The Ministry of Science and ICT (MSIT) has selected Naver Cloud, Samsung SDS, and Elice Group as the participating companies for a government project aimed at securing, establishing, and supporting the operation of advanced Graphics Processing Units (GPUs). This initiative is a core component of building an "AI highway" to invigorate South Korea's artificial intelligence (AI) ecosystem and elevate the nation to one of the top three AI powerhouses globally.

Through this collaboration with the three chosen Cloud Service Providers (CSPs), a total of 9,704 advanced GPUs will be acquired and deployed. These resources are intended to support AI innovation across both the private and public sectors. The project aims to empower enterprises and research institutions to bolster their AI capabilities through AI model and service development and AI R&D, while simultaneously enabling the participating CSPs to accumulate experience in building and operating advanced GPU infrastructure, thereby enhancing their AI infrastructure service competitiveness.

The selection process, conducted from March 12 to April 13, involved an open call to which five cloud companies responded. Evaluation criteria included business understanding and execution capability, preparation and competitiveness, efforts towards AI ecosystem development, and operational capability and business management. Following document suitability review and on-site inspections covering physical infrastructure (rack space, electrical facilities, cooling systems) and network environments (internet lines, network security), three companies were ultimately chosen after a project cost deliberation and adjustment phase.

The 9,704 GPUs to be secured comprise 2,016 units of Vera Rubin and 7,688 units of B300 models. Specifically, Naver Cloud will secure 1,008 Vera Rubin and 3,112 B300 GPUs; Samsung SDS will acquire 1,008 Vera Rubin and 2,016 B300 GPUs; and Elice Group will obtain 2,560 B300 GPUs. The MSIT deliberately selected these latest high-performance models, Vera Rubin and B300, due to their superior price-to-performance ratio, despite recent increases in GPU deployment costs driven by rising memory and storage prices. This strategic choice is expected to yield computing performance exceeding initial targets. The secured GPUs are estimated to provide performance equivalent to approximately 19,000 B200 units, representing about a 30% increase over the original goal of 15,000 B200 equivalents. The next-generation Vera Rubin GPU, in particular, is anticipated to significantly enhance bandwidth and computational speed compared to existing models, leading to reduced data bottlenecks, increased parallel processing of user requests, and a substantial decrease in the time and cost required for AI model training.

The utilization plan for the 9,704 GPUs is bifurcated:

• Public/Government Use (6,376 units): 2,016 Vera Rubin and 4,360 B300 GPUs will be dedicated to supporting independent AI foundation model development, national AI projects, and AI model/service development and advancement for industry, academia, and research institutions. This allocation aims to stably meet large-scale AI innovation demands, such as hyper-scale AI model training, from both the private and public sectors.
• CSP Self-Utilization (3,328 B300 units): These GPUs, accounted for as operational costs, will be utilized by the CSPs to provide computing resources for the AI industry and research through cloud-based GPU services. They will also be used by CSPs to advance their own AI models and services. This dual-purpose utilization is expected to establish a virtuous cycle, enabling CSPs to expand their cloud-based GPU services, strengthen their large-scale AI infrastructure building and service capabilities, enhance their internal AI competitiveness, and ultimately contribute to bolstering the nation's overall AI development and utilization capabilities. CSPs are also exploring additional ways to actively contribute to the AI ecosystem, including providing AI services to the private sector and sharing their R&D cases in self-utilized AI services.

The government plans to initiate GPU procurement orders with the selected CSPs in June, with B300 services progressively launching within the year as installation is completed. Vera Rubin services are projected to commence sequentially in the first half of 2027, considering its market release schedule. Lee Jong-ho, the Minister of Science and ICT, emphasized that these advanced GPUs, including Vera Rubin, will be a critical driving force in elevating AI research and development speed and technological capabilities, supporting AI innovation and growth for domestic companies and research institutions by securing world-class AI infrastructure capabilities.