SK and GS Plan Massive AI Data Centers Amid Overcapacity Concerns
South Korea's SK and GS pursue AI data center projects up to 17.4GW, prompting expert warnings over real demand, GPU procurement hurdles, and asset depreciation
Image source: https://www.enewstoday.co.kr/news/articleView.html?idxno=2481819
According to industry reports on October 20, 2026, South Korea's SK and GS groups are advancing massive artificial intelligence data center initiatives under a government mega-project framework. However, mounting concerns over actual computational demand and hardware availability have triggered calls for thorough verification, as rising token usage does not inherently secure high facility utilization or operational profitability for domestic infrastructure.
Infrastructure Roadmaps: SK's 15GW and GS's 2.4GW
SK and GS plan long-term capacities of 15GW and 2.4GW respectively, which reach 18.4GW when combined with Naver's 1GW proposal. Led by SK Telecom, SK plans to open an initial 5GW facility in 2029 and expand to 15GW by 2035, leveraging domestic HBM production, partner equity, and project financing. Meanwhile, GS aims to construct a 1.2GW hub in Donghae by 2028, later scaling to 2.4GW alongside power and cooling ecosystems.
Domestic Demand Ceilings and Overseas Leasing Risks
Experts question whether actual demand can support this physical expansion. Professor Lee Seung-hyun of Hanyang University ERICA noted that domestic public sector demand will likely need only 1GW to 2GW even under aggressive adoption, meaning surplus capacity depends entirely on foreign clients. Seok Kwang-hoon of the Energy Transition Forum warned that if foreign tenants face funding difficulties or delay GPU installations, Korean operators supplying land and power could bear debt service and upkeep costs without receiving expected rental income.
Hardware Procurement and Phased Deployment
Hardware availability presents another critical bottleneck. Professor Lee estimated that a 5GW facility requires roughly 3 million GPUs, underscoring that power and land alone cannot provide compute services. Furthermore, deploying hardware ahead of firm commitments risks rapid asset depreciation; high amortization can create deficits despite rising revenue. Analysts advise cultivating domestic AI software and pacing construction strictly to signed contracts and verified GPU delivery schedules.