South Korea picks three operators to give every resident free AI

South Korea has chosen the three groups that will give every resident a free AI assistant with no usage caps. TechSpot reports the winners were selected on August 29, 2026, and they are led by SK Telecom, KT and Kakao. The condition attached to the money is the part worth studying: most of the usage has to run on models built in Korea.

The program is run by the Ministry of Science and ICT. Sources use two names for it. The Next Web calls it "AI for Everyone," while TechSpot, gHacks and Cryptopolitan call it "AI for All."

How the operators were picked

The tender was public and short. According to The Next Web, bidding opened on July 13, 2026 with an August 11 deadline. Cryptopolitan reports that six groups applied and three were selected.

The winners are the country's two largest telecom companies plus the operator of Kakao, the messaging app most Koreans use daily. The state is not paying for everything. The Next Web reports that the selected companies must match the government's GPU support with their own funding.

Hardware support is where the reports disagree. The Next Web says each consortium receives up to 512 Nvidia B200 GPUs. TechSpot and gHacks describe up to 512 B200 chips distributed across the three operators in total. That is a threefold difference in the compute behind the service, and it is unresolved.

The Korean-model quota is the real policy

The headline is free AI. The policy is industrial. Cryptopolitan reports that at least 80% of the service has to run on Korean models, and it breaks the requirement into two parts.

• At least 50% must run on the operator's own sovereign foundation models. A foundation model is the large base model an assistant is built on.
• At least another 30% must come from other domestic companies.

Foreign systems are not banned outright. Cryptopolitan reports they are permitted only where "minimally necessary," and without government subsidy. The Next Web describes the same structure as a 50% domestic floor with an additional 30% requirement.

Read next to the country's current habits, the quota is a deliberate correction. The Next Web cites usage of 23.45 million for ChatGPT, 8.45 million for Gemini and 2.41 million for Claude. The state is buying distribution for Korean models rather than waiting for them to win it.

What residents get, and when

Access is unlimited. TechSpot quotes the terms plainly: "Users will have unlimited access, with no token limits under the government-supported plans."

The service is not app-only. Cryptopolitan reports delivery through standalone apps, phone calls and text messages, which reaches people who never install anything. It also says SK Telecom is wiring the assistant into the Government24 portal and the PASS identity app. That app covers 83 official certificate types. The same report counts 20 direct partner organizations and 11 more agreed by memorandum.

Item	Detail
Public beta	September 2026 (Cryptopolitan says late September, with SK Telecom targeting October)
Full launch	Before the end of 2026
Program horizon	Through 2030, per The Next Web
Who pays after 2026	Operators must build their own revenue from 2027, per Cryptopolitan
Population covered	52 million per The Next Web; 51 million per Cryptopolitan

Named use cases are ordinary government errands. The reports list booking medical appointments, searching for housing, getting tax advice, checking eligibility for government support, and finding study material.

One budget figure needs care. TechSpot and gHacks both cite about 10 trillion won, or $7.2 billion, for South Korea's AI-related spending in 2026. That is the national figure, not this program's. The gHacks report states clearly that the ministry has not disclosed what this program costs.

What this means for developers

A government just wrote model provenance into a production requirement, and that is an engineering problem before it is a political one. Meeting an 80% quota split 50-30 means routing every request by who built the model, then proving the split afterwards. If you build serving infrastructure, expect requests for per-request model attribution, logging and reporting. Compliance-aware routing is becoming a feature, not a config file.

If you train or fine-tune models, notice what just became valuable. National procurement is now a distribution channel, and it rewards a model for being domestic as much as for being good. Any team building sovereign models in any country should read this tender as a template.

For anyone selling a consumer AI app in Korea, the floor price has moved to zero. Cryptopolitan reports operators must fund themselves from 2027, so watch what appears then: ads, paid tiers, or a renewed subsidy. That is the date when the real business model shows up.

Finally, do the capacity arithmetic yourself before believing "unlimited." The chip allocation is reported as 512 B200s, either per operator or in total, against a population above 50 million. Ask what unlimited means in practice when the queue is long, and treat generous free tiers elsewhere with the same question.