SK Hyper, the AI data center subsidiary SK Telecom incorporated in July, disclosed Thursday that its first build phase will require roughly $350 billion, a figure that encompasses not just land, construction, and electrical infrastructure but the full range of compute hardware that customers would deploy inside its facilities.
CEO Chung Suk-geun laid out the numbers on Bloomberg TV's "Bloomberg Tech: Asia" on October 23, describing a 5-gigawatt first phase and 15-gigawatt target by 2035 that would make South Korea one of the largest AI compute markets in the world. The financing strategy behind the headline number is designed to avoid the overbuilding risk associated with large-scale data center programs: SK Hyper plans to raise capital in demand-verified tranches rather than on projected demand alone, adjusting each phase of investment as hyperscaler bookings become concrete.
Chung was direct about where the real difficulty lies. Site access and power supply are manageable in Ulsan, South Korea's heavy-industrial southeastern port. "Securing actual demand in the global market and raising the capital needed to build the data centers will be the biggest challenges," Chung said. The company is targeting its first outside investment close before the end of 2026, with additional tranches expected over a three-to-four-year period.
SK Hyper's buildout is organized around three regional clusters, each centered on a gigawatt-scale campus. The first, and the project furthest along, is in Ulsan, roughly 305 kilometers (190 miles) southeast of Seoul. SK Group has been developing the Ulsan site since September 2025 in partnership with Amazon Web Services, which is committing $4 billion of the facility's estimated ₩7 trillion ($5.29 billion) total cost. The campus is being built around an initial deployment of 60,000 GPUs, making it Korea's largest GPU cluster, with a phased target of approaching 900 megawatts of capacity and an ultimate aim of more than 1 gigawatt on site.
From that southeastern base, SK Hyper plans to add gigawatt-scale campuses in the Chungcheong region in the center of the country, then in the southwestern Honam area. The full 15-gigawatt buildout would extend through 2035, with the 5-gigawatt first phase expected to begin delivering capacity by 2029, subject to actual customer demand.
SK Telecom itself has committed to invest up to ₩750 billion (approximately $567 million) in SK Hyper in phases through 2030, providing the subsidiary's foundation capital. That figure is seed funding relative to the investment that would follow: the ₩464 trillion ($350 billion) first-phase estimate implies an average of roughly $70 billion per gigawatt when compute hardware is included, consistent with SK Telecom's own engineering estimate that a 1-gigawatt AI data center campus requires approximately ₩70 trillion (approximately $53 billion) in construction-only investment.
How LNG cold energy cooling changes the economics at gigawatt scale
Ulsan's most important technical advantage is its physical proximity to a liquefied natural gas regasification plant, an advantage that capital investment alone cannot replicate elsewhere.
SK Gas operates an LNG combined-cycle power plant adjacent to the Ulsan Mipo National Industrial Complex where SK Hyper is building. When LNG, stored at -162°C / -260°F, is warmed and vaporized into pipeline gas, the process releases an enormous quantity of cold energy, approximately 830 kilojoules per kilogram of LNG processed. At a large-scale terminal, this cold energy represents hundreds of megawatts of available thermal capacity. Historically, this cold has been dissipated into seawater as waste. LNG cold energy recovery instead routes it through heat exchangers into a chilled-water loop, providing a near-zero-marginal-cost cooling source for industrial facilities next door.
For a hyperscale AI data center, cooling is not a secondary concern. Thermal management typically consumes 30 to 40 percent of total facility energy, and the standard industry efficiency metric, Power Usage Effectiveness (PUE), measures how much energy a facility spends on overhead (cooling, lighting, power conversion) for every unit of energy that reaches the actual compute hardware. World-class hyperscale facilities achieve a PUE of 1.1 to 1.2; a facility running conventional chillers at scale is more likely to operate around 1.4 to 1.6. The difference, multiplied across a gigawatt of AI compute, amounts to hundreds of megawatts of overhead power that either needs to be generated or does not.
LNG cold energy from an adjacent terminal eliminates the compressor energy that chillers require, effectively driving the cooling contribution to PUE toward zero marginal cost. Japan has used this technique near LNG terminals for decades. Osaka Gas's Senboku facility pioneered industrial LNG cold recovery in the 1970s, and NTT and SoftBank have operated data center cooling systems near LNG terminals since the 2000s. The Ulsan site is positioned to apply the same physics at a scale that no competing location in Japan or Taiwan can currently match, because the adjacent LNG plant eliminates piping distance that limits the approach elsewhere.
Chung noted the advantage plainly, describing how SK Gas's plant provides stable, large-volume power while also enabling "the use of liquefied natural gas cold energy for cooling," which Chung described as "a meaningful efficiency advantage in a business where electricity and thermal management dominate operating costs."
Financing model that tracks confirmed demand
The $350 billion first-phase figure is notable for what it includes: compute hardware, including servers and GPU racks, alongside land, construction, electrical infrastructure, and substations. This makes the estimate closer to a total cost of ownership figure than a conventional capital expenditure number for physical facilities.
The financing structure is deliberately designed to track confirmed demand rather than move ahead of it. SK Hyper plans to share the total project cost with data center customers and outside investors, raising tranches over a three-to-four-year period following the first outside investment, expected in late 2026 or early 2027. The company is in active discussions with both global big tech firms and financial investors. Chung rejected the idea that capital market volatility would derail the fundraise. "There is a lot of uncertainty in the capital markets, but AI model developers and service providers can feel customer demand rising sharply," Chung said.
This demand-linked approach is a form of project finance risk management: each capital tranche is drawn down only as customer commitments become concrete, so that the overbuilding risk common to speculative infrastructure programs is converted into a series of staged options based on a verified order book. The difficulty is that it requires hyperscaler customers to make capacity commitments years ahead of need, which is the commercial challenge Chung has described as the primary obstacle.
Chung also said revenue could begin as early as 2027, well ahead of the 2029 target for phased capacity delivery, suggesting some initial workloads may be hosted on the growing Ulsan campus before the broader 5-gigawatt program reaches completion.
Why Japan and Taiwan are the target customers
The case SK Hyper is making to global compute buyers rests on the size of the domestic market: South Korea's entire current market is 1 gigawatt. The 5-gigawatt first phase alone is five times that figure, and filling it with domestic enterprise and sovereign AI demand is not commercially viable.
"Given the scale of the domestic IT industry, Korea has not been an attractive location in the global data center industry until now," Chung said. The proposal is to attract inference and training workloads from neighboring markets, specifically Japan and Taiwan, by offering what neither can easily provide at the moment: abundant, stable power on politically neutral, technologically sophisticated territory, combined with a conglomerate that controls semiconductors, energy, and telecommunications within the same group.
SK Hynix, SK Group's memory chip business, is the world's dominant supplier of high-bandwidth memory, the HBM chips inside every NVIDIA H100 and H200 GPU. No competing data center developer globally controls both the chip supply and the physical compute infrastructure in the same corporate family. Chung cited this directly: "Until last year, the key competitive edge was who had secured the most graphics processing units. Now what matters is who can generate tokens at the lowest cost and achieve the highest efficiency."
The sovereign AI dimension reinforces the proposal to regional buyers. "Many countries and companies are worried about relying solely on specific AI models developed overseas," Chung said. "They need not only their own AI models but also sovereign AI infrastructure they can control directly at home." Korea's location, close to Tokyo and Taipei, outside China's sphere of control, and integrated into global semiconductor supply chains, positions it as a politically neutral compute hub for companies that want to diversify away from US-controlled hyperscalers without moving to Chinese alternatives.
Chung leads the Global Telco AI Alliance, a consortium of major carriers including KDDI, Deutsche Telekom, and e&, a role that gives SK Hyper direct channels into the exact class of corporate AI infrastructure buyers it needs as principal customers for its capacity. Before joining SK Telecom, Chung led Naver's Clova AI product unit, one of Korea's most significant commercial AI deployments; Chung's KAIST computer science degree and Georgetown MBA give credibility on both the technical and financial sides of a deal that will require both.
Regulatory changes in South Korea in 2026
The timing of SK Hyper's capital raise is not incidental. South Korea spent 2026 building a regulatory and financial framework specifically designed to make this kind of project feasible.
The most significant move was legislative. The National Assembly passed the Special Act on Promotion of the Artificial Intelligence Data Center Industry on May 7; the law was promulgated June 9 and takes effect on March 10, 2027. The legislation introduces a one-stop approval system with a permit-timeout mechanism, exempts certain non-capital-region AI data centers from power-grid impact assessments, and eases facility installation requirements. It is one component of what President Lee Jae-myung's administration has branded the "Three Mega Projects for Korea's Great Leap Forward," a national AI strategy combining semiconductors, physical AI, and AI data centers.
That framework took a further step on October 22, the day before Chung's Bloomberg TV appearance, when President Lee chaired the second meeting of the Regulatory Rationalization Committee and the Ministry of Science and ICT announced a new package of measures. The ministry said it would minimize permits required to build AI data centers, exclude unmanned facilities such as server rooms and cooling equipment from building-size calculations, and reduce parking, EV charging, and ancillary requirements that have historically added cost and time to large structures. The government also announced a new ₩200 billion (approximately $151 million) program to localize the supply chain for AI data center materials, parts, equipment, and software, including a testbed cluster for verifying domestic solutions alongside commercial facilities.
Special-zone designations, which confer levy reductions on operators, are expected in the first half of 2027, with community benefit plans required as a mandatory condition to manage local opposition in host regions. Proposed enforcement decrees were expected to be issued for public notice at the end of October. The South Korea national plan extends to 18.4 gigawatts across SK Group (15GW), GS Group (2.4GW), and Naver (1GW).
Asia's AI infrastructure competition and Korea's position
SK Hyper is building amid accelerating regional competition. Asia-Pacific data center capacity is projected to grow substantially through 2030, with AI training and inference workloads driving a structural shift in how compute capacity is allocated and where it is located. Western hyperscalers and Chinese cloud leaders together have committed over $160 billion to regional AI infrastructure across secondary markets in Asia.
Japan and Taiwan, the two demand markets SK Hyper most directly wants to attract, are themselves investing in domestic compute. Japan is doing so through government-backed initiatives including its ABCI 3.0 sovereign AI supercomputer, while Taiwan is expanding hyperscaler cloud regions tied to its semiconductor ecosystem. But neither has Korea's combination of available industrial land, adjacent LNG energy infrastructure, and conglomerate-scale execution capacity positioned at the same moment.
The AWS $4 billion Ulsan commitment signals external validation from a relevant customer. AWS has committed to deploying 60,000 GPUs in an initial cluster while pledging further AI-focused infrastructure investment in the region. OpenAI, meanwhile, is separately advancing plans for a Stargate AI data center in Korea in partnership with Samsung Electronics and SK Hynix, deepening the country's profile as a destination for frontier AI compute.
The scale SK Hyper is proposing, 15 gigawatts and $350 billion for the first 5 gigawatts, places Korea alongside the largest announced AI infrastructure programs anywhere in the world. The company knows the challenges are real: demand commitments from global customers, not construction, are the limiting factor, and Chung has said so plainly. But the combination of a purpose-built subsidiary, an enabling legislative environment, a government treating AI infrastructure as a national priority, and an existing principal customer in AWS, supported by an LNG cooling advantage and a demand-linked financing model designed to track rather than guess the market, gives SK Hyper a position that few competitors at this scale have established simultaneously. The next few years will show whether that is enough to fill 15 gigawatts, starting with whether Chung closes the first outside investment before the end of 2026.
Currency conversions in this article are approximate and based on rates as of October 23, 2026.
Frequently asked questions
What is SK Hyper, and how does it relate to SK Telecom?
SK Hyper is a wholly owned subsidiary of SK Telecom, South Korea's largest wireless carrier, formed on July 23, 2026. It was spun out specifically to develop and operate AI data centers (AIDCs) at scale, separating the capital-intensive infrastructure buildout from SK Telecom's core telecom operations. The parent company has committed up to ₩750 billion (approximately $567 million) in seed funding through 2030, with the bulk of the $350 billion first-phase investment expected to come from external data center customers and financial investors.
What makes LNG cold energy cooling different from standard data center cooling?
Conventional data centers use electrically powered chillers to remove heat from server hardware, accounting for 30 to 40 percent of total facility energy. LNG cold energy cooling replaces that electrical load with thermal energy that would otherwise be wasted: when liquefied natural gas is warmed from -162°C (-260°F) to pipeline temperature, the cold it releases is captured via heat exchangers and routed into the data center's cooling loop at near-zero marginal cost. For a 1-gigawatt facility, the difference between a conventional PUE of 1.5 and an LNG-assisted PUE of 1.1 can represent over 400 megawatts of saved overhead power, capacity freed for additional compute instead. The Ulsan campus benefits because SK Gas's plant is adjacent, eliminating the distance limitation that constrains this technique elsewhere.
What does South Korea's new AI data center law actually change for operators?
The Special Act on Promotion of the Artificial Intelligence Data Center Industry, passed by the National Assembly on May 7, 2026, promulgated June 9, and taking effect March 10, 2027, streamlines the permitting process through a one-stop approval system with binding permit timelines, exempts qualifying non-capital-region facilities from power-grid impact assessments, and eases facility installation requirements that previously applied to large structures. A second regulatory package announced on October 22 further reduces the permits needed for AI data centers, excludes unmanned server rooms and cooling equipment from building-size calculations, and cuts parking and ancillary-facility requirements. Special-zone designations, which will offer levy reductions to operators, are expected in the first half of 2027. Together, these measures reduce the time and cost between a signed hyperscaler deal and operational capacity.
How is SK Hyper planning to fill 15 gigawatts when South Korea's entire current data center market is 1 gigawatt?
The explicit strategy is to attract workloads from Japan and Taiwan, two markets with large AI compute demand but limited ability to deploy capacity at industrial scale quickly. South Korea offers stable power in a politically neutral jurisdiction (a concern for companies reluctant to depend solely on US-controlled hyperscalers or Chinese alternatives), physical proximity to Japan and Taiwan, and SK Group's integrated control of chip supply (SK Hynix HBM memory), energy (SK Gas LNG), and telecommunications (SK Telecom). The demand-linked financing model, drawing capital only as customer commitments arrive, means SK Hyper is not betting on filling 15 gigawatts before locking in infrastructure spend: it is offering to build to confirmed demand. Revenue is expected to begin as early as 2027, well ahead of the 2029 target for formal first-phase capacity.
Original article: techtimes.com
English brief by Korea Wind Intelligence, koreawindintel.com
