AI Data Center Map of Korea

The map marks four large projects under construction or in development (Chapters 02–05), four large operating assets (Chapter 00), and two reference cases for power approval and financing (Chapters 01 and 06), each placed on its host city or county. Power figures follow each company's or agency's official announcement; because the bases differ (utility supply capacity, project scale, and so on), they are not summed. Kakao Ansan and Yongin Deokseong-ri, whose power figures are not discussed in the text, are shown by location only.

We compared the four sites in Paju, Ulsan, Haenam, and Gumi on secured power, tenant demand, capital sourcing, and operating targets. This is neither a census of all data centers in Korea nor a representative sample. Because business models and the items included in investment figures differ, investment size alone cannot rank the projects economically. Power status is summarized from each project owner's official announcements, and projects at different administrative stages are not placed on the same footing.

Sources: ^{[9][10][11][12][7][13][8][14][15][3]}. Capacities follow company and government announcements and are distinct from IT load available to servers. The Ulsan schedule is as announced in June 2025. Because the items included in each project's investment differ, per-MW costs cannot be compared directly.

New power supply for data centers in the Seoul metropolitan area is severely constrained. Of the 522 preliminary technical review applications filed between August 2024 and March 2026, 279 (53.4%) were ruled not feasible for supply. Of the 24 that reached the main review, only 10 passed (including one in Seoul). Outside the metropolitan area, by contrast, 187 of 214 applications were notified that supply was feasible^[1][2]. The 10 final approvals in the metropolitan area amount to 1.9% of the 522 cumulative applications. Because applications still under review are included, this ratio should not be read as the approval probability of any individual project.

As of the end of March 2026, 9,583MW of the 63,846MW in cumulative nationwide applications remained in the queue, with supply feasibility not yet decided^[22]. Meanwhile, the GS AI data center in Donghae, Gangwon, passed the power supply review for its Phase 1 (100MW) in September 2026 and is preparing to break ground in November^[23].

The power system impact assessment is a procedure that examines in advance how large-scale power demand would affect the national grid^[4]. After its legal basis, the Special Act on Distributed Energy Promotion, took effect on June 14, 2024^[24] the power authorities have compiled the relevant statistics since August of that year^[1]. The detailed operating rules were issued for public notice in July 2026, but as of September 15 the final notice had not been confirmed as enacted^[4].

Separately, the Special Act on the Promotion of the AI Data Center Industry (AIDC Act) takes effect on March 10, 2027. The Act includes an exemption from the power system impact assessment for new AIDC construction outside the metropolitan area, for the expansion of operating facilities, and for the conversion of existing data centers to AI use. The capacity threshold for eligibility is to be set by presidential decree^[5][25]. The Ministry of Science and ICT held a public forum on the subordinate legislation on September 9, 2026^[6]. Whether the Ulsan, Haenam, and Gumi projects qualify for the exemption will therefore depend on the capacity threshold in the enforcement decree and on each project's specific conditions.

Business model: leasing. Revenue comes from leasing data center space and supporting infrastructure to customers^[14]. Rent per unit and the leased capacity that generates revenue are the key performance metrics.

Status and facts: LG Uplus broke ground on the 200MW Paju AIDC in May 2025. Building 1 is targeted for completion in June 2027, with all four buildings to come online in stages by 2028^[14][15]. Board-approved investment totals KRW 1.9645 trillion: KRW 615.6 billion for Phase 1 and KRW 1.3489 trillion for Phase 2^[8]. The company has said it will fund the investment from its own cash without external borrowing^[14]. Power supply of 200MW has been confirmed, and according to the company, Building 1 was fully leased before completion^[3]. As of June 5, 2026, construction was about 20% complete^[15].

Analysis: Because the contracted capacity (MW), term, rent per unit, and the pre-leasing status of Buildings 2 to 4 are undisclosed, the exact operating break-even point cannot be calculated^[14][3]. Even though the company has pointed to rising colocation rents, the level of rent actually contracted is unconfirmed^[14]. Keeping the ramp-up period, the lag between signing and actual revenue, as short as possible is what eases the early financial burden.

Business model: commercial (SK–AWS partnership). The project is jointly pursued by SK Group and Amazon Web Services (AWS)^[11]. Because the lease and revenue-sharing terms are undisclosed, it cannot be compared directly with a conventional leasing business.

Status and facts: SK Telecom, SK Ecoplant, AWS, and Ulsan Metropolitan City held the groundbreaking ceremony on August 29, 2025. SK Broadband and other affiliates participate in the project, and SK Ecoplant is responsible for infrastructure construction. In June 2025, SK and AWS had signed the definitive agreement to build the center^[11]. The targets announced by the Ministry of Science and ICT at the time were initial operation of Phase 1 (41MW) in November 2027 and full operation of 103MW in February 2029. The KRW 7 trillion total project cost includes AWS's USD 4 billion investment^[7].

According to press reports in September 2026, power has been secured for the initial facilities now under construction. SK, however, has not officially disclosed the exact capacity, citing a non-disclosure agreement (NDA)^[12]. SK Group Chairman Chey Tae-won said on September 11 that discussions on an additional expansion of about 900MW had made substantial progress^[26], while Korea Electric Power Corporation (KEPCO) explained that no formal administrative application for additional supply had been filed^[12]. Discussions on expansion do not amount to final approval of power supply.

Analysis: It cannot be assumed that the partnership with AWS translates directly into fixed lease income for SK^[11]. Because the breakdown of total investment (construction, equipment, and so on) is not clearly separated, direct comparison with other projects' facility investment is limited.

Business model: public infrastructure. A public-private joint venture that provides AI computing resources to industry, academia, and research institutes^[9]. Alongside usage-fee revenue, the delivery of public computing capacity is a central measure of performance.

Status and facts: The National AI Computing Center in Solaseado, Haenam, is being built with 40MW of utility supply capacity and 15,000 AI chips. The groundbreaking ceremony was held on August 3, 2026, and completion is targeted for 2028^[9].

On project cost, the Ministry of Science and ICT announced KRW 2.5 trillion or more (2026–2030), while Haenam County announced KRW 2.9 trillion (June 2026)^[9][10]. The difference in items covered by the two figures has not been clearly established. Meanwhile, Koark, the special purpose company (SPC) that owns the project and in which Samsung SDS is the largest shareholder, has filed a power use application with KEPCO. From 2027, part of the computing capacity is to be supplied early using Samsung SDS's existing data center resources^[9].

Analysis: Haenam offers usage vouchers and fee discounts to startups and research institutes^[9]. Performance should therefore be judged not only on lease income but also on whether the policy goal of supplying public AI infrastructure is met. Filing a power use application also does not mean that supply approval has been granted.

Business model: owner-operated services. Samsung SDS operates the data center infrastructure itself and provides AI and GPU cloud services to customers^[13]. Service revenue and equipment investment determine the profit structure.

Status and facts: Samsung SDS decided to build a 60MW AI data center on the site of its Gumi Plant 1 and, on January 2, 2026, disclosed an investment of KRW 427.3 billion. This amount covers building and facility construction, and further investment may follow as AI infrastructure such as GPUs is added^[13][16]. The original plan targeted groundbreaking in July 2026 and operation in March 2029, but as of September 15 neither the actual start of construction nor the result of the power system impact assessment had been confirmed.

Analysis: Building and basic facility investment alone cannot support a simple comparison of unit costs or cost advantages against other data centers. Gauging the profitability of the AI and GPU services Samsung SDS will operate itself requires a combined look at the service revenue outlook and the scale of additional equipment investment, including high-cost GPU servers.

Comparing break-even points requires a precise understanding of which costs are included. The 'operating break-even' used in this report means the level at which revenue covers pure operating expenses (OPEX) and interest on borrowings. This differs from accounting break-even, which reflects depreciation, and must also be distinguished from the payback period, the time it takes to fully recover the initial capital invested.

Rents and contract terms. Colocation rents in the Seoul metropolitan area rose from about KRW 140,000 per kilowatt (kW) per month in 2019 to about KRW 250,000 in 2025. Of new metropolitan supply, the share committed through occupancy or pre-leasing was high, at 99.7% in 2024 and 99.4% in 2025, and 16% and 23.5% of the supply scheduled for completion in 2027 and 2028, respectively, is already pre-leased^[2]. Because the time remaining to completion differs, however, the gap in pre-leasing rates alone is not enough to conclude that demand has weakened.

In the domestic market, the common contract structure discounts rent in the early period and then fixes the annual escalation at about 2%. Large global contracts of 10MW or more, by contrast, are increasingly set for terms of 10 years or longer with 3% annual escalation or indexation to the consumer price index (CPI)^[2].

Hanam IDC, for example, charges rent per kW on the committed capacity and settles electricity separately per kilowatt-hour (kWh)^[27].

Reference cases: Gwacheon IDC and Hanam IDC

Gwacheon IDC: In July 2025, LS Securities estimated that Gwacheon IDC would need an average annual utilization rate of about 25% to break even. The estimate takes into account rising depreciation and the interest burden on borrowings, and differs from the actual outcome^[28]. This utilization rate cannot be applied to other data centers as is.

Hanam IDC: According to Macquarie Korea Infrastructure Fund's investor relations materials for the first quarter of 2026, Hanam IDC has signed leases covering 99% of its target IT load (25.44MW). The company expects contracted rent on the full target IT load to begin accruing around mid-2027^[29]. Even after leases are signed, there is a lag before revenue actually flows in, and that point is not the same as reaching break-even.

Financing. Even with power approval and tenants secured, a project can stall if financing fails. In 2026, a data center development in the metropolitan area triggered an event of default (EOD) after it failed to pay the land balance on time, despite having signed a 20MW power use contract with KEPCO, and creditors demanded early repayment^[30]. By contrast, the Yongin Deokseong-ri AI data center (total project cost about KRW 1.4 trillion), which has not disclosed an anchor tenant, secured KRW 710 billion in main project financing (PF) in April 2026^[30]. This is why the mere disclosure of a tenant should not be taken as a guide to financing prospects.

What to check for each project

Paju AIDC: The actual contracted rent per unit and initial discount terms, the contracted volume and term, and the pre-leasing rate of Buildings 2 to 4 need to be established. The timing of rent collection after Building 1 is completed, and the level of operating costs and interest on borrowings, should also be confirmed.

SK AIDC Ulsan: Whether the Phase 1 facility now under construction begins normal operation should be checked. The SK–AWS contract structure needs to be clarified, whether it is a plain lease or a revenue-sharing arrangement, along with the split of facility investment and operating costs. The power application and approval status for the additional 900MW expansion, and whether the investment is finalized, should also be monitored.

National AI Computing Center, Haenam: Before operation, final approval of the 40MW power supply and adherence to the completion schedule should be confirmed. After operation, the level of public subsidies, the fee structure, the actual computing capacity delivered and its utilization, and the cost burden of fee subsidies warrant review.

Samsung SDS Gumi: The actual start of construction and passage of the power system impact assessment should be confirmed first. Once those two conditions are established, the scale of additional GPU and server investment, the status of customer demand, and the trend in AI and GPU service revenue should be tracked.

When a data center reaches break-even is determined by rent per unit, occupancy, and financing terms. Even for facilities of the same size, the point at which operating costs and interest are covered varies widely with the rent actually contracted, when revenue begins accruing on contracted capacity, and the interest burden of financial borrowings. Even with pre-leases signed, delays in tenants' actual move-in and rent collection inevitably add to the early financial burden^[27][28][29].

① Limitations of the analysis and data

• Scope of verification: We compared four projects under construction or in development that were presented in public announcements at 40MW or more. The definition of capacity differs by project. The reference date of the analysis is September 15, 2026. For Gumi, the start of construction is unconfirmed. Gwacheon IDC and Hanam IDC are used as reference cases for break-even estimates and the timing of lease revenue.
• Selection criteria: Paju, Ulsan, Haenam, and Gumi were selected from projects for which the owner, investment plan, and operating or completion target could all be confirmed in public sources. The four cases were used to compare the leasing, commercial partnership, public infrastructure, and owner-operated service models. They are neither a survey of all domestic projects meeting the capacity threshold nor a representative sample.
• Cases excluded from the detailed comparison: The detailed comparison by business model was limited to the four projects above. GS Donghae is introduced as a reference case that passed the power supply review. GAK Sejong, Kakao Ansan, Hanam IDC, and Epoch Anyang, which are already in operation, are treated as reference cases separate from the new projects. Yongin Deokseong-ri is introduced only as a financing case.
• Scope of market data: Rents, pre-leasing rates, and contract practices are secondary citations via KHARN, cross-checked against the CBRE Korea original (pp.20, 24, 31)^[2][31]. Pre-leasing rates are measured on the IT load (MW) of metropolitan supply assets and do not represent all AIDCs in Korea. Whether rents are quoted per kW of IT load and whether electricity is included are not specified in the original either.
• Discrepancies between sources: The project cost of the National AI Computing Center differs by source: the Ministry of Science and ICT announced KRW 2.5 trillion or more, while Haenam County announced KRW 2.9 trillion.
• Limits of applying the reference cases: The approximately 25% for Gwacheon IDC is the average annual utilization needed to break even, as estimated by LS Securities in July 2025 from cost and revenue projections. It is neither an actual outcome nor an industry-wide benchmark. The 99% for Hanam IDC is a lease signing rate, and mid-2027 is the company's expected point at which rent on the full contracted load begins accruing. Because the two cases differ in metric and cost scope, the figures cannot be compared directly^[28][29].
• Scope of investment figures: Haenam includes AI chip procurement, while Gumi is centered on the building and basic facilities. For Ulsan, the items included in the total investment could not be separated either, so investment amounts cannot be compared directly across projects. Paju's KRW 1.9645 trillion is the sum of the approved amounts for Phases 1 and 2. Whether the separately reported land purchase of KRW 105.3 billion is included in this sum could not be confirmed from the original disclosure. We therefore did not add the land purchase to the total, nor did we present the sum as a total including land.
• Limits on procedural status: Whether the power system impact assessment notice has been enacted, whether Samsung SDS Gumi has broken ground, and whether GAK Sejong Phases 2 and 3 have started construction could not be confirmed as of September 15, 2026.
• Nature of sell-side materials: The Samsung Securities and LS Securities materials^[27][28] are brokerage research covering the respective stocks. They were used to confirm contract structures and the basis of break-even estimates; investment opinions and target prices were not cited.

② Basis for key figures

Ratios and totals we calculated ourselves are distinguished from estimates and projections presented by sources.

③ Methodology

• Source hierarchy: Official materials from companies and regulators take precedence, followed by raw data and aggregator databases, specialist media, syndicated media, and community sources, in that order.
• Source classification: A source is classified as primary only when the original document was verified directly. Where the original could not be obtained, the citation is marked as secondary and the fact that the original was not obtained is stated.
• Reference and retrieval dates: Values fixed at a specific point in time carry a reference date; dashboard figures that change in real time carry a retrieval date.
• Handling of conflicting figures: Where differing figures are found, both are presented. When a reference value must be chosen, disclosure of the calculation scope, clarity of the retrieval date, and separability of line items take precedence over the prominence of the outlet, and the reason for the choice is stated in the text.
• Our own estimates: The formulas for ratios and totals we calculated ourselves, and the basis of estimates and projections presented by sources, are set out in Appendix ②.
• Legal and regulatory interpretation: Regulator materials are cited only to the extent of explaining the general classification of structures and their potential applicability, and are not extended into an official classification or approval of any specific product. Descriptions of product structures and regulatory interpretation are kept in separate paragraphs.
• Conflict of interest disclosure: The same disclosure wording is placed in a fixed position at the top of every report.