South Korea's Ministry of Science and ICT selected SK Telecom, Kakao, and KT consortia on August 28, 2026, as operators for the national 'AI for All' program, which promises free access to domestic AI models without token limits for approximately 51–52 million citizens. This is the first G20 initiative in which the state subsidizes computing infrastructure to transition the entire population from ChatGPT and foreign services to national models. The beta version will launch in September, full launch is promised by the end of 2026, and from 2027, servicing the entire country will be funded from the budget.

image

What happened

The Ministry of Science and ICT concluded the selection process and appointed three consortia led by SK Telecom, Kakao, and KT as operators of the free national AI service. The state will allocate a total of 512 Nvidia B200 accelerators to them in 2026, and from 2027 will take on the cost of servicing the entire population from the budget. The operators' paths to users are different: SK Telecom is building 'agentic' AI accessible via phone calls and SMS; Kakao, together with LG Uplus, is integrating a chatbot and AI agents into the KakaoTalk messenger; KT is connecting the service to the Daum portal and platforms like Danawa, Musinsa, and Zigbang. All three services promise to integrate with government systems.

Context

The program is designed as a response to the dominance of ChatGPT and Chinese services: Seoul aims to retain AI demand within the country and strengthen digital sovereignty by switching the population to domestic models through free access. The announcement contains no statements about new architectures, training methods, or benchmark results, so 'AI for All' is better read as a state financing and distribution program rather than a research breakthrough. Its essence is that the state subsidizes both layers — computing infrastructure and the channels where users already spend their time: messengers, portals, and telephone networks. Basic access to AI in the country is turning from a paid subscription into a default service.

Why this matters for the industry

For Korean vendors, the program opens a distribution channel of national scale: access to users comes not through a separate app, but through surfaces where they already are — messengers, portals, and telephone lines. Agent integrations with government systems set a ready-made standard for the market of government service integrators. Teams building products for the Korean market should consider the free government channel as a competitor in 'ask — find — book' scenarios from the beta. The anchor price is also changing: consumers get used to free unlimited access, and pressure on paid chat subscriptions will begin before the full launch. If the experiment shows retention and quality, 'AI as a public utility' will turn from a national project into a replicable template and create budget demand for inference and its optimization in other countries.

Why this matters for users

From the September beta, Korean users will get unlimited bots within familiar apps: in messengers, on portals, and via regular calls or SMS. Practical scenarios shift to dialogue: booking a doctor's appointment, figuring out taxes, or finding housing can be done through the same bot, without switching between sites and queues. Subscriptions and token payments are not assumed — access is by default for all citizens of the country. Readers outside Korea will not have direct access, but the experiment itself is worth watching: this is the first case where a state on a national scale is transitioning its population to free national models, and it will show whether 'default government AI' becomes a repeatable model for other countries.

What is still unknown / limitations

The service does not yet exist — neither the beta nor the full launch, so there is nothing to check quality and reliability on right now. Specific models are not named: the developer, model size, training methods, and results on open benchmarks are unknown, so any claims about the quality of domestic models are untestable before the beta. The unlimited promise runs into the unspoken economics of inference: the allocated computing quota for the real demand of the entire country may not be enough without extremely aggressive serving — batching, quantization, speculative decoding, or switching some requests to compact models. The evaluation methodology and program metrics have not been announced in advance, so observation will be closer to operational monitoring than to checking promised claims. A quiet revision of terms is also possible — throttling, traffic prioritization, or rolling back some promises; the original WSJ report is only available by subscription, so some details rely on secondary retellings.

Sources

Author

Look at AI, editorial team