South Korea’s Ministry of Science and ICT is expected to name two or three winning operators for its “AI for All” program before the end of August — and the outcome will settle one of the most consequential distribution questions in public AI policy: does government-mandated AI reach more people when it arrives as a new app to download, or when it appears invisibly inside platforms citizens already open every day? KT Corporation placed its bet publicly on August 26, when the country’s largest wired carrier released the full details of its 16-company consortium and identified its answer: go where the users already are.
KT’s strategy is deceptively simple. Rather than routing users to a new AI chatbot — or even to KT’s own “A.” competitor — the telco plans to embed domestic AI models directly into eight consumer-facing platforms that Koreans already use for shopping, studying, searching the web, finding homes, and managing money. The consumer experience does not change: users open Musinsa’s fashion app, or Mathpresso’s Qanda education platform, or Kakao’s Daum search portal, and AI capabilities appear within those familiar interfaces. No download. No new account. No learning curve. The AI is invisible inside existing platforms.
That framing matters because Korea’s own data reveals a stubborn adoption problem. Approximately 23 million Koreans currently use generative AI, meaning roughly 29 million residents — more than half the adult population — remain untouched. Deputy Prime Minister and Science Minister Bae Kyung-hoon, a former director of LG AI Research, framed the gap plainly at the program’s July launch: ensuring AI access for all citizens had become an urgent task, he said, so that the AI divide would not compound into social and economic inequality. KT’s theory is that the reason those 29 million have not adopted AI is not interest or cost — it’s friction. And the way to eliminate friction is to remove the requirement that they ever encounter a new product at all.
Korea’s AI for All: What the Government Is Actually Building
The program’s ambitions are extraordinary by any global standard. South Korea’s Ministry of Science and ICT launched the “AI for All” initiative on July 13, making South Korea the first G20 nation to give all 52 million of its citizens free AI access, with no usage caps.
The program has two phases. The first is a general-purpose AI chatbot with no usage caps — a direct counter to the usage limits that foreign services like ChatGPT impose on free-tier users. The second, targeted for 2027, is more ambitious: a proactive AI agent that scans a citizen’s individual circumstances, identifies government benefits, scholarships, and administrative programs they qualify for, and submits applications on their behalf. The government envisions one AI agent per citizen as the long-term goal.
The structural mandate attached to the program is what makes it technically consequential. Winning operators must route at least 50% of user queries through their own certified Korean foundation model, and at least 30% more through other South Korean-built models — putting the domestic AI floor at 80% of total usage. Foreign models can fill the remaining gap, but only without government subsidy. The Ministry is providing 512 Nvidia B200 GPUs to cover initial compute for the beta launch, with national budget support beginning in 2027 and the service committed free through 2028.
Six consortia submitted bids by the August 11 deadline: SK Telecom, KT, Kakao, ESTsoft, MKD, and Xenon. The ministry confirmed all six submissions on August 18. The selection decision is expected before August 31 — within days of this publication.
How Do You Reach 29 Million People Who Ignored Every AI App? KT’s Answer
KT’s consortium is structured as a deliberate response to the technology adoption problem identified in Everett Rogers’ foundational work on diffusion of innovations. Rogers found that technologies diffuse faster when they are “compatible” with existing behaviors — when adopters do not need to learn new workflows or abandon familiar platforms to access the new capability. The “AI Inside” approach maximizes compatibility by delivering AI zero-friction to users who already visit the partner platforms daily.
In education, the anchor partner is Mathpresso’s Qanda platform, which holds approximately 95 million registered users globally and around 8 million monthly active users. The platform built its position on a distinctive optical character recognition feature: students photograph a math problem and receive step-by-step solutions within seconds. KT and Mathpresso are not new partners — KT invested $8 million in Mathpresso in a prior funding round and co-developed MathGPT, a 13-billion-parameter math-specialized language model, alongside Mathpresso and Upstage in January 2024. Under the AI for All plan, Qanda would extend that foundation into hyper-personalized learning paths that adapt continuously to each student’s progress, questioning patterns, and knowledge gaps. EBS — Korea’s national public broadcaster and educational institution — adds a complementary layer of AI-driven educational services on top.
In fashion and e-commerce, Musinsa brings approximately 17 million registered users and around 8 million monthly active users to the consortium. The platform would add personalized recommendation engines and conversational AI product search, making fashion discovery feel less like a search query and more like a dialogue with a styling assistant. ConnectWave, which operates Korea’s largest price comparison portals Danawa and Enuri as well as the merchant tools platform Makeshop, joins as the e-commerce infrastructure partner.
In information and search, KT’s partnership with Daum — one of Korea’s original and most widely used web portals, now operating as an Upstage subsidiary following Upstage’s May 2026 acquisition from Kakao — would bring AI-powered answer synthesis to a large existing user base. For real estate, Zigbang would use AI to help users research properties and navigate decisions in one of Korea’s most consequential consumer spending categories. BC Card adds AI capabilities to financial services and payments. Saltlux contributes AI language technology, and DeepSearch adds deep financial data analysis.
The model and infrastructure layer below all these consumer touchpoints consists of Upstage (one of Korea’s top-three finishers in the government’s separate sovereign model competition), Motif Technologies, and NC AI on the model side, with ActionPower for voice recognition and Lablup and Rebellions on the infrastructure and chip side.
KT’s Full-Stack Infrastructure Pitch
KT describes its contribution as an AI full-stack approach — domestic AI chips, foundation models, cloud platform, and telecommunications network, all operating as the invisible connective tissue beneath the consumer platform touchpoints.
The chip layer is Rebellions, a Korean AI semiconductor startup in which KT is an investor. Rebellions’ ATOM-MAX inference NPU — already deployed in government public safety infrastructure in South Korea — provides the compute for inference workloads. On the model side, Upstage’s Solar Pro and KT’s own Mi:dm models handle the Korean-language reasoning at the core. KT’s cloud infrastructure provides the API integration layer: a consumer platform sends a query to KT’s orchestration platform, which routes it to the appropriate domestic model and returns the result inside the platform’s native user interface. The user never leaves the app they opened.
The Ministry’s 512 B200 GPUs allocated for the beta launch offer roughly 4.6 exaFLOPS of aggregate FP8 compute — sufficient for a beta pilot but not for 52 million simultaneous users at scale. Operators are expected to supplement with their own private infrastructure from 2027 forward, when the national budget begins covering service-delivery costs.
Three Theories of AI Distribution in the Same Competition
The six-consortium competition is inadvertently running a real-world experiment among rival theories of how governments and companies achieve population-scale AI adoption.
SK Telecom’s consortium — approximately 19 organizations including Liner, Shinhan Card, Hana Card, T Map Mobility, and Goodoc — has built its strategy around its A. (A dot) assistant app, which already counts 10 million monthly active users and has become a leading AI service in Korea’s B2C market. SKT plans to distribute the national AI service through that app, leveraging its 22.5 million mobile subscriber lines as the distribution channel. SKT’s AI business head Kim Ji-hoon stated the company will “deliver services that all South Koreans can easily use, based on our full-stack AI capabilities spanning AI models, infrastructure, and service operations, along with nationwide telecom customer touchpoints.”
Kakao’s consortium — a large grouping including LG AI Research, LG Uplus, LG Electronics, LG CNS, and Seoul National University Hospital — is built around a different existing platform: KakaoTalk, South Korea’s dominant messaging service with more than 50 million users. Kakao will coordinate the consortium and deploy its Kanana AI model, while LG Uplus handles service validation, LG AI Research supplies the K-EXAONE foundation model, LG Electronics contributes device integration, and LG CNS manages public system integration.
KT’s model occupies a third structural position. Where SKT is asking users to engage with A. more deeply, and Kakao is routing AI through KakaoTalk conversations, KT is embedding AI directly into the domain-specific platform where a user is already accomplishing a task — shopping for a dress, solving a math problem, searching for an apartment. The AI serves the task the user is already doing, rather than requiring them to switch contexts to an AI interface. This approach maps most directly to Rogers’ “compatibility” dimension: the user perceives no new product to learn and no behavior to change.
The Ministry of Science and ICT is expected to select two or three of the six consortia, meaning more than one distribution theory may ultimately shape how Korea’s national AI service reaches its residents.
What It Would Take for This to Work
KT’s platform-embedding strategy carries a structural advantage and a structural risk.
The advantage is frictionlessness: the users who have most resisted AI adoption — older Koreans, rural residents, those classified as digitally vulnerable — are disproportionately likely to resist a new app download but far less likely to notice when a platform they already use quietly begins offering AI-assisted recommendations, answers, or applications. If the goal is reaching the 29 million Koreans who have not yet tried AI, removing the requirement to try something new is the most direct path.
The risk is quality. The domestic model mandate requires routing 80% of queries through Korean foundation models — primarily in the 7-billion to 32-billion parameter range — rather than through frontier commercial models operating at orders of magnitude more capability. Within Korean-language and Korean-cultural tasks, the performance gap may be modest. For complex multi-step reasoning or queries that require broad world knowledge, the gap may be visible. Whether that gap matters for the specific tasks Musinsa users, Qanda students, and Daum searchers bring to these platforms is the defining empirical question of the program’s first year.
The beta service is scheduled to launch in late September, giving the first real-world evidence within weeks of this writing.
Does the Platform Model Change How AI Spreads Globally?
Korea’s AI for All competition is worth watching for a reason that has nothing to do with Korea specifically. Most governments attempting to deliver AI at national scale are planning to build new products — apps, portals, government chatbots — and then distribute them through channels they hope citizens will adopt. Korea is the first large democracy to test whether the distribution problem can be short-circuited entirely, by routing AI through the commercial platforms that citizens already use without being asked to change anything.
If KT’s consortium wins and its platform-embedding model proves effective at reaching the population that standalone AI apps could not, it creates a template for how governments elsewhere can approach the AI equity gap: not by building new AI-first products, but by finding the platforms where digitally hesitant citizens already live and adding AI there.
The decision arriving this week will determine which of Korea’s three competing theories gets its first national-scale test. By late September, when the beta service launches, early evidence will begin to emerge. By year-end, when the full service is expected to go live for all 52 million citizens, South Korea will have produced the most detailed data set yet on whether the hardest AI adoption challenge is building the model or solving the last-mile distribution problem.
Frequently Asked QuestionsWhat is Korea’s “AI for All” program, and when will it launch?
“AI for All” (모두의 AI, also called ModuAI) is a South Korean government initiative committing to provide all 52 million residents with free, unlimited access to an AI assistant and public-service agent, built primarily on domestically developed Korean AI models. The Ministry of Science and ICT opened bidding in July 2026, received proposals from six consortia by the August 11 deadline, and is expected to select two or three winning operators before the end of August 2026. A beta chatbot service is planned for late September 2026, with a full service — including an AI agent that proactively identifies government benefits for individual users — targeted for year-end 2026 and into 2027. The service will remain free through 2028 under government funding.
How does KT’s approach differ from the approaches of SK Telecom and Kakao?
The three leading consortia represent three distinct distribution theories. SK Telecom is routing the national AI through its existing A. assistant app, which already has 10 million monthly active users, and plans to distribute further through its 22.5 million mobile subscriber lines. Kakao is distributing through KakaoTalk, South Korea’s dominant messaging platform used by more than 50 million people, with its Kanana AI model and the K-EXAONE model from consortium partner LG AI Research providing the underlying capability. KT is embedding AI directly inside consumer-facing platforms — Musinsa, Qanda, Daum, Zigbang, BC Card — so users encounter AI while doing what they already do, rather than adopting a new AI interface. The platforms approach most directly targets the population that has resisted every AI app so far.
What is the 80% domestic model requirement, and does it affect service quality?
The Ministry of Science and ICT requires that at least 50% of all queries be routed through certified Korean foundation models from the government’s sovereign AI competition, and at least 30% more through other Korean-built models — creating a mandatory domestic floor of 80%. Korean sovereign models currently operate primarily in the 7-billion to 32-billion parameter range. Within Korean-language tasks, the performance gap relative to frontier commercial models may be modest; for complex multi-step reasoning, the gap is more likely to be visible. Whether that gap matters for the specific tasks the “AI for All” service is designed to support — civic information retrieval, government benefit navigation, learning assistance, and consumer search — is the central empirical question that the beta launch in late September will begin to answer.
Why does Korea’s approach to government AI differ from what the US and EU are attempting?
Most governments in Europe and North America are using AI within existing government workflows — administrative automation, fraud detection, benefit processing — rather than delivering AI directly to citizens as a public service. The US government has moved toward AI adoption inside agencies but has no equivalent of a free, unlimited AI utility for citizens. The EU’s AI Act governs how AI is used but does not mandate government AI provision. South Korea is the first G20 nation to commit government funding to providing AI access as a basic public service on par with public broadband or public health infrastructure, with a specific domestic technology mandate requiring that Korean-built models power the majority of the system.