KAIST said on Jan. 30 that the AnyBridge AI team, led by computer science professor Jong-se Park (박종세), developed next-generation AI infrastructure software that can provide LLM services efficiently by integrating various AI accelerators rather than relying only on GPUs.

The technology won the top prize in Kakao’s “4 Science and Technology Institutes x Kakao AI Fostering Project.” The project is an industry-academia cooperation programme jointly 추진 by Kakao and four science and technology institutes, including KAIST, GIST, DGIST and UNIST.

The organisers selected top teams after a comprehensive evaluation of the technology and business feasibility of prospective startup teams based on AI technology. AnyBridge AI won the top prize and will receive 20 million won in prize money and up to 35 million won in Kakao Cloud credits.

AnyBridge AI focused on a structural limitation in which most LLM services rely on expensive GPU infrastructure, causing operating costs and power consumption to rise sharply as service scale expands. The researchers analysed that the cause was not specific hardware performance, but the absence of a system software layer that can efficiently connect and operate diverse AI accelerators.

AnyBridge AI proposed an integrated software stack that can serve LLMs in the same interface and runtime environment regardless of accelerator type. It pointed to the existing GPU-centric LLM serving structure and presented “multi-accelerator LLM serving runtime software” as a core technology, receiving a high evaluation for enabling multiple types of AI accelerators to be used together in a single system.

KAIST explained that the technology can enable a flexible AI infrastructure architecture that is not dependent on a specific vendor or hardware, while allowing selection and combination of the most suitable AI accelerators depending on workload characteristics.

Park said, "This award recognises the need for system software that integrates diverse AI accelerators, moving beyond the limitations of GPU-centric AI infrastructure." He added, "We will develop it into a core technology for next-generation LLM serving infrastructure through cooperation with industry partners."