[Digital Today reporter Chi-gyu Hwang] The AI Policy Cooperation Committee under the Korea Artificial Intelligence and Software Industry Association (KOSA) said on Feb. 19 it published a "Public Sector GPU Utilisation Strategy Report" laying out specific steps for South Korea to leap into the global top three AI powers (G3).

The report presents, from an industry perspective, ways to efficiently use the volume of graphics processing units (GPUs) the government will secure by 2030.

The report pointed out that GPUs have a short lifespan of 3 to 5 years and that effective demand for use is insufficient compared with the scale of infrastructure. It warned that if utilisation cannot be maximised from the early stage of adoption, assets purchased with massive budgets could risk turning into scrap. It recommended boldly shifting the weight of national policy away from the existing race to possess infrastructure and toward a race to use it on industrial worksites.

The report's key messages can be summarised in 4 points.

First, the government should become the "First Customer" and drive the early market. The report stressed the need to remove market uncertainty by mandating the adoption of domestically made AI in public sectors such as administration and defence. It also highlighted the need to create a new "one-stop package" to support the entire process, from diagnosis to deployment, for small and medium-sized manufacturing companies with low AI adoption rates.

Second, the budget structure focused on hardware purchases should be overhauled to recognise the value of software and data. The report proposed creating a "Rolling Review" track in government support programmes so companies can respond in time to business needs. It also suggested establishing a multi-year support system that guarantees an uninterrupted research environment for up to 3 years (2+1 years) for companies with strong performance.

Third, the training and inference stages should be strategically separated to raise infrastructure efficiency. The report recommended concentrating Nvidia GPUs on high-difficulty model development (training), while making the use of domestically produced NPUs the principle at the public-facing service (inference) stage to support early reference cases for domestic chips.

Fourth, practical AI engineering talent and "supercomputing architects" should be fostered. The report emphasised up-skilling that equips industry veterans with domain knowledge with AI capabilities. It also said the state should systematically train architects who can optimally design and operate large-scale GPU clusters.

Lim Woo-hyung (임우형), chairman of the AI Policy Cooperation Committee and co-head of the LG AI Research, said, "Now that securing GPUs is coming into view, this is a golden time to realise a leap to AI G3." He said, "A 'public-private one team' play is more desperately needed than ever, where the public sector proactively serves as a priming pump and the private sector responds with creative engineering."