In the government's second round of support for graphics processing units (GPUs), first-round winner Naver Cloud and first-round loser Coupang are re-entering the bidding. While many view Naver Cloud and Samsung SDS as leading contenders, attention is on whether Coupang's renewed bid could change the race.

The Ministry of Science and ICT said 5 companies participated in the tender for the 2026 project to strengthen the foundation for using AI computing resources (support for securing, building and operating GPUs).

In the first round, 4 companies participated: Naver Cloud, NHN Cloud, Kakao and Coupang. Three companies were selected, excluding Coupang. A total of 1.46 trillion won was投入 to secure 10,080 B200 units and 3,056 H200 units. Among the first-round beneficiaries, only Naver Cloud said it intended to join the second round, and Coupang, which failed in the first round, is also trying again. KT Cloud, Samsung SDS and Elice Group are joining as new participants, creating a five-way contest.

According to the ministry notice, the project sets 5 key directions: economic feasibility, large scale, direct clustering, latest GPUs and launching services within the year. The evaluation totals 100 points: project readiness and competitiveness (50 points), efforts to develop the AI ecosystem (26 points), understanding of the project and execution capability (12 points), and operating capability and project management (12 points). In the build-plan excellence category (32 points), which carries the highest weighting, at least 1 cluster for government use must be configured with 256 servers (2,048 GPUs) or more, with preference given to larger clustering. Faster service launch schedules also earn additional points. The structure effectively favors large infrastructure operators with data center space and large-scale clustering capabilities.

Among systems integration (SI) companies, Samsung SDS was the only participant. It was recently selected as preferred bidder for a 2.5 trillion won project to build a national AI computing center, and it has a track record of supplying the latest resources such as H100 through a high-performance computing (HPC) support project run by the National IT Industry Promotion Agency (NIPA). The industry believes it likely proposed for this project citing strengths including operational know-how from its Dongtan AI-only data center and capabilities to launch Korea's first B300-based GPU-as-a-service offering.

KT Cloud operates 16 data centers, 8 in the capital area and 8 outside it, and has the largest scale among domestic cloud service providers (CSPs) by physical infrastructure. In November last year it opened the Gasan AI Data Center in Gasan-dong, Geumcheon-gu, Seoul, with a total floor area of about 11,046 pyeong and IT capacity of 26 megawatts. It is also pushing ahead with building new centers, including Bucheon Samjeong (48 megawatts), aiming to secure 500 megawatts of power capacity by 2027.

Naver Cloud entered the second round based on its participation in the first. With the government specifying GPUaaS operating track record as an eligibility requirement, Naver Cloud secured a large-scale GPU operations reference by directly building and operating 3,056 H200 units through the first-round project.

But there are also calls for more companies to get a chance. An official at a company participating in the second-round tender said, "Given the evaluation structure, it is true that existing operators with operating records have an advantage over new entrants, but as winning a government project becomes a good reference for domestic operators, fairness should be considered so that opportunities go to various companies."

Coupang, which failed in the first round, is also in focus. With NHN Cloud and Kakao widely reported to have opted out of the second-round tender due to difficulties such as securing additional data center space, Coupang's position appears to stand out relatively because it has secured space in central Seoul.

Coupang is operating GPU infrastructure by leasing a data center in Yangjae-dong, Seoul, owned by Singapore data center specialist Empyrion Digital. The facility has a total floor area of 60,000 square meters and power receiving capacity of 40 megawatts, and liquid cooling that supports 130 kilowatts per rack was introduced on the seventh and eighth floors. It is assessed as advantageous in securing engineers and in latency due to its central Seoul location. Also drawing attention is Nvidia's announcement last month of building an "AI factory" for Coupang Intelligent Cloud (CIC). Industry observers believe a large DGX SuperPOD purchase contract would have accompanied the announcement, since Nvidia presented Coupang as a customer case. Coupang has not disclosed direct investment in AI infrastructure such as data centers, but it disclosed capital expenditure of $1.246 billion, about 1.65 trillion won, as of 2025.

Elice Group is the only startup among the 5 companies. It operates portable modular data centers (PMDCs) and highlights that it can build within 3 to 4 months by simultaneously carrying out site preparation, module production and IT equipment installation. It has previously been selected as a supplier for a GPU leasing project led by the ministry, and supports the full lineup of B200, H100 and A100. The company said it has also completed a PMDC design that supports the next-generation GPU Vera Rubin NVL72.

The ministry plans to confirm the project operator in May after selecting an evaluation committee, conducting evaluations in April and May, and carrying out on-site inspections of data centers. The government did not disclose in advance the number of companies to be selected in this tender, but explained that, assuming no issues, it could select a minimum of 1 and a maximum of 5.

A ministry official said, "The evaluation criteria are completely separate from the first project and are unrelated to whether a company re-enters," and added, "After the evaluation, we will verify through on-site inspections that there are no differences from the proposals, and then plan to make final selections through adjustments such as negotiations, taking into account that each operator's scale differs."