韩国向5200万人提供免费无限AI，但真正关键在于80%国产模型配额

South Korea just did something no other country has done Let's start with the facts—I checked these one by one (sources cross-verified across multiple credible outlets, linked at the end): South Korea's Ministry of Science and ICT (roughly equivalent to a tech ministry) has finalized the three operating consortiums for its "AI for All" program: SK Telecom, KT, and Kakao. Beta starts in September, full launch by year-end. The core pitch is two words: free, unlimited. Doesn't matter what you earn, how old you are, or whether you're tech-savvy—all 52 million people in South Korea get direct access to generative AI, and officials are explicit: "no token limit." The government has allocated 512 NVIDIA B200 chips to these three companies this year, and starting in 2027, the national budget will even absorb part of the cost of running this nationwide service. That makes Korea the first G20 country to do this. (Quick reality check here: "unlimited" is the official claim, not physical reality. Feeding 52 million people off 512 chips necessarily means tiered rate limiting, or dumping most requests onto lightweight models—there's no way everyone gets truly unlimited access. Finding that hidden trapdoor is exactly what I plan to test myself later.) Every headline is chasing "free" and "unlimited." But free and unlimited is just the hook to get 52 million people flowing in—the real move that turns that traffic into a moat is buried in the next section. The real catch: that 80% There's a rule that every article glosses over, and it goes like this: Every operator must route at least 50% of queries to its own "certified Korean sovereign foundation model," and at least another 30% to other Korean companies' models. Combined, the floor for domestic models is 80%. I saw a viral social post claiming "at least 50% on domestic models"—that number is way too conservative. The actual floor is 80%. This is the soul of the entire program. "Free and unlimited" isn't charity—it's an incentive. The government is essentially saying: I'll provide the compute, the money, the gateway, and funnel tens of millions of citizens through it in one shot—but the AI you use has to be Korean, eight times out of ten. In other words, Korea isn't running a "welfare program"—it's running industrial policy: using everyday national usage to forcibly prop up the demand side for domestic models. And from an engineering standpoint, that 80% lines up neatly with the earlier question of "how do you feed 52 million people with 512 chips?" The sensible architecture is a smart router up front that judges difficulty: the huge volume of repetitive, trivial requests gets offloaded to cheap domestic small language models (SLMs), and only genuinely hard queries get escalated to premium models. To be clear—512 chips is just seed compute for a pilot; serving 52 million people almost certainly requires tiered rate limiting as a matter of physics. SLM routing isn't what makes it "truly unlimited"—it's the lever that determines how far the free tier can stretch. And here's where the policy quota (80% domestic) and the cost structure (SLM routing saves compute) line up perfectly: Korea doesn't need domestic models to beat GPT, it just needs cheap-and-good-enough domestic small models to absorb 80% of citizens' daily usage—and the compute that saves is what funds the free tier. Why do this? Understand "sovereign AI" and it all makes sense This looks strange in isolation, but drop it into the bigger 2026 trend and the logic clicks. By 2026, "sovereign AI" has gone from a conference buzzword to a line item in G20 national budgets. Global spending on this is projected to blow past $100 billion: France launched its "Projet Voltaire" sovereign cloud, deploying 50,000 NVIDIA H200 chips in the first wave, with plans to double that by 2027. India selected 12 domestic foundation models from over 500 proposals, backed by 38,000 GPUs. Canada committed C$925 million over five years, and the EU mobilized €20 billion for AI Gigafactories. The reasons countries are racing for sovereign AI boil down to three things: fear of supply chains being choked off by frontier US labs, the need to protect data sovereignty under local language and regulatory requirements, and treating AI as a strategic industry worth nurturing. But here's the key difference—most countries are fighting on the supply side: building data centers, stacking GPUs, training their own models from scratch. Korea's move is unusual in that it's attacking the demand side. A model that nobody uses is dead, no matter how good it is. Korea's read is honest: domestic models might catch up technically, but they can't out-compete the user base and habits ChatGPT has already built. So what do you do? The government pushes 52 million people into the domestic ecosystem with free, unlimited access, creating something domestic models need that ChatGPT can't give them—massive, real, everyday usage. Everyone's stacking on the supply side. Korea is the only one so far playing this bold on the demand side. What's actually in Korea's hand Two lists need untangling here, or things get confusing fast: the service operators are the SKT/KT/Kakao consortium (responsible for delivering AI to citizens), while the "sovereign foundation models" they're required to route to come from a separate, parallel national competition. The two lists overlap (SKT is on both), but the roles differ—one is "the pipe that delivers AI to your front door," the other is "the water flowing through that pipe." That "sovereign foundation model" national competition is currently in elimination rounds—narrowed from five teams to four, with only two surviving by 2027. The visible players include LG's EXAONE, Naver's HyperCLOVA X, Upstage's Solar Pro, SK Telecom's A.X series, and Kakao's Kanana. Right now, LG AI Research, SK Telecom, and Upstage are leading into the third round. SK Telecom just unveiled A.X K1—Korea's first hyperscale model at the 519-billion-parameter class, clearly aiming for a top-three spot globally. One detail worth noting: Naver got cut in January—not because the tech was bad, but because HyperCLOVA X had "data independence" concerns (it used frozen encoder weights from Alibaba's Qwen). In the logic of sovereign AI, "how pure is your model's bloodline" apparently matters more than "how high does it score on benchmarks." The 80% domestic quota in "AI for All" is what feeds this exact pool of national-team models. The government is running a competition to pick sovereign models on one hand, while guaranteeing those models real traffic through free, nationwide service on the other—supply and demand, worked on simultaneously. That's the whole play. So what does this mean for a regular person building with AI? If you're a freelancer, a solo founder, or anyone making a living with AI, don't file this away as unrelated foreign news. I think there are three concrete takeaways worth internalizing: One: Korea's "small models handle the grunt work" architecture is exactly the cost-cutting move you should be copying. As I broke down above, Korea can only afford "unlimited for everyone" because domestic small models absorb 80% of requests. The same logic applies directly to a solo operation—stop routing every single task to the most expensive GPT. The right move is to build a hybrid routing layer up front: dump the grunt work—classification, extraction, rewriting, first drafts—onto cheap small models (even locally-run, heavily-subsidized open-source ones), and reserve the premium API calls for the truly hard 20% that needs real reasoning. This isn't magic—open-source tools like RouteLLM and LiteLLM exist specifically for this. One routing layer, and your cost for the same output can drop several times over. Korea just demonstrated this architecture with a national budget; you can use it to cut your own costs. Two: the moat is shifting upward, toward your data and your vertical know-how. When a country turns "having access to generative AI" into a public utility like tap water, "I have a ChatGPT account" stops being an advantage. Everyone has a generic chatbot. What's actually valuable is solving a specific problem in a specific industry, fed by data nobody else can get their hands on. Worth noting: Naver got cut from the sovereign model national team not because its tech was weak, but because of doubts about its data lineage. "Where does your data run, and whose regulations does it answer to" is a question that's going to come up more and more from clients and governments alike, no matter what you're building. Three: the real frontier is "agents that get things done," not "AI that chats well." The most important design choice in Korea's service is that it plugs directly into government systems to handle tasks on your behalf: booking medical appointments, filing taxes, searching for housing. That's not Q&A—that's "carrying a task from start to finish." But from my own experience building agent systems, the moment an agent can actually take real action (submitting an application, placing an order, deducting a payment), the cost of a mistake jumps from "answered one question wrong" to "did the wrong thing." So the hard part of task automation was never getting a model to talk well—it's plugging it into real systems without causing damage. The real engineering challenge is permission boundaries and rollback, not how smart the model is. Whether this nationwide task-automation rollout in Korea breaks anything is exactly what I want to watch most closely. I'm going to test this myself—here's what I'm verifying No amount of policy analysis beats actually using the thing. I'm in Korea, so the moment the September beta opens, I'll sign up and test it myself, then write up the first-hand results in a follow-up piece. Here's what I'll be watching (also listed below in case you're in Korea and want to test alongside me): Is "unlimited" really unlimited, or is there a hidden rate limit? Free things usually have a trapdoor somewhere. Which model does it actually route to, and how much worse is the quality compared to ChatGPT? I'll throw the same prompts at it in Chinese, English, and Korean. With 80% running on domestic models, can users actually tell the difference? Does the task-automation feature actually work? For the parts that plug into government systems—booking a doctor's appointment, checking your tax filing—can it actually complete the task end to end, or does it just spit out an explanation? How friendly is the UX for people who aren't tech-savvy? This program claims to serve "people who can't install software"—that's the claim most likely to turn out to be empty. What data does it collect from me? The flip side of plugging into government systems is privacy. Whether a country handing AI to its citizens as a public good is a smart move won't be clear for another two or three years. But the direction is obvious: the AI competition is shifting from "whose model is strongest" to "who can get the most people using AI, every day, with the least friction." Korea is betting on the latter. I'll report back after testing. Next piece: is this gamble real, or is it just for show? Sources (cross-verified across multiple outlets): Korea Herald — Korea picks SK Telecom, Kakao, KT to build free nationwide AI services The Korea Times — SKT, KT, Kakao consortiums selected for free AI service for public TechSpot / Decrypt — free access with no token limits The Next Web — 52 million citizens, domestic models KED Global / Korea Herald — sovereign foundation model national team (LG, SKT, Upstage, etc.) Light Reading — South Korea enters second phase of sovereign AI project References 韓國政府砸 10 兆韓元要請全民免費用 AI：不限 Token 吃到飽，扶植本土模型 | 動區動趨-最具影響力的區塊鏈新聞媒體 韓國將為所有 52 百萬公民提供免費 AI 訪問權限 韓國推出「全民 AI」：為 5,200 萬人提供免費聊天機器人使用權 | Gate 新聞 Originally published at Judy AI Lab. Visit for more articles on AI engineering and development.