Machinepower Index / Dossiers / South Korea
Rank 2–3 across the scoring rules tested. How this is tested
South Korea
Chokepoint Leverage — For nations holding a step in the chain nobody can route around, which their government could switch off: trade access for compute, investment and protection, and stay irreplaceable.
How the Score, Potential and bottleneck cost are worked out: methodology.
The three layers
Energy, grids, data centres and silicon — the physical capacity to run models at scale.
Compute tracked by Epoch AI: South Korea controls 5,200 H100-equivalents, wherever the sites are; has 1,100 on its soil with the owner not reported. Hosted compute is shown, not scored.
Models, alignment capability, standards influence and the talent that produces them.
The capacity to wire AI into the state and the economy while holding public consent.
Where the score comes from
| Layer | Sub-dimension | Basis | Score |
|---|---|---|---|
| Watts | Planning & Permitting | Assessed | 7.4 |
| Firm Power Capacity | Dataset | 6.4 | |
| Sovereign Compute | Dataset | 7.0 | |
| Hardware & Silicon Access | Dataset | 10.0 | |
| Weights | Frontier Model Output | Dataset | 6.7 |
| Alignment Capability | Assessed | 7.4 | |
| Regulatory Standing | Dataset | 9.1 | |
| Talent Density | Dataset Sources differ | 5.8 | |
| Will | Policy Capacity | Dataset | 9.6 |
| Public Sector Adoption | Dataset | 7.9 | |
| Workforce Transition | Dataset | 8.1 | |
| Public Trust | Dataset Sources differ | 6.8 |
Assessed marks a measurement we judged rather than took from a dataset: 2 of the 12 here.
Sources differ marks a cell where two sources disagree by more than a point. Talent Density: Stanford's count and the developer-based estimate differ by more than a point (5.1 against 6.4); the cell is their average. Public Trust: Ipsos and Melbourne/KPMG differ by more than a point (7.7 against 5.9); the cell is their average.
Sources
Share of adults agreeing that products and services using AI have more benefits than drawbacks; the Index scores public trust as 1 + share ÷ 10. Survey of 23,532 adults across 32 countries, March to April 2026.
View source →Announced at APEC (October 2025): ~50K GPUs for the National AI Computing Center plus allocations to Samsung, SK, Hyundai and Naver Cloud AI factories. The release names no total programme cost; Hyundai's own commitment is ~$3B.
View source →The 29 October 2025 memorandum pairs coordinated AI export approaches with cooperation between the US Center for AI Standards and Innovation and Korea's AI Safety Institute. The $350B investment figure belongs to the separate trade package, not this document.
View source →Korea remains one of few nations maintaining a competitive local-language foundation model ecosystem with dominant regional search share.
View source →Used for the workforce measurement. The IMF combines schooling, pay, labour mobility and active labour-market policies; the Index rescales it so the weakest of 173 economies scores 0 and the strongest 10.
View source →A cross-check for planning and permitting, which stays assessed: it scores building permits and electricity connections for ordinary firms, not grid-scale connections for data centres.
View source →Fossil, nuclear, hydro and other dispatchable plant; averaged with generation for firm power.
View source →Second source for talent, averaged with Stanford's count.
View source →Second source for public trust, averaged with Ipsos.
View source →Capital is shown beside the Score, not scored.
View source →On the register — what the world cannot get anywhere else
South Korea holds one of the register's chokepoints: a position nobody else can route around, which its government could actually shut off. See the whole chain.
Samsung and SK Hynix held roughly four fifths of high-bandwidth memory output in 2026, down from almost all of it in 2024. The gap went to Micron, whose supply chain touches Korea nowhere: memory from Taiwan and Japan, packaging in Taiwan and, from 2027, Singapore. An accelerator without this memory is not an accelerator, and shortages have already cut Nvidia's 2026 shipment forecasts. This is the weakest entry here and it is getting weaker: Seoul designates the stacking technology as strategic but exempts the finished chips from export control, so every restriction ever applied to Korean memory has been an American one.
TrendForce →Bottlenecks
These hold South Korea below its Potential. Its two lowest scores cost it the most: talent density at 5.8 (part of Weights) and firm power capacity at 6.4 (part of Watts).
Assessment
A material hardware titan in high-bandwidth memory (HBM) and microchip fabrication with Samsung and SK Hynix. Its primary vulnerability is dependence on US model ecosystems and software weights, alongside heavy energy demands for hyper-foundries.
The art of the possible
Korea already makes the memory the frontier runs on, and that is the position to deepen rather than dilute: high-bandwidth memory now binds frontier training as tightly as logic does, and two Korean firms supply almost all of it. The 260,000-GPU build-out and HyperCLOVA-class domestic models are worth having because they make that position harder to bypass, not because Korea needs one of everything. The trade is simple: memory supply for frontier access and security guarantees, priced while two firms still hold the market.
Strategic play — Chokepoint Leverage
Decide whether the memory chokepoint belongs to Seoul or is on loan from Washington. Two Korean firms supply almost all of the high-bandwidth memory frontier training depends on; price that in frontier access and security guarantees, and build domestic models only where they make the position harder to bypass.
Leverage — the nearest position. Holds the memory chokepoint and has never turned it. Every restriction ever applied to Korean memory was written in Washington.
Open South Korea in the live Index → — every source behind every number, and the comparison tool.