South Korea's Ministry of Science and ICT has named the operators of its "AI for All" program: three consortia led by SK Telecom, KT, and Kakao beat out six applicants for the contract. The service will be free to every citizen, with no token limit, entering a closed beta in September and launching in full by the end of the year — running entirely on large language models developed in Korea.
The government's compute contribution is 512 Nvidia B200 GPUs, split among the three winners. Starting in 2027, nationwide operating costs will be covered by the state budget. Korea's AI-related budget for 2026 is 10 trillion won, roughly $7.2 billion, triple the previous year's figure.
512 chips don't stretch far per person
Do the rough math first. 512 B200s split three ways comes to a little over 170 chips per consortium. With a population of about 51 million, that works out to 0.1 chips per 10,000 people — a fraction that would look thin next to any frontier lab, where a single large inference cluster can run to tens of thousands of chips.
So the phrase "no token limit" needs to be read in context. Free and unlimited here rests on time-slicing and queuing, not on stacking hardware. The closed beta in September and the push to launch by year-end give the three consortia a bit over three months to find the concurrency level at which the experience starts to break down.
The member lists say more than the model lists
The makeup of the three consortia is more revealing than the model rosters each has published.
SK Telecom's group brings in Liner, Shinhan Card, Hana Card, Tmap Mobility, and SK Broadband, alongside medical-appointment platform Goodoc, education company Elice, and health-data firm Soundable Health. KT's consortium includes chipmaker Rebellions, Upstage, Daum, fashion e-commerce player Musinsa, BC Card, real-estate platform Zigbang, and public broadcaster EBS. Kakao's group pairs medical-AI company Lunit and Seoul National University Hospital with Shinhan Bank, LG AI Research, LG CNS, LG Electronics, and chipmaker FuriosaAI.
Payments, real estate, medical appointments, education, retail — this lineup isn't building a single chatbot; it's laying a conversational front end over public services. The listed use cases back that up: booking a doctor's appointment, finding an apartment, tax questions, corporate tax calculations, checking eligibility for government subsidies, and recommending educational content, all routed through government systems.
Naver is absent from every group. Korea's largest search and AI company sat out its own government's flagship national AI project.
All three are drawing from the same pool of models
The model lineups each consortium announced look varied at a glance, but the overlap runs deep. K-EXAONE 2.0 appears on both SK Telecom's and Kakao's lists; Motif 3 is shared by SK Telecom and KT; the Solar series shows up in two of the three. SK Telecom's flagship, A.X K2, is the largest by parameter count among the three at 688 billion.
Seen from another angle, Korea only has so many home-grown foundation models to draw on, and the three consortia are largely recombining the same pool. What looks like a three-way competition on the surface is built on shared underlying supply.
The market most willing to pay for AI just got a free one from the government
One comparison stands out: 25% of Koreans have paid for an AI service at some point, versus just 2% in the United States. At the same time, more than 20 million Koreans already use free generative AI tools.
In a market with some of the world's highest willingness to pay for AI, the government is turning around and making its own service free. In the short term, that squeezes the subscription business of domestic AI companies. The longer bet is different: get homegrown models into the daily routines of 51 million people first, build user numbers and habits, and worry about revenue later.
Sources: Ministry of Science and ICT project announcement, The Korea Times, CocoLoop, TechSpot; consortium membership, the 512-B200 allocation, and Korea's AI budget figures were cross-checked against two media reports.