My view: what the money should buy, and what it could crowd out

This chapter is my view. The facts in it are cited; the judgments are mine.

Access is not adoption

The operators were chosen on “service convenience and user-acquisition strategy”, quality and safety, and contribution to the ecosystem. Nothing on outcomes.^[18] At the kick-off the minister gave his own test: “What matters most is how easily citizens reach the service, how often they come back, and how much it actually helps their lives.”^[12] Two of the three are reach and frequency, which are easy to count. No one has said how the third will be measured, and users are the cheapest thing to buy: put a free chatbot in front of people who already use KakaoTalk or a phone, and monthly users follow. Whether they get anything done is the question the budget should answer. Completed tasks can be gamed too, by splitting tasks, counting trivial successes or recruiting the users who would have succeeded anyway, so the count needs a definition by task class and a quality check behind it.

Free also doesn’t close the gap the ministry worries about. In the 2025 Internet Usage Survey, 55.5% of Koreans had never used generative AI.^[77] In KISDI’s panel, 59.2% of non-users called the cost a burden and 63.5% said they don’t know how to use it or find it hard to learn.^[74] A free app removes the first barrier, not the second. My judgment is that the benefit will follow the existing gradient unless someone counts who is using it, not only how many: new users from the 55.5%, by age and region.

What I would measure

The FinOps Foundation separates resource-efficiency metrics, such as cost per token, from business metrics, such as cost per case resolved.^[80] The programme has published neither, and the second is the one that matters. On my scorecard the cost side carries the budget in GPU-years, about 2,512 in 2027 under the proposal, and the full cost per successful outcome: GPUs, operations, engineering, safety and support together, not the lease alone. The outcome side carries tasks completed by class, with errors, reversals and rework counted against them; net time saved by citizens and staff after they have checked the answer; and additionality, the benefit over the public-service channels and free chatbots that already exist. A lower raw cost per completed task proves nothing until the task mix and quality are held constant.

Three more lines: GPU status, measured separately as delivered, commissioned, allocated, available, busy, and doing useful work within a service level (of the 35,200 B200-equivalent GPUs the state has secured elsewhere, one outlet reports 11,600 in service, a deployment status, not a utilization^[4]); queue time and speed at the busiest hour; and guardrails on prompt data. Moving GPUs by usage rewards whatever is counted. If the count is monthly users and repeat visits, the cheapest way to win GPUs is engagement; count quality-checked tasks and peak-hour service, and the GPUs follow usefulness instead. The reallocation formula hasn’t been published, so that choice is still open. The full scorecard is in the appendix.

The subsidy goes through the platforms that already have the users

The state didn’t pick three AI companies; it picked three distribution networks. KakaoTalk has 49.6 million monthly users in Korea, 96% of the population, and Kakao plans to make it the front door for the chatbot and its agents; KT has the Daum portal, SK Telecom the phone.^{[18, 71]}

Subsidising established platforms could speed adoption, and that case is a good one: the people who have never used AI already use KakaoTalk and a phone. It could also squeeze smaller unsubsidised rivals. Wrtn, one of the best-known Korean AI startups, had 380,000 app users in September 2026 and now competes with a free, unlimited chatbot on state GPUs in the apps everyone already has.^[69] Whether that takes Wrtn’s users is a risk to watch, not a result: nothing has launched, and nothing has been measured. There is evidence on the other side. Some smaller firms are inside the consortia: Liner in SK Telecom’s, Upstage and Motif in KT’s.^[21] The budget’s third line, 50 billion won ($37.1 million), is reported as money to develop finance, medical and public-service agents and connect them to the three platforms, which is funding for specialists to build on the subsidised capacity.^[11] And the rule that 30% of the output tokens come from other Korean companies’ models is a lasting duty to share. None of this guarantees fair terms, but a crowd-out verdict has to weigh it.

What I would ask in return

In return for the GPUs, I would ask the winners to open up: APIs and integration points that smaller companies can build on, including those that lost or never bid, so that they build on the subsidised capacity instead of competing against it. An open API is a proposal, not a cure. It would need terms: who gets access, at what price and quotas; who funds the capacity behind third-party traffic; nondiscrimination against the platform’s own agents; portability; security, liability and service levels. Without them, openness could deepen dependence on the same three distributors, and the common app the ministry is discussing guarantees no neutral defaults either.^[81] The tender asks for less: it scored the bidders’ ecosystem plans, 20 of the presentation round’s 100 points, and nothing binds after selection; the FAQ sets no duty to open APIs, and the winners’ marketplaces are still plans.^{[86, 89]}

The payoff the state should want is productivity, not engagement: Korea is projected to have 32% fewer people of working age in 2050 than in 2025, and an ecosystem of many services on shared capacity is how a free model turns into value for those who remain.^{[87, 91]} The number to watch is in Kakao’s own app: Kakao is keeping ChatGPT inside KakaoTalk after AI for All launches, so the subsidised Korean service and the foreign one it is meant to replace will sit in the same place, and users will choose.^[8] That is a fairer test than most subsidies get.