SK Ulsan AI Data Center Nears 900 MW, but Capacity Still Has to Become Compute

SK Group says the SK Ulsan AI data center is approaching a 900-megawatt expansion plan, far beyond the 100-megawatt project first announced with Amazon Web Services. Chairman Chey Tae-won disclosed the figure at the 2026 Ulsan Forum, according to reporting published September 13. He also said discussions with additional global technology companies were progressing quickly.

The number changes the scale of the story. This is no longer only an AWS facility scheduled to begin operating in the second half of 2027. SK is pitching Ulsan as the first anchor in a much larger Korean AI infrastructure network. Yet the 900-megawatt figure describes an expansion plan, not operating computing capacity available today.

That distinction creates the central tension. SK has land, energy businesses, telecommunications networks, memory expertise, and an established cloud partner. It must still secure power, customers, financing, equipment, and staged construction before those advantages become usable AI capacity. The pressure is therefore on SK to convert an ambitious campus plan into contracted infrastructure that customers can deploy.

The SK Ulsan AI Data Center Has Outgrown Its Original Plan

The important change is not a completed 900-megawatt facility, but the speed at which SK has enlarged the project’s intended footprint.

Chey told reporters that the Ulsan expansion had “almost come all the way to 900 megawatts.” His remarks followed continued discussions with large technology companies. He said SK expected to announce further details soon, according to the latest Ulsan expansion coverage.

SK and AWS began with a more defined arrangement. Their 2025 agreement called for SK to construct the data center and AWS to establish an AI Zone inside it. An AI Zone is local infrastructure that provides AWS computing, networking, and managed AI services while keeping workloads in South Korea.

Operations were expected to begin in 2027 under a 15-year strategic partnership. SK said the site would support dedicated AWS hardware, high-speed UltraCluster networking, Amazon SageMaker, Amazon Bedrock, and Amazon Q. The arrangement combined a physical data center project with a specific cloud-service deployment.

That initial commitment gave Ulsan something many proposed AI campuses lack: a recognized anchor customer and a defined service model. AWS was not simply lending its name to a property development. It planned to operate cloud and AI infrastructure within the campus while SK supplied construction, connectivity, energy coordination, and related capabilities.

SK Telecom later described Ulsan as a facility that could grow to one-gigawatt scale. In July 2026, the company said it wanted more than two gigawatts across South and North Gyeongsang provinces. It placed that regional cluster inside a national ambition totaling as much as 15 gigawatts.

Chey’s latest statement narrows the gap between that one-gigawatt aspiration and a more immediate proposal. However, it does not explain how much of the nearly 900 megawatts has approved grid access, committed financing, signed customers, or a construction schedule. It also does not replace the phased opening of the original facility.

Earlier project materials described a much smaller first operating stage. The first phase was associated with approximately 41 megawatts, followed by a later expansion above 100 megawatts. These figures can coexist with the new claim if SK treats 900 megawatts as the campus’s planned development envelope.

They should not be treated as interchangeable. An energized building that can support installed servers is different from land reserved for future data halls. A signed power agreement is different from an application awaiting review. A campus master plan is different from contracted information technology load, meaning the electricity consumed by computing equipment rather than the wider property.

SK’s wording matters for that reason. “Nearing 900 megawatts” appears to describe progress in assembling expansion partners or plans. It does not mean the site currently operates at that level. Construction on the initial facility remains underway, with the first service expected during the second half of 2027.

The news is still consequential. Moving from a 100-megawatt opening concept toward nearly 900 megawatts would turn the SK Ulsan AI data center into infrastructure built for multiple large customers. It would also make power procurement and tenant commitments more important than the design of any single building.

Why Ulsan Fits SK’s AI Infrastructure Strategy

Ulsan gives SK an unusual way to combine cloud demand with assets spread across telecommunications, semiconductors, construction, and energy.

The project brings together several SK businesses. SK Telecom leads the broader data-center strategy. SK Broadband contributes network infrastructure and operating experience. SK hynix sits upstream as a major supplier of high-bandwidth memory, which feeds data to AI accelerators at very high speed.

Other affiliates address physical delivery. SK Ecoplant can support engineering and construction. SK’s energy businesses provide experience with gas, electricity, and industrial utility systems. SK AX contributes information technology integration and data-center operations.

This structure does not guarantee successful execution. It does reduce the number of essential capabilities that SK must source from unrelated providers. The group can coordinate land, power, networking, memory, construction, and enterprise technology within one corporate family.

Ulsan strengthens that model because it is already one of South Korea’s major industrial centers. The city hosts energy, chemical, automotive, and shipbuilding operations. Those industries create potential customers for AI inference, industrial simulations, process optimization, predictive maintenance, computer vision, and robotics.

Chey framed Ulsan as a testing ground for manufacturing AI. That goal links the computing campus with nearby factories rather than treating it only as remote cloud capacity. SK Innovation’s Ulsan complex already applies AI-assisted monitoring to refining and chemical operations, providing one example of the industrial demand SK wants to cultivate.

The geographic choice also responds to South Korea’s concentration problem. Much of the country’s existing data-center market sits in or around greater Seoul, where proximity to customers has traditionally outweighed regional diversification. Power availability, permitting delays, and local opposition have made that concentration harder to sustain.

A large Ulsan campus would move new computing demand toward an established energy and industrial base. It would still need low-latency links to Seoul and other population centers. SK Telecom and SK Broadband can address that requirement through dedicated data-center interconnection routes, which connect distant facilities using high-capacity fiber networks.

AWS adds another layer. The original AI Zone partnership said Korean organizations could run AI workloads locally while using the company’s cloud services. That supports data residency, the practice of storing and processing information within a specified jurisdiction.

Local processing can matter to government agencies, regulated industries, and enterprises with sensitive operational data. It also gives AWS a way to expand Korean AI capacity through infrastructure developed with a domestic group.

AWS is therefore both a customer and a reference point. Its involvement validates the initial phase, but a campus approaching 900 megawatts would need demand extending beyond one announced partnership. Chey’s promise of more global technology partners suggests SK recognizes that requirement.

SK Telecom’s broader plan also depends on partnerships. The company says a typical gigawatt-scale AI data center requires financing from strategic investors, long-term customer contracts, and project-finance structures. That model ties each stage of construction to credible future revenue.

The strategy resembles an infrastructure platform more than a conventional telecommunications expansion. SK would supply sites and operating systems while cloud providers, model developers, governments, and enterprises consume computing capacity. Success depends on coordinating those commitments before expensive equipment and power infrastructure sit idle.

This is why the 900-megawatt statement deserves attention even before completion. It signals that SK believes discussions have advanced enough to enlarge the Ulsan plan. The next disclosures must show whether that confidence rests on signed commitments or preliminary interest.

Nearly 900 Megawatts Creates a Power and Tenant Test

SK’s main opponent is the gap between announced capacity and infrastructure that has power, customers, hardware, and revenue.

Electricity is the first constraint. AI servers run accelerators at high utilization and concentrate far more power inside each rack than conventional business computing. Cooling systems, network equipment, and backup infrastructure add further demand.

A campus approaching 900 megawatts would represent a substantial continuous load. The exact relationship between campus capacity and IT load remains undisclosed, so the figure should not be converted directly into a GPU count. Different cooling designs, redundancy standards, accelerator generations, and utilization rates produce different results.

SK argues that Ulsan offers favorable energy conditions. The group has access to gas infrastructure, industrial sites, and affiliates with experience managing large energy systems. Chey also emphasized the need for stable power during his forum remarks.

Those advantages do not remove the grid process. South Korean developers must pass power-system impact assessments before receiving large new supplies. Regional projects have encountered uncertainty during those reviews, even as national policy encourages construction outside greater Seoul.

As of March 2026, supply availability had not been confirmed for 9,583 megawatts of data-center applications nationwide, according to reporting based on Korea Electric Power Corporation figures. The grid review backlog shows why announced capacity cannot automatically be counted as future operating capacity.

Busan and Ulsan had a lower reported share of unresolved applications than several other regions. Even so, a project of this size needs more than a generally favorable regional outlook. It requires specific connections, generation arrangements, transmission capacity, redundancy, and delivery dates.

Cooling is the second physical test. The original Ulsan facility was designed around hybrid cooling, which combines different techniques to remove heat more efficiently across varying workloads. Higher-density AI systems increasingly use liquid close to chips because air alone becomes less effective as rack power rises.

A staged campus can adapt its cooling design as hardware changes. It can also create integration risk if later data halls use different power and thermal architectures. SK must balance rapid construction against the possibility that accelerator requirements shift before each phase opens.

The third constraint is equipment. AI capacity needs accelerators, high-bandwidth memory, networking switches, storage, and electrical hardware. SK hynix gives the group exposure to a critical memory component, but it does not supply every layer of a computing cluster.

SK Telecom has also partnered with Nvidia on Korean AI infrastructure. That relationship provides another potential route to accelerators and reference architectures. AWS, meanwhile, can deploy its own Trainium and Inferentia chips alongside other supported hardware inside its cloud environment.

These options can reduce dependence on a single accelerator route. They also make the project more complex. Different customers want different chips, software frameworks, networking designs, security controls, and commercial terms.

The decisive constraint is contracted demand. A developer can reserve land and design electrical capacity years before customers occupy it. An anchor tenant signs for a meaningful block of capacity, giving lenders and developers evidence that the project can produce revenue.

AWS provides that role for the initial AI Zone. The identity, capacity commitment, and timing of the additional partners mentioned by Chey remain unknown. Until SK discloses those details, the 900-megawatt figure is best read as an advanced expansion target.

That does not make it empty. Large infrastructure projects often begin with land, power studies, customer negotiations, and phased financing long before full construction. It does mean readers should distinguish development progress from operating compute.

South Korea’s AI Ambition Raises the Competitive Stakes

Ulsan is part of a national contest to secure computing capacity, but SK is competing for the same power, hardware, capital, and customers as every other operator.

SK Telecom has framed its 15-gigawatt plan as an effort to make South Korea an Asian AI infrastructure hub. Under the roadmap, the company would start with Ulsan, expand beyond two gigawatts across the southeastern region, and add capacity elsewhere in the country.

The company said it aims to activate five gigawatts in stages beginning in 2029, with a longer-term path toward 15 gigawatts. Its national buildout announcement explicitly identified power, sites, anchor tenants, and operating systems as essential parts of the plan.

That roadmap is much larger than the capacity SK operates today. It also reflects a broader change in competition. Telecommunications companies once treated data centers mainly as network-adjacent real estate. AI turns them into capital-intensive computing factories where electrical engineering, chip supply, cloud software, and customer utilization determine returns.

Domestic rivals approach the market from different positions. Naver operates its own cloud platform and data centers, including the GAK Sejong facility. Its model connects infrastructure with Korean-language AI services, enterprise software, search, and sovereign computing requirements.

KT and LG Uplus combine telecommunications networks with enterprise data-center businesses. Samsung SDS provides cloud and enterprise infrastructure through another large Korean corporate group. International operators and cloud providers add competition for tenants, equipment, land, and specialized staff.

SK’s differentiation is its corporate span. It can connect SK Telecom’s networks, SK hynix memory, AWS cloud services, and group energy assets. The Ulsan project tests whether that full-stack argument produces better execution than a more focused cloud or colocation strategy.

AWS is also expanding its Korean presence through spending and infrastructure beyond this one site. Its Ulsan partnership helps SK attract global cloud demand, but it also gives AWS influence over the service layer. SK must ensure that its role extends beyond building facilities for a larger platform.

That tension is particularly relevant to sovereign AI. South Korea wants local computing that supports domestic companies, public institutions, and sensitive workloads. Infrastructure located inside the country advances that goal, even when a foreign cloud provider operates part of it.

However, physical location does not answer every sovereignty question. Customers also care about control over encryption keys, software dependencies, model access, operational authority, and applicable contracts. A Korean data center can improve data residency without making its entire technology stack domestic.

The competitive stakes reach beyond South Korea. Japan, India, Malaysia, Singapore, and other Asian markets are attracting new data-center investment. Each offers a different mix of power, connectivity, regulation, land, customers, and political support.

Ulsan’s industrial base offers a distinctive demand story. It could support both regional cloud workloads and factory-focused AI deployments. Yet South Korea has higher construction costs and tighter land and grid conditions than some alternative markets.

Latency also matters. Training a model can occur far from end users if networks move data efficiently. Interactive inference, which produces responses from a trained model, benefits more from proximity to customers and applications.

SK therefore needs a balanced customer base. Large training clusters can absorb substantial power but produce uneven project cycles. Enterprise inference can create steadier demand, though it requires successful adoption across many organizations.

Chey acknowledged the commercial issue directly. He warned that large AI investments need business models capable of producing returns and supporting reinvestment. That concern prevents the story from becoming a simple race for the biggest announced number.

The industry’s challenge is not only to build capacity. It must keep expensive accelerators busy with workloads customers value enough to fund. Ulsan will pressure SK’s competitors if it secures that demand at scale, not merely because its master plan contains a larger megawatt figure.

What the 900 MW Figure Does Not Yet Prove

The expansion claim leaves four unresolved questions: what has been contracted, what has been approved, what will open on schedule, and who will pay to use it.

First, SK has not published a detailed breakdown of the nearly 900 megawatts. The figure could include the initial AWS facility, future data halls for other tenants, supporting infrastructure, or capacity under negotiation. Public reporting does not establish how much represents firm customer commitments.

Second, the power status remains unclear. SK has highlighted Ulsan’s energy infrastructure, including nearby gas resources and group affiliates with relevant capabilities. It has not publicly mapped the latest figure to approved grid connections or dedicated generation schedules.

Third, the financing structure has not been disclosed. SK Telecom has said large projects will require a mix of corporate capital, strategic investment, long-term contracts, and project financing. Those sources become easier to secure when anchor customers commit to long leases or minimum usage.

Fourth, the operating timeline remains staged. The initial SK Ulsan AI data center is expected to begin service in the second half of 2027. A campus approaching 900 megawatts would almost certainly take longer to complete than that first phase.

This schedule distinction is important because hardware evolves faster than buildings. Data-center shells, substations, cooling plants, and fiber routes must serve several generations of accelerators. Designs that optimize for today’s rack density can require modification before later phases open.

Construction also faces supply-chain pressure. Transformers, switchgear, generators, cooling equipment, and specialized labor can have long lead times. A project with power approval can still miss its target if critical electrical systems arrive late.

Environmental performance deserves scrutiny as well. SK has promoted energy-efficient design and access to stable gas-based power. Gas can improve local supply reliability, but it still produces carbon emissions unless paired with credible mitigation measures.

AI customers increasingly track the carbon intensity of their computing. Cloud providers also carry climate commitments that influence procurement and facility design. SK will need to explain how the enlarged campus aligns with those requirements as its power demand grows.

Water use is another open question. Hybrid and liquid cooling can improve thermal performance, but water consumption varies significantly by system and climate. Public materials have not provided a complete water plan for a campus near the newly discussed scale.

Local economic claims also require careful treatment. SK has previously projected substantial direct and indirect employment as the site expands. Construction creates a large temporary workforce, while highly automated data centers usually employ fewer permanent workers than factories consuming similar land or power.

The more durable regional effect could come from adjacent businesses. Cloud teams, AI developers, equipment suppliers, maintenance specialists, and industrial software companies can form around reliable computing infrastructure. That outcome depends on Ulsan attracting users rather than functioning only as a remote server location.

None of these uncertainties disproves SK’s plan. They define the difference between a plausible expansion and a completed one. Chey’s remarks suggest negotiations have advanced, but the next announcement must provide enough detail to evaluate the quality of that progress.

Readers should also resist converting megawatts into intelligence or economic output. Power capacity measures an input. Useful AI output depends on accelerator efficiency, software, networking, utilization, model quality, and customer applications.

A smaller cluster running continuously on valuable workloads can outperform a larger but underused campus financially. Conversely, a well-filled 900-megawatt site would give SK and its partners a significant foundation for Korean cloud and AI services.

The right interpretation is therefore conditional. SK has assembled a credible starting position in Ulsan and expanded its ambition rapidly. The project’s real scale will become visible only through binding commitments and physical delivery.

Three Signals Will Show Whether SK Can Deliver

Partner commitments, power milestones, and the 2027 opening will determine whether the SK Ulsan AI data center becomes infrastructure or remains an oversized plan.

The first signal is a named global technology partner with a defined role. Chey said further announcements should come soon, making this the most immediate test. A useful disclosure would identify whether a partner is an anchor tenant, investor, chip supplier, cloud operator, or construction collaborator.

Those roles carry different weight. A vendor agreement can help delivery but does not prove customer demand. An anchor tenancy with committed capacity would strengthen the business case and support financing for additional phases.

A vague memorandum would provide less evidence. It could mark the beginning of negotiations without establishing how much capacity the company will use. Readers should watch for contract duration, opening phase, capacity allocation, and any minimum commitment.

The second signal is confirmed power delivery. SK does not need to energize 900 megawatts at once, but it should publish a credible sequence connecting each construction phase to electricity supply. Grid approvals, substations, on-site generation, transmission work, and redundancy plans will matter more than another rounded capacity target.

Power disclosure would also clarify the meaning of Chey’s figure. If the company connects the expansion to secured or scheduled supply, the plan becomes more concrete. If it continues discussing capacity without corresponding energy milestones, the verification gap grows.

The third signal is the initial opening in the second half of 2027. The first phase will test whether SK, AWS, contractors, and regulators can deliver the facility on schedule. It will also reveal what hardware and cloud services are available when customers arrive.

That launch should be judged by usable capacity, customer activity, and service availability. A ceremonial opening matters less than energized data halls running production workloads. Evidence of manufacturing customers would further support Chey’s vision for Ulsan as an industrial AI test bed.

The opening will not validate all 900 megawatts. It will establish whether the delivery system behind the larger plan works. On-time operation would strengthen confidence in later phases, while delays would make the expansion schedule harder to defend.

For developers and enterprise buyers, the practical question is where reliable regional compute will be available and under whose software stack. More Korean capacity can improve workload placement, latency, procurement options, and data residency. It can also intensify competition among cloud and infrastructure providers.

The reported expansion deserves attention because SK has more than an empty site proposal. AWS has a defined initial role, construction is underway, and SK controls businesses across several required infrastructure layers. Those facts give the ambition a credible foundation.

The unanswered questions are equally material. Nearly 900 megawatts is not yet operating capacity, and SK has not disclosed enough to treat it as fully contracted. Watch the promised partner announcement first, the power sequence second, and the 2027 launch third. Together, those signals will show whether the SK Ulsan AI data center is becoming a functioning AI hub or simply a very large number on a development map.