KT (030200.KS) has developed a next-generation AI technology that improves performance through data calibration rather than model retraining, offering a way around the technical limits of running costly computing hardware to train AI on new data. The advance could provide a technical foundation for the "token factory" plan that KT's chief has put forward as a new business.
KT presented the research at CVPR 2026, an international AI conference held in the United States in June, telecommunications industry sources said on the 11th. Developed jointly with Chung-Ang University, the Electronics and Telecommunications Research Institute (ETRI) and others, the technology is called HyCal.
HyCal's core strength is that AI does not need a separate training process when handling new data. Rather than fine-tuning — training on additional data in a specific field to adapt a model to a purpose — or expanding parameters, KT's data calibration technology alone raises AI performance. In effect, it is an efficiency technology that reduces the burden of training and operating AI.
Analysts say it is particularly effective for vision-language models (VLMs), which must take in various types of data. Medical imaging and satellite images differ in data volume and visual patterns, making it difficult for a single AI to process them together. Such gaps widen whenever the client, industry or use context changes. But applying KT's technology, which calibrates data values, processes the data within the embedding space — the numerical representations an AI uses to understand data — already learned inside the model, eliminating the need for retraining.
"We can reduce computing costs, such as the GPUs and memory needed to retrain an entire model each time a new type of data emerges," KT said. "We expect the cost savings to be larger in environments with frequent, repeated model updates."
KT is reviewing whether to apply the research to the next-generation version of its own AI model, Mid:m K. Because Mid:m K is being expanded as a multimodal technology that understands different forms of data together — text, images and audio — HyCal is expected to be highly useful.
The telecommunications industry is watching whether HyCal will become the technology that turns KT's token business plan into reality. A token factory is a type of AI operating platform that automatically selects the most efficient model among multiple AI systems, measures usage in units of tokens, and handles optimization, billing and settlement. As performing the same task with less computation emerges as a core competitive edge in AI, KT's AI efficiency technology could be offered as a new alternative, analysts say.
According to Goldman Sachs, the spread of AI agents is expected to drive global monthly token usage up 24-fold, from 5 quadrillion this year to 120 quadrillion by 2030. "Many companies are busy looking for efficient ways to use tokens to lower their AI costs," a telecommunications industry official said. "AI operating efficiency technology will grow more important as it ties into the token economy."
