KOSA Publishes "Public Sector GPU Utilization Strategy Report"

The AI Policy Cooperation Committee (Chairman Lim Woo-hyung, Co-Director of LG AI Research Institute) under the Korea Artificial Intelligence and Software Industry Association (Chairman Cho Jun-hee, KOSA) announced on Thursday, February 19 that it has published the “Public Sector GPU Utilization Strategy Report” containing specific implementation plans for Korea to leap forward as one of the world’s top three AI powerhouses (G3).

This report presents, from the perspective of industry, ways to efficiently utilize the volume of graphics processing units (GPUs) the government will secure by 2030. It specifically points out the short lifespan of GPUs, at 3-5 years, and the lack of actual demand relative to the scale of the infrastructure. It warns that if utilization rates are not maximized from the outset, these massively invested assets risk becoming mere scrap metal. Accordingly, it recommends a bold shift in national policy focus from the existing competition for infrastructure ownership to a competition for industrial utilization.

The AI Policy Cooperation Committee presented four key strategies in its report.

First, the government must become the "first customer" and drive the early market. He emphasized the need to eliminate market uncertainty by mandating the adoption of domestically produced AI in public sectors such as administration and national defense. He also emphasized the need to establish a "one-stop package" that supports the entire process, from diagnosis to implementation, for small and medium-sized manufacturing companies with low AI adoption rates.

Second, the budget structure, which is currently heavily focused on hardware purchases, must be reformed to recognize the value of software (SW) and data. To enable companies to respond to business needs in a timely manner, a "Rolling Review" track for government-funded projects should be established. Furthermore, a multi-year support system should be established to guarantee a seamless research environment for up to three years (2+1 years) for high-performing companies.

Third, to improve infrastructure efficiency, the learning and inference stages should be strategically separated. While NVIDIA GPUs should be heavily invested in high-difficulty model development (training), domestically produced NPUs should be used for public service (inference) to ensure early references for domestically produced chips.

Fourth, we must cultivate practical AI engineering talent and "supercomputing architects." He emphasized the need for upskilling industry veterans with domain knowledge to acquire AI capabilities, as well as systematically developing architects capable of optimally designing and operating large-scale GPU clusters at the national level.

Lim Woo-hyung, Chairman of the AI Policy Cooperation Committee and Co-Director of LG AI Research Institute, said, “Now that securing GPUs has become visible, it is the golden time to realize the AI G3 leap forward,” and “The ‘public-private one team’ play, in which the public sector takes the lead and the private sector responds with creative engineering, is more urgent than ever.”

KOSA Chairman Cho Jun-hee said, “We must secure global export competitiveness by building a ‘full-stack AI’ package armed with our country’s world-class ‘manufacturing data.’” He added, “The association will take the lead in creating a healthy ecosystem where the government budget does not stop at hardware such as GPU purchases, but is properly recognized for the value of software, such as through payment of AI process costs.”

The full report can be found on the KOSA website .

• See more related articles