KT says only 7 of Token Factory's 73 AI models run on its own hardware as it plans more GPUs

Tech Times reported on November 11 on a striking ratio inside KT's Token Factory: the platform offers 73 AI models, but only seven of them run on KT's own hardware.

The same report notes that KT plans to add more GPUs. Read together, the two facts describe a gap between the catalogue a platform can assemble and the compute it actually owns.

Token Factory sells model access by the token, and inference capacity is the core cost of that business. A low share of in-house hardware usually means inference leans on external compute, which shows up in unit economics, latency and margins.

For a telecom operator, owning GPU clusters is about more than cost. Local inference touches data residency, regulatory compliance and the service guarantees enterprise customers expect, and it is the most tangible layer of the sovereign AI story.

That also explains the GPU shopping. Aggregating models is a commercial exercise that can be done in a quarter through partnerships; expanding compute runs through procurement, data halls, power and installation.

Three things to watch: which seven models are hosted in-house (likely the most heavily used or most sensitive), the size and source of the new GPU fleet, and whether KT discloses what share of total inference its own hardware carries. Those numbers say more about Token Factory than the length of its model list.