This is Part 2 of Korea’s AI Bet, a five-part series on how Korea is treating AI as a national project rather than a personal gadget. Part 1 argued that Korea decided AI is infrastructure. This part asks the harder question: what does it actually take to build that infrastructure — and pay for it?
“Free AI for every citizen” is an easy sentence to say. It is an extraordinarily expensive thing to build. Behind that one promise sits a stack of at least eight separate construction projects — money, a model, chips, power, data, people, law, and a way to actually deliver it — and most countries attempt only one or two of them. On June 28–29, 2026, Korea did something few governments have dared: it put a number on nearly all of them at once, unveiling its largest-ever AI and semiconductor package — a national plan reported at roughly $576 billion when government and private commitments are combined, including some ₩800 trillion (~$518 billion) in new chip fabs from Samsung and SK and a separate ₩550 trillion (~$356 billion) AI data-center build-out aiming to lift capacity from 8.4 gigawatts in 2029 toward 18.4 gigawatts by 2035. (Headline figures vary by outlet because they mix corporate capex, national totals, and decade-long cumulative spend; treat the components, not the single number, as the signal.)
The headline question is who pays for all this. The more revealing one is what the money actually buys — because “free AI for everyone” rests on at least eight separate bills, and how Korea covers each is where it parts ways with every other country. Here is the anatomy.
The eight things a country must build to give away AI
Strip the politics away and “free national AI” is an engineering and economics problem with eight inputs: (1) money to pay for it sustainably; (2) a model to run; (3) compute — the chips and servers; (4) power and data centers to house and feed them; (5) data to train on; (6) talent to build and maintain it; (7) law and governance to set the rules; and (8) a distribution layer to actually put it in citizens’ hands. Skip any one and the promise collapses: a brilliant model with no power to run it is a press release; abundant compute with no talent to wield it is a warehouse of idle silicon. Most governments pick the one or two inputs they are already good at and call it an AI strategy. Korea is the rare case attempting all eight simultaneously — and the way it does each one, and where it quietly falls short, is where the real story lives.
Money: a country without a sovereign fund mobilizing its own citizens
Korea’s headline is a five-year ₩100 trillion (~$72 billion) AI push, with the 2026 AI budget alone at roughly ₩10.1 trillion (~$7.3 billion) — a record, up more than 200% year on year. But the structure matters more than the size. Of the ₩100 trillion, only about ₩35 trillion (~$25 billion) is direct government money (₩30 trillion national, ₩5 trillion local). The other ₩65 trillion (~$47 billion) is private capital the state intends to induce — not spend.
How do you induce ₩65 trillion? Not the way the United States does it (pure private capital chasing returns, as in the $500 billion Stargate venture) and not the way China does it (the state simply orders it). Korea’s mechanism is a third path: citizen-participation infrastructure funds with profit-sharing and tax incentives, backed by policy finance that absorbs first losses so private money feels safe following the government in. In plain terms, a country with no oil and no sovereign wealth fund is mobilizing its own citizens’ savings and pension capital to build AI capacity. Resource-rich states write a check from a national fund; the United States lets private markets price the risk. Korea, lacking both, has engineered a middle structure where the state takes the first loss so that ordinary investors will take the next one. It is an unusually democratic way to finance a national bet — and also its weakness: most of that ₩65 trillion is projected, not banked, and it only materializes if private investors actually find the returns attractive once the subsidies thin out.
The model: a state-run tournament, not a national lab
To give citizens AI for free, you need a model you control. Korea’s answer — branded “AI for All” (모두의 AI), built on a homegrown model line — is being produced in a way that is itself distinctive. The government did not stand up a single national laboratory. It ran a competition. The Ministry of Science and ICT selected several private consortia in 2025, then narrowed the field through staged evaluations — Naver Cloud and NC were cut in early 2026; a newcomer, Motif, was added — leaving a working set of teams (LG AI Research, SK Telecom, Upstage, and Motif) competing in six-month cycles toward a final shortlist around 2027, supported by roughly ₩530 billion (~$380 million). This competition’s winning model is meant to power the separate, free public-facing “AI for All” service; free access under that banner is planned through 2028 under government funding, after which it is meant to transition to industry.
The design choice is the point: competition among private champions, then selection and state support, rather than a monopoly lab — and a homegrown model rather than a rebranded foreign one, which is what gives Korea data sovereignty over the system millions will use. The honest caveat: the model’s real-world performance against the global frontier is still unproven, and the launch date remains a moving target.
The hardware paradox: owning the chokepoint, renting the engine
Here Korea is both uniquely strong and quietly dependent. The world’s most advanced AI runs on a specialized memory called HBM, and Korean firms — Samsung and SK Hynix — make the large majority of it (a share often cited near 90%, though Micron’s gains have pulled the combined figure into the high-70s-to-80s range in 2026). In effect, no one else’s AI works without Korean parts. That is a genuine, hard-to-copy chokepoint.
And yet Korea imports the engines that memory feeds. The accelerators — Nvidia’s — are bought, not built: a nation-wide deal for roughly 260,000 Blackwell GPUs, split across the government (~52,000), Samsung, SK and Hyundai (~50,000 each), and Naver (~60,000). So the paradox: Korea manufactures the memory the world’s GPUs cannot run without, then buys those GPUs back. The dependence even loops inward — a large share of SK Hynix’s revenue now tracks a single customer, Nvidia. And it is exposed: when Washington restricts frontier access — as it did in mid-June 2026, cutting off the very top tier of some US models worldwide — even close allies feel the gate.
This is the part that explains every other country’s strategy too. No nation is self-sufficient across the full AI hardware stack, so national AI is really a procurement portfolio: own the chokepoint you happen to hold, buy what you can’t make through alliances, and build what you can control at home. China, cut off by sanctions, is forced into full self-reliance — its best domestic accelerators perform at a fraction of the top Western chip, and without access to the most advanced chip-making equipment it is stuck a process generation behind. The United States owns most of the stack — chips, frontier models, and capital — and is genuinely weak only in manufacturing, which it is now paying others to fix. Taiwan holds an entirely different chokepoint: TSMC’s leading-edge logic, without which even Nvidia’s designs are just blueprints. Each country, in other words, bets the hand it was dealt. Korea’s distinct hand is memory plus a willingness to buy and partner on its own terms — and the June 28–29 announcements are, in effect, the country doubling down on that hand: more domestic fabs to defend the memory chokepoint, more imported compute to cover the gap, more data centers to host it all.
Power and data centers: the bottleneck nobody puts on the poster
Money and chips are useless without somewhere to plug them in, and this is where Korea’s plan meets physics — and geography. The AI Data Center Industry Promotion Special Act (the AIDC Special Act), passed on May 7, 2026 and taking effect around February 2027, fast-tracks permits and steers data centers out of the capital region, tellingly exempting non-capital data centers from grid-impact assessment. The June 28–29 package put that geography on a map, pointing much of the build at the southwest (Honam): Samsung’s roughly ₩425 trillion (~$308 billion) regional commitment is anchored by a new chip fab in Gwangju’s Cheomdan-3 district, bundled with an AI data center and solar-and-hydrogen energy (SK Hynix is still weighing a site in Jangseong, South Jeolla, against going abroad), and the government has promised 6.3 gigawatts of power to the region’s semiconductor cluster. Separately, the government’s flagship national AI computing center is going up at Solar-Sido in Haenam — a Samsung SDS-led consortium (with Samsung C&T, Naver Cloud, Kakao and others), ₩2.5–2.9 trillion (~$1.8–2.1 billion, figures vary by outlet), 15,000-plus GPUs by 2028, with an OpenAI–SK data center slated for the same site. Tellingly, Solar-Sido was chosen precisely because it sits on a renewable-energy base, and nearby the 8.2-gigawatt Sinan offshore wind complex (₩48 trillion, ~$35 billion, targeted for 2030) anchors the region’s clean power.
That geography is itself contested. The government frames the southwest tilt as balanced regional development — chip fabs in Honam, packaging in the Chungcheong region, materials-and-equipment hubs in the southeast — but the opposition calls it political favoritism, noting that Honam is the administration’s traditional base, that chip-site selection normally takes five to seven years rather than weeks, and that one group chairman had reportedly been cool on a Gwangju fab just two months earlier. Politicians from the southeast (Daegu–Gyeongbuk) and Chungcheong have pushed back hard. For a plan that needs to survive across administrations, turning AI infrastructure into a regional-spoils fight is its own kind of risk.
Read the fine print, though, and the bottleneck is real — just relocated. Most of that wind is still on paper: only about 1% is generating today, and renewables remain under ~11% of the national mix (the official plan reaches ~21% only by 2030). The deeper constraint isn’t generation but transmission: the southwest already produces more electricity than it can ship to where the demand is, curtailing the equivalent of roughly nine nuclear reactors’ worth of output. The government itself concedes renewables alone won’t cover AI’s appetite, adding nuclear reactors and a small modular reactor to the data-center plan. So the honest version isn’t “Korea will run AI on clean energy”; it’s “Korea is betting it can wire a renewables-plus-nuclear southwest to demand faster than its grid has ever managed.” Power is the hardest of the eight inputs — the one money builds slowest. A fab takes a few years; a grid that can feed 18 gigawatts of always-on compute, sited where communities accept it and clean enough for global customers, is the work of a decade.
Data and talent: what money can’t buy quickly
Two inputs resist being bought. The first is data. A sovereign Korean model needs Korean-language training data, and Korean is a “mid-resource” language — far less text than English or Chinese — so some public datasets even lean on translation. Korea actually ranks first in the world for volume of open public data, but quality is still maturing, and there’s a legal hole: Korea’s Copyright Act has no text-and-data-mining exception, leaving model training in a copyright gray zone while an amendment waits.
The second, and sharper, is talent. Korea ranks first in the world in AI patents per capita — two years running. Yet on net talent flow it sits near the bottom of the OECD: more AI researchers leave than arrive, and the outflow is widening (the number of PhD-level researchers intending to emigrate rose from 592 in 2023 to 709 in 2025). One flagship university left three-quarters of its graduate science-and-engineering seats unfilled in 2025. The driver is a salary gap of roughly four to one against the US, widening at the senior level. The government’s counter-funding to retain researchers is real but small — on the order of a single large US lab’s annual talent budget. You can fund chips and fabs in a single announcement; you cannot manufacture senior researchers on the same timeline, and you certainly cannot outbid Silicon Valley for them line by line. This is the quietest threat to “AI for All”: a country that leads the world in producing AI patents while steadily losing the people who would build the national model those patents describe.
Law, safety, and the last mile
The seventh input is the rulebook, and Korea wrote a distinctive one. Its AI Basic Act — in force since January 22, 2026, Asia’s first comprehensive national AI law to actually take effect (the EU’s AI Act came first globally) — leans toward promotion plus trust rather than prohibition, with a one-year grace period and penalties capped around ₩30 million (~$22,000) — a small fraction of the EU’s. Above the agencies sits a Presidential Council on National Artificial Intelligence Strategy (its official site, aikorea.go.kr) whose recommendations carry real weight. On safety, the law mandates labeling of AI-generated content, and in mid-June 2026 the Ministry of Science and ICT signed AI-safety and cybersecurity arrangements with frontier labs opening up in Seoul. The trade-off is candid: a promotion-first law with light penalties raises real questions about deterrence.
The eighth input — distribution — is the one Korea is best equipped for and most likely to underestimate. The country has world-leading broadband and smartphone penetration (nationally above 90%, though usage among people in their sixties is markedly lower — smartphone ownership near 80%, broadband/internet use closer to 60–65%), so the last mile is, technically, already built, and inclusion for older and vulnerable users is an explicit design goal. But access is not the same as use: Korea’s own retreat from AI digital textbooks in 2025 showed that putting a tool in front of people doesn’t guarantee they adopt it. And it’s worth being precise about what already exists versus what’s promised — for example, Seoul’s AI program for lower-income residents is a city-level pilot for a vulnerable bracket, not the nationwide free service still to come.
What Korea is actually doing differently
Step back and the pattern is clear. Korea is not first in any single one of the eight inputs — the US has more capital and compute, China more forced self-reliance, others more energy. What is close to unique is that Korea is attempting all eight at once, as a hybrid: owning the one chokepoint it holds (memory), buying and partnering for what it can’t make (GPUs, frontier safety know-how), building a model of its own, and then setting universal, inclusive, free access as the actual goal. Most major powers pursue the first six inputs to win a market or a race; very few pair “build the whole stack” with “give it to everyone.” That combination — chokepoint ownership + diversified procurement + sovereign model + inclusion as the objective — is the genuinely Korean move.
It is also a bet that can lose, and a fair reading has to hold both sides. The optimistic case isn’t naive: Korea has done this before — it treated broadband as infrastructure in the early 2000s and seeded gaming, fintech and the K-content wave on top of it — and for an economy whose working-age population is shrinking fast, the IMF estimates AI could lift potential growth by up to roughly 0.4 points. Those are real stakes, not slogans. The skeptical case is just as concrete: Korea trails badly on raw compute scale, its homegrown model is far likelier to land in the world’s upper tier than at the very frontier (the government’s own working target is reportedly a top-tier-but-not-number-one model), its researchers are leaving, much of the ₩100 trillion is induced rather than spent, and the power isn’t secured. The structure also leans on a ninth input the eight-part list understates — international alliances — because the GPUs and frontier safety know-how must be bought or partnered for, on terms a shift in US policy can tighten overnight; and a plan this centralized is fast but fragile, where a single weak link or a change of administration before 2028 could stall it. The verdict analysts from Carnegie to Brookings converge on is that full AI sovereignty is a fantasy and resilience is the realistic prize: Korea is unlikely to become literally self-sufficient, but it has a credible path to a durable upper-tier role — borrowing the frontier while owning the chokepoint. Whether capital can close the gaps in talent and electricity as fast as it closes the gap in fabs is the question the next two years will answer.
What it means if you build in Korea
For a founder or resident watching from outside, the practical reading is two-sided. The upside: the cost floor for using AI in Korea is being pushed toward zero, the country is wiring more compute and (eventually) a free national model into the environment, and the talent pool — if it can be kept — is unusually AI-fluent. If you’re weighing the move, our free-AI guide maps what a foreign resident can actually use today, and the D-8 founder visa guide covers the entry route. The caution: the marquee pieces — the homegrown model, truly universal free access, secured power — are still ahead, not behind, so build your plans on what exists now, not on the press-release timeline. Part 1 of this series explains why Korea treats all of this as infrastructure rather than a product; this part is the bill for that decision — and Korea has just signed it.
Stay updated
We track how Korea’s AI bet unfolds — what gets funded, what launches, what slips, and what each move means for foreigners building here. No spam, just the updates that matter.
Disclaimer: This post reflects the author’s experience and publicly available information as of 2026. It is general information, not legal, tax, or immigration advice. Rules and rates change — verify current details with the relevant authority (NPS, NTS, MOJ) or a licensed professional before acting.
