark), expert review (35 points) and user evaluation (25 points), where members of the public and experts used the models directly. A parameter is one of the learned numbers inside a model, and the count is a rough guide to size, not to quality.

The same report says the ministry would raise GPU support for the three teams from 768 Nvidia B200 units in the first half of 2026 to 1,000 units in the second half. That is a concrete answer to how national compute reaches the model teams: a fixed allocation of leased, government supported GPUs per stage. No monetary amount was given, and the date of the third evaluation was not announced. The earlier Financial News article of 5 April 2026, which said SK Telecom had reached the second phase with A.X K1, fits this sequence.

Why does the programme want shared GPUs? Training a large model needs a lot of computing power for a long time, which is expensive for a single company. Nvidia's release says researchers, enterprises and startups can use the sovereign infrastructure to build models and applications. That is a stated intention, and the release gives no rules on allocation, pricing or access. Beyond the per stage GPU allocations to the model teams, public rules for rationing the wider shared capacity had not been published as of 19 October 2026.

Money is part of the story. The Korea Times reported on 29 August 2025 that the cabinet proposed 10.1 trillion won for AI in the 2026 budget, more than triple the 3.3 trillion won of 2025, inside a record 728 trillion won budget. The same report says the government planned to mobilise 100 trillion won of growth funds over five years, including 1 trillion won in 2026. Those are proposal figures; the budget the National Assembly passed is not cited here, and the growth funds are not described as AI only. Financial News adds that the government's aim is to be "top 3 in artificial intelligence", and describes an "AI Highway" strategy as a campaign pledge of the President, alongside a target of about 100 trillion won in private investment over five years in AI data centres and power grids.

Layer 3: home-grown chips

Most AI accelerators in the programme are Nvidia products. The government also wants Korean chips used. Light Reading reported on 30 June 2026 that the planned AI data centres are meant to test and deploy neural processing units (NPUs) from local startups, naming FuriosaAI and Rebellions, and that the government plans to export turnkey AI data centre designs. An NPU is a chip designed specifically for AI calculations. The article says nothing further on volumes or timing.

Rebellions shows what that layer looks like in money terms. The Next Web reported on 30 March 2026 that the company, founded in 2020, closed a 400 million dollar pre-IPO round at a valuation of about 2.34 billion dollars, led by Mirae Asset Financial Group and the Korea National Growth Fund, which put in 250 billion won. That brought its total funding to 850 million dollars, after a 250 million dollar Series C in September 2025. The report says Rebellions merged with Sapeon, SK Telecom's AI chip unit, in 2024, and that its Rebel100 NPU uses a chiplet design, meaning several smaller dies joined in one package, linked by the UCIe interconnect standard and paired with HBM3E memory. The company also showed rack and pod platforms built around the chip. An IPO is planned, with no date set.

Two links between layers show up here. The memory on Rebel100 is HBM, the subject of our post on Korea's memory makers, and the chief executive is quoted in the same report saying memory "is not very easy to get". So the Korean chip startup competes for the same scarce memory as the GPU makers. The state's National Growth Fund also appears as an investor, which shows how public money flows into the chip layer.

Funding and valuations of private companies are unaudited and change. For FuriosaAI, this post relies only on Light Reading's mention. For a wider view of the accelerator market, including custom chips for training and inference, see our AI chip market share explainer.

The rules: the AI Basic Act

Korea also has a national AI law. Cooley's summary of 27 January 2026 says the Act on the Development of Artificial Intelligence and Establishment of Trust, the AI Basic Act, took effect on 22 January 2026. MSIT was finalising enforcement decrees, which set technical detail.

According to Cooley:

• All operators must tell users when AI generated sound, images or video are hard to tell apart from human made content, and must label content produced by generative AI.
• High-impact AI is AI that significantly affects human life, safety or fundamental rights. Cooley lists healthcare, energy, transportation, hiring and biometric analysis. Operators must assess whether a system qualifies before deployment, explain outcomes in a meaningful way, protect users, keep human oversight and document their safety measures.
• High-performance AI is defined in an MSIT legislative notice as AI trained with cumulative compute of at least 10 to the power of 26 floating point operations. Operators may be required to run a risk management plan and report to MSIT, and Cooley says further requirements are expected.
• The Act applies to AI systems outside Korea that affect Korean users or markets. A foreign operator with no Korean office must appoint a local agent if its prior year revenue exceeds 1 trillion won, its AI service revenue exceeds 10 billion won, or it has more than 1 million average daily users in Korea in the three months before the prior year ended.
• Administrative fines are capped at 30 million won, about 21,000 US dollars in Cooley's conversion, and MSIT has said it will give businesses a year before imposing them. Cooley does not say when that year starts.
• Governance includes a National AI Committee chaired by the president, an AI Policy Center and an AI Safety Research Institute.

For a company building on the sovereign stack, the practical point is that the Act sits on top of the infrastructure. A model trained on national GPUs and offered to users in Korea can still carry transparency duties. Whether a given product is high-impact depends on how it is used, and companies should read the decrees and take legal advice instead of relying on a summary.

What is firm and what is uncertain

Firm: Nvidia published the five line split of the over 260,000 GPUs; the foundation model project started with five named teams and, after the January and August 2026 evaluations, three remain; the AI Basic Act is in force.

Uncertain: how many GPUs are installed; where the government's share is deployed; when the third evaluation happens and which two teams will remain; how the wider national compute will be allocated beyond the model teams; whether Korean NPUs reach large deployments; and private company valuations. The evaluation results here come from press reports of MSIT announcements, so ministry level detail should be checked against MSIT's own notices.

What to watch next

• MSIT's official count of GPUs installed, and the opening of the National AI Computing Center.
• The third evaluation of SK Telecom, LG AI Research and Upstage, and the narrowing to two teams.
• Published rules for sharing national compute beyond the model contest.
• Whether Korean NPUs run production workloads in the planned data centres and at what scale.
• Enforcement decrees under the AI Basic Act and the end of the fines grace period.
• How the National Growth Fund is allocated, and whether Rebellions sets an IPO date.

For teams reviewing the models and packages these programmes add, Safeguard offers SBOM and dependency scanning, with a free scan at https://gold.safeguard.sh/scan.

Quick check

Sources

• Nvidia, press release on Korea sovereign AI infrastructure (30 October 2025)
• Data Center Knowledge, South Korea to boost AI infrastructure with 260,000 Nvidia chips (3 November 2025)
• Light Reading, South Korea plans massive 18.4GW AI data center buildout by 2035 (30 June 2026)
• Fnnews, Korea AI data centre policy and power (5 April 2026)
• Cooley, South Korea's AI Basic Act overview and key takeaways (27 January 2026)
• Inven Global, LG AI tops Korea's sovereign AI project as Naver Cloud fails to advance, MSIT confirms (15 January 2026)
• Digital Today, SKT, LG and Upstage advance to third evaluation of sovereign AI (18 August 2026)
• The Korea Times, Korea to more than triple AI budget in 2026 (29 August 2025)
• The Next Web, Rebellions closes 400M pre-IPO round at a 2.34B valuation (30 March 2026)

Last reviewed: 19 October 2026.