Seoul's 80 percent quota turns Korean model builders into each other's wholesale customers

Divide the state's chip grant three ways and each consortium gets about 171 B200s ^[1] to serve a population of 51 million ^[4] with no token limits ^[5], which works out to one state-supplied GPU per 99,609 citizens ^[2]. The Carnegie Endowment reported in June that South Korea has no major publicly known operational AI chip clusters, even as SK Hynix and Samsung supply the high-bandwidth memory inside much of the world's AI hardware ^[19], so the grant is a down payment on inference rather than the inference bill, and SK Telecom's plan to build a gigawatt-class AI cloud on Nvidia's DSX system from 2027 ^[20] indicates where the balance is expected to sit. The ministry has still not put a figure on the programme ^[7].

Measured against peers, the spend is modest and pointed in a different direction. President Lee Jae Myung's administration has earmarked about $7.2 billion (10 trillion won) for AI in 2026, three times the previous year's commitment ^[15], which implies roughly $2.4 billion in 2025 ^[7], amounts to about 3.1 percent of the $232 billion the European Commission set out to mobilise through InvestAI from February 2025 ^[16][5], and runs about 5.8 times the $1.25 billion India approved for its IndiaAI Mission in March 2024 ^[17][6]. Brussels and Delhi are funding machinery. Seoul is funding the finished product, already in citizens' hands ^[18]. The allocation is legible in what it forgoes: money that could have bought clusters in a country that has none ^[19] is buying free serving instead.

The demand side is where I would stress-test this. About two-thirds of South Koreans have tried AI and roughly 23 million use generative tools regularly, while only about 1.8 million people in the country pay for anything ^[11][13], a conversion rate near 7.8 percent ^[4], and the ministry has set no uptake target ^[14]. SK Telecom's route to habit is delivery rather than novelty: A.X K2 and other Korean systems reachable by app, phone call and text message for older users ^[9], wired into Government24 and the PASS identity app for 83 types of official certificate ^[10]. Certificates are a chore people already perform, which is a better wedge than a chat box.

This is probably wrong, but the 80 percent floor reads less as protectionism than as procurement, a state guaranteeing serving volume to domestic labs whose audience mostly does not pay ^[13]. Or rather, the more interesting version: the floor prices a model's distribution before anyone has agreed what the model is worth.

It can bend elsewhere. The foreign carve-out is defined only as what the ministry considers minimally necessary ^[3], a phrase that can widen without an announcement. Uptake may stall against free foreign tiers, in which case the chip grant gets renewed rather than retired and the 2027 handover slips. Or the cross-purchase rule works, and two or three Korean labs end up billing volume they can eventually price. I would take the thesis as wrong if operators show real paid conversion off a free unlimited tier by the end of 2027, or if the public-service budget line shrinks instead of growing.