Free AI for a nation · Part 4 of 7
Who pays for what, and what the deal rewards
The state leases the GPUs and grants each consortium 10 billion won; the operators carry the service and must find revenue against rivals that are already free. What that rewards is a hypothesis about incentives, not a measured outcome.
Three ledgers are easy to run together: the state’s, the operators’ and society’s. This chapter keeps the first two apart; chapter 6 is about the third.
What the state gives. This year, 512 government-held B200s in kind, as a managed cluster run by a supplier the ministry hasn’t named, valued by the call at 6.6 million won per GPU a month, hosting included.[86] For 2027, under the proposal, a one-year lease of about 2,000 more and 10 billion won ($7.43 million) per consortium for “system operation, securing stability and collaboration with individual participating companies”.[81] From March 2027, an official told Financial Post Korea, GPUs may move to whichever consortium shows better usage.[11] The GPUs are a cost to the state and an opportunity cost of capacity; they are not cash the operators pay.
What the operators carry. The free chatbot and public agent with no usage limit at least until 2028, the 50% and 30% model shares (chapter 5), a separate app, at least 10% of the project cost themselves, and everything the GPUs don’t cover: storage, servers, network, staff, model updates, safety work and support.[28, 86, 89] In 2026 there is no money for any of it.
What running it costs, nobody publishes, and the answer turns on whether the 2027 lease includes hosting, which isn’t published either. If it does, as this year’s cluster does, space and power are inside the lease and the grant pays for the layer around the GPUs, which no reference price covers. If it doesn’t, reference prices give a scale: space and power for one B200 in Seoul come to about 12% of the lease price, and SemiAnalysis’s 2023 figures for an H100 server put hosting at 21% of capital plus hosting, or 27% of capital alone.[90, 92] For a third of the lease that is 6.9 to 12 billion won a year, about the size of the grant, in that scenario only. Neither scenario prices engineering, safety and support, and adding hosting to a lease that includes it would count it twice (appendix).
Revenue, in two columns. Permitted, by the call and the FAQ: the prompts users type, which operators may use as data; paid advanced features; and any revenue model, ads, partnerships, add-ons or business sales, as long as it doesn’t undermine the free service.[22, 24, 27] Observed: nothing yet, because the service hasn’t launched, and the citizen’s alternatives are already free. For scale, Kakao earns 32,230 won a year per KakaoTalk user from Talk Biz ads and subscriptions, a whole messenger’s revenue, not margin, and not what an AI service inside it can expect; the state’s subsidy at the end-2026 targets is 16,667 won per user.[85] At the state’s lease price and chapter 2’s full-load scenario, a chat user’s output costs about 606 won of lease a year and one coding-session task a day about 58,000 won: the state’s lease allocated to tokens, not an operator’s cash cost or a measured cost of the agents the consortia promise.
So what does the deal reward? A hypothesis, with its conditions. If GPUs follow monthly users and repeat visits, and if revenue follows prompts or ad impressions, then the operators are paid for engagement rather than finished tasks, and the citizen whose agent books the trip and is done uses the most capacity and brings the least of either. Both ifs are open: the reallocation formula isn’t published, and nobody has shown that agents yield less valuable data than chat; my coding session left a long trace, and a booking agent generates intent and transactions that a chat never does. The risk is real enough to design against, with chapter 6’s measures. It is not an outcome anyone has measured.