The capacity number carries this program, so it is worth taking apart. One B200 producing about 10,000 output tokens a second, delivered in 50-token-per-second slices, serves 200 people at once, which is where the 51,000 simultaneous-user figure for a 256-GPU operator comes from [8][19]. Run those same 256 processors for a full day and you get roughly 221 billion output tokens; divide by the assumed peak factor of six, then by the 2,000 tokens each user is assumed to consume daily, and 18.4 million daily users falls out [9][18]. Every term in that chain is a planning input, and the estimate's own caveat says as much: it holds only for lightweight optimized models, short requests and no early surge [10].
The hardware has since been divided in a way the model did not anticipate. The July 17 scenario sized a single operator at 256 processors, while the state is supplying 512 across three consortia, an average of about 171 each, and the assumptions predated the selection of three operators [8][2][11][17]. On the benchmark arithmetic, the whole national fleet supports on the order of 102,000 users at the same instant [20]. That ceiling is adequate if demand spreads across the day the way the model assumes, and it is the entire exposure if it does not.
The domestic-usage floor is the part of the design that sets the cost curve. Qualifying Korean foundation models must meet government independence standards, and foreign models may cover limited functions but draw no government funding for doing so [4][5]. The standard lever for holding an inference bill flat is routing each workload to the cheapest model that clears the quality bar; here the funded majority of traffic is committed in advance to a supplier list chosen on industrial-policy grounds. Seoul has taken the domestic-capability side of that trade, and the price of it will show up as forgone efficiency rather than as a line item.
A skeptic would say unlimited is a soft word, since more than 20 million South Koreans already use free generative AI without any state involvement [7], and none of them use it continuously. That is right, and it is why a no-limits base tier is deliverable at all. But with a fixed fleet and no cap on the basic chatbot or the public-service agent [12], rationing still has to land somewhere, and what is left is model size and response speed. Fifty tokens a second implies about ten seconds to finish a 500-token answer [21], and nobody has committed to that as a service level, least of all for the phone and text channels meant for people who struggle with digital services [14].
The financing sequence is the durable part. State hardware first, partial operating support a year later, operator-generated revenue after that [2][3], with paid specialized and higher-performance tiers permitted above the free floor [12]. That makes a publicly funded free tier the acquisition channel for three companies' paid products, which can be a defensible bargain. It cannot be audited as one, because the ministry has published no cost and set no adoption target [6], and a cost per user needs a denominator. Agreements are due in September 2026 with a launch by year-end [13], which is the window in which that number is still a negotiating question rather than an appropriation.
