A Different Kind of AI Bet
Most national AI strategies in 2026 are arguments about rules or arguments about models. The EU is racing to enforce the AI Act. The US is fighting itself over federal versus state authority. China is routing around export controls to keep its labs supplied with chips.
South Korea just did something else entirely. President Lee Jae-myung unveiled a 10-year national AI plan worth roughly 1,350 trillion won — about $880 billion — and almost none of it is aimed at building a frontier lab to compete with OpenAI, Anthropic, or Google DeepMind. It's aimed at owning the physical inputs those labs, and everyone else's, already depend on.
Those two sub-totals overlap and compound rather than sitting side by side as clean, separate buckets — official breakdowns of a decade-long, multi-agency plan like this rarely resolve to tidy arithmetic. The signal is the scale and the target, not the precision of the split.
Why Memory, Specifically
To understand the plan, you have to understand where the actual bottleneck in AI sits right now. It isn't raw GPU count. It's high-bandwidth memory (HBM) — the memory chips that sit directly next to every AI accelerator, feeding it data fast enough to keep it from sitting idle.
Samsung and SK Hynix are already among the dominant global suppliers of HBM. That's not a coincidence Seoul is leaning into — it's the entire strategy. South Korea isn't trying to out-build OpenAI's model roadmap. It's positioning itself as the country nobody in the global AI race can route around, regardless of which lab or which model architecture eventually wins.
Physical Bottlenecks Outlast Model Cycles
Frontier model architectures shift every few months — a new release can reshuffle the leaderboard overnight. Memory and compute bottlenecks move on a completely different timescale, measured in years of fabrication capacity and capital expenditure. A country that controls the scarce physical input holds leverage no matter which model wins the next benchmark cycle.
How This Compares to Everyone Else's Strategy
Line up the major AI strategies currently in motion and South Korea's plan looks structurally different from all of them:
- The United States has no federal AI law. States like California, Colorado, and Texas are legislating independently, and Washington is spending political energy on federal preemption fights rather than building state-owned physical capacity.
- The European Union is spending its capital on the EU AI Act's compliance deadlines — high-risk system obligations and general-purpose model provider sanctioning power both land around August 2, 2026. That's a rules-first strategy, not an infrastructure-first one.
- China is racing on chip access through export-control workarounds and domestic substitution, reacting to constraints rather than pre-committing capital at this scale.
- South Korea wrote a check for physical capacity — memory, power, land, buildings — a decade out, and left the model layer entirely to the private labs already competing for it.
The Timing Isn't Random Either
This plan lands just after a chip-sector selloff rattled confidence in AI infrastructure valuations, and days after SK Hynix priced one of 2026's largest AI-related IPOs — a $28 billion listing that gave public markets their first direct read on how investors price the AI memory chip supply chain specifically, separate from GPU makers.
Seoul isn't reacting to that volatility. If anything, the plan is a bet that the volatility is noise and the underlying demand curve — more inference, more agents, more enterprise AI in production — is real and durable enough to justify a decade-long commitment.
There's a Domestic Logic, Too
South Korea has watched a smaller neighbor build an entire geopolitical identity around chip dominance. Taiwan's position in the global economy, and the strategic protection that position buys it, is built almost entirely on TSMC's fabrication lead.
Korea is a small, export-dependent economy with a shrinking working-age population. It doesn't have unlimited paths left to remain indispensable to the global economy for the next generation. Owning a scarce input to the most capital-intensive buildout in the history of the technology industry is one of the few plays still available — and arguably one of the more durable ones.
What to Watch Next
- Whether the memory bottleneck actually holds. If HBM supply loosens faster than expected — through new entrants or alternative architectures — the leverage this plan is built on weakens.
- Whether other governments copy the model. If South Korea's bet pays off, expect other mid-sized economies with a chip or energy advantage to make similar infrastructure-first commitments instead of chasing regulation or frontier labs.
- Whether $880 billion actually gets deployed on schedule. Ten-year national plans routinely survive changes in government, budget cycles, and shifting priorities in name only. The real test is capital actually breaking ground, not the announcement.
Frequently Asked Questions
My Take
Trying to out-build OpenAI, Anthropic, or Google DeepMind from a standing start in 2026 is a losing bet for almost any country — the capital, talent, and data advantages are already too concentrated. Trying to become the supplier nobody in that race can function without is a fundamentally different, and arguably smarter, bet.
Model leaderboards will keep reshuffling every few months for the foreseeable future. Physical bottlenecks — memory, power, land — will not move nearly as fast. South Korea just spent a decade's worth of political capital betting on the slower-moving, harder-to-replace layer of the stack. Whether that's foresight or an expensive bet on demand that cools before the checks clear is the question the next ten years will answer.
Related Articles:
