South Korea finalizes free nationwide AI access with an 80% domestic model quota

South Korea signs off on free national AI service

South Korea's Ministry of Science and ICT has settled on the three consortiums that will run its "AI for All" program: SK Telecom, KT and Kakao. According to a detailed write-up on dev.to, a beta opens in September with a full rollout targeted by the end of the year, extending free generative AI access to the country's roughly 52 million residents regardless of age or income, with officials saying usage will carry no token cap.

The same report says the government has allocated 512 NVIDIA B200 chips to the three operators this year as seed compute, and that from 2027 the national budget will begin absorbing part of the service's running costs — a combination the author describes as a first among G20 nations.

The write-up itself flags a caveat: "unlimited" is the official positioning rather than a physical guarantee, since serving an entire population off 512 chips almost certainly implies tiered rate limiting or shifting most requests onto smaller, cheaper models.

The 80 percent domestic floor

The detail sitting under the free-access headline is a quota. Each operator must route at least half of all queries to its own certified Korean sovereign foundation model, and at least a further 30% to models from other Korean companies — an 80% floor for domestic AI in total.

The dev.to author reads this less as social welfare than as industrial policy: the state supplies compute, funding and a distribution channel, and in exchange guarantees domestic models tens of millions of everyday users they would struggle to win from ChatGPT on their own. The quota also lines up with the economics — routine requests can be offloaded to inexpensive domestic small language models, reserving premium compute for genuinely hard queries, which is how a free tier becomes plausible at national scale.

A parallel race to pick the models

Separately from the operators, the government is running a contest to certify its sovereign foundation models, and the two efforts interlock: the 80% quota feeds guaranteed traffic to whichever models survive. Per the write-up, the competition has narrowed from five teams to four, with only two expected to remain by 2027. Visible contenders include LG's EXAONE, Naver's HyperCLOVA X, Upstage's Solar Pro, SK Telecom's A.X series and Kakao's Kanana, with LG AI Research, SK Telecom and Upstage reportedly ahead going into the third round. SK Telecom has also unveiled A.X K1, a roughly 519-billion-parameter model described as Korea's first at hyperscale class.

One notable casualty is Naver, cut in January. The reason given was not benchmark performance but "data independence": HyperCLOVA X had used frozen encoder weights from Alibaba's Qwen, which failed the sort of lineage scrutiny that sovereign AI efforts invite.

A demand-side move in a supply-side race

Per the dev.to analysis, sovereign AI spending worldwide is projected to pass $100 billion. France is deploying 50,000 NVIDIA H200 chips in the first wave of its Projet Voltaire cloud; India picked 12 domestic foundation models from over 500 proposals, backed by 38,000 GPUs; Canada has committed C$925 million over five years; and the EU has mobilized €20 billion for AI gigafactories. Nearly all of these efforts build supply — chips, data centers, homegrown training runs. Korea's move is unusual in attacking demand instead: however good a domestic model becomes, it cannot outbuild the user habits ChatGPT has already locked in, so the state is manufacturing usage at population scale.

Why it matters

If the reported details hold — and they come from a single, though carefully sourced, community write-up — this is the first national AI policy to convert an entire population into guaranteed demand for domestic models, pairing a supply-side model competition with compelled traffic on the demand side. For policymakers, it is a template other mid-sized AI nations could copy. For builders, it is a live demonstration of routing architecture: cheap small models absorb the routine bulk while premium APIs handle the hard remainder. And for users, it is worth remembering that free and unlimited from 512 chips is a policy claim rather than physics; where the practical limits surface should become clear once the beta opens.