KT underwrites a free AI consortium with GPUs it bought beyond its government allocation

Seven applicants on top of 16 members makes 23, about 44 percent more than today ^[19]. Because 16 plus five already clears the target KT gave, it can decline two of the seven and still get past 20 ^[20]. "We plan to review additional partnerships with these companies and will continue to accept applications from others," Lee said ^[5].

What the members get is placement. Ieum Inside links a general-purpose agent built on real-time search with services specialized for more than 10 fields, including public administration, education, finance and shopping ^[6]. A KT official said users "will be able to enjoy the same AI experience naturally within the services they already use, without having to install a separate app" ^[8]. In the Zigbang case KT described, the agent volunteers that a listing "is 50 million won below yesterday's recorded sale price and was listed only two days ago" ^[17].

Those placements generate tokens on KT's own platform. Token Factory runs domestically developed models including Solar, Motif and Mi:dm alongside Rebellions' neural processing units ^[9], and KT says separating prompt processing from response generation improves tokens per minute by about 31 percent ^[10], while prompt compression can cut input tokens by up to 80 percent ^[11]. Set that against the reserve. Lee said, "We technically hold a buffer of about 20% in case traffic surges, and we also have GPUs on hand for when demand exceeds that" ^[13]. A 20 percent buffer above serving load is roughly one GPU in six sitting idle at normal traffic ^[21], and the 31 percent throughput gain covers that reserve with about 9 percent left over ^[22].

The reserve is KT's own. "To be genuinely committed to AI for All, KT has secured additional GPUs beyond those provided by the government," Lee said ^[12]. KT did not say how many it bought or what a paid tier would cost ^[23].

The consumer route to revenue turns on a decision someone else makes. "If at some point the government opens a monetization model to the general public as well, we plan to separate out premium public services and provide them separately," Lee said ^[14]. KT said it will consider charging for business-use or premium public services once they launch ^[16]. The item with a date on it is a PC version of an AI agent for professionals in January, and the paid model there arrives "if needed" ^[15].

I would expect the business tier to arrive first and the consumer premium tier to arrive late or never, because financial groups and reservation platforms are not queuing for a free assistant, they are queuing for a customer touchpoint they can eventually pay to keep ^[4]. The counter-argument is that KT never prices a subscription at all, since the return arrives as usage inside the Daum search portal, member service touchpoints and Genie TV ^[7], all places KT already bills someone. If the January professional product ships free with no upgrade path, that second reading wins and the GPU purchase is a distribution cost.