SK Ulsan AI Data Center Nears 900 MW, but Profitability Is the Real Test

SK Group says the SK Ulsan AI data center project has expanded toward 900 megawatts, even as Chairman Chey Tae-won warns about investment returns.

That combination matters more than the headline capacity. SK is accelerating one of Asia’s largest planned concentrations of AI computing infrastructure. Its chairman is simultaneously questioning whether AI services can generate enough profit to support the capital flowing into them.

Chey made that tension explicit at Ulsan Forum 2026. He argued that AI investment needs sustainable business models and warned that insufficient returns could produce an investment bubble. Amazon Web Services remains the anchor partner for the initial Ulsan facility, but SK expects other global technology companies to join later.

The latest figure also requires careful interpretation. The first facility was introduced as a roughly 100 MW development, with operations planned for 2027. Nearly 900 MW appears to describe a much broader expansion around that foundation, rather than a fully financed facility approaching immediate completion.

The SK Ulsan AI data center has therefore become a test of two competing timelines. Computing capacity, substations, and industrial sites must be developed years in advance. Reliable demand and profitable AI applications often become visible much later.

The SK Ulsan AI Data Center Has Grown Beyond Its Original Scale

The important change is not a routine construction update. SK has moved Ulsan from a single data center project toward a gigawatt-scale computing cluster.

At the September 2026 forum, Chey said capacity associated with the Ulsan development had expanded to nearly 900 MW. He also said SK expected to announce additional partnerships with global technology companies as the project advanced.

His comments followed a visit to the construction site on September 10. SK executives from its telecommunications, energy, and industrial businesses joined the inspection, according to a site visit account published by Yonhap News Agency.

The expansion builds on an agreement announced in June 2025. SK Group and AWS established a 15-year strategic relationship covering an AWS AI Zone in Ulsan. An AI Zone brings dedicated computing, networking, and managed cloud services closer to customers that need local infrastructure.

Under the original AWS partnership, SK would lead data center construction. AWS would install and operate cloud and AI capabilities for customers in South Korea.

The first phase was presented as a facility costing about 7 trillion won. Ulsan officials said it would use GPUs and other AI processors for model training and inference. They initially expected construction to start in 2025 and commercial operations to follow in 2027.

Later municipal materials described a 41 MW first operating phase and 103 MW at full completion in 2029. SK communications generally referred to a 100 MW-class facility. Those figures are broadly consistent with one physical campus being delivered in stages.

The new 900 MW number is different. It pushes the project close to the gigawatt threshold, where infrastructure planning becomes more similar to a major industrial complex than a conventional data center.

However, SK has not publicly provided a complete breakdown connecting every megawatt to a construction phase, financed building, signed customer, or grid connection date. The Korea Times account says capacity “under construction” has approached 900 MW, while earlier official disclosures describe a smaller initial facility and future expansion.

That distinction matters. Announced capacity can include planned phases, adjacent sites, reserved power, or a wider cluster. Operational capacity means servers are installed, connected, commissioned, and serving workloads.

Readers should therefore treat 900 MW as SK’s updated project scope, not as 900 MW of computing already available. The figure demonstrates ambition and planning momentum, but it does not settle the delivery schedule.

The project still represents a major escalation. A 100 MW deployment can support a substantial AI cloud region. A 900 MW cluster requires far more land, grid coordination, cooling equipment, network capacity, financing, and customer commitments.

That larger scale creates the central conflict around SK’s strategy. The company must reserve infrastructure before global customers make final placement decisions. Waiting for demand would reduce financial risk, but it could also leave SK without capacity when buyers need it.

Why Ulsan Became SK’s Starting Point

Ulsan gives SK an unusual combination of industrial land, energy operations, manufacturing demand, and access to international connectivity.

The southeastern city already hosts large petrochemical, automotive, shipbuilding, and energy facilities. Those industries offer potential customers for manufacturing AI, including systems used for process optimization, inspection, simulation, and predictive maintenance.

SK also has an established local footprint. Its affiliates operate energy and chemical facilities in the region, while SK Telecom supplies network and data center expertise. SK hynix gives the group direct exposure to high-bandwidth memory, which feeds data to AI processors.

This arrangement supports SK’s preferred “full-stack” infrastructure pitch. The group can combine power development, telecommunications, memory, cloud operations, construction, and industrial applications without relying on one subsidiary.

Ulsan’s government identified several other advantages when it announced the initial project. They included large industrial sites, water access, nearby energy infrastructure, and proximity to submarine cable landing facilities around Busan.

Submarine cables matter because Ulsan’s market extends beyond local manufacturers. A regional AI hub needs low-latency connections to customers and other cloud regions across Asia-Pacific.

Energy integration is equally important. Ulsan officials have discussed LNG infrastructure, renewable energy, hydrogen, and waste-cold-energy systems that can assist data center operations. These resources do not eliminate environmental costs, but they give planners more options than a congested urban site provides.

The location also fits South Korea’s regional development goals. Much of the country’s current data center activity remains concentrated around Seoul and the surrounding Gyeonggi region. Grid constraints and community resistance have made further concentration harder.

Moving capacity south can distribute electricity demand and attract technology investment outside the capital region. It also aligns computing infrastructure with industrial data generated by factories, ports, and energy facilities.

That industrial connection is central to Chey’s argument. He said manufacturing AI requires datasets that individual companies or regions cannot assemble alone. He called for broader cooperation capable of supporting scale.

The practical use cases are easier to understand on a factory floor. A refinery can use sensor histories to identify abnormal equipment behavior. A shipyard can analyze production schedules and component movement. An automaker can train computer-vision systems for quality inspection.

These workloads do not automatically require every dataset to leave its owner’s systems. Companies can use secure cloud environments, controlled data-sharing structures, and domain-specific models. Still, each approach needs governance before industrial data becomes reusable at scale.

Ulsan gives SK access to the industries that can test these systems. It does not guarantee that companies will share data, redesign workflows, or pay enough for the computing involved.

That is why the location offers both opportunity and pressure. SK must turn nearby industrial activity into recurring AI workloads. Otherwise, Ulsan’s advantages remain inputs without a profitable output.

Scale Is Pressuring SK to Secure Power and Customers Early

A project approaching 900 MW cannot advance through server purchases alone. Power access and long-term customers become the controlling variables.

AI data centers place dense, variable loads on electricity systems. Training and inference clusters can change consumption rapidly as processors coordinate large workloads.

South Korea’s government started formal grid-code work in August 2026. The effort addresses standards and connection strategies for AI data centers, including millisecond-scale power fluctuations measured in megawatts.

A grid code defines how large electricity users must behave when connecting to the power system. It can include voltage support, fault response, load controls, and technical requirements that protect other customers.

For SK, securing nominal power capacity is only the beginning. The company must coordinate substations, transmission access, backup systems, and delivery schedules. It must also show that individual data halls will not destabilize the surrounding grid.

Cooling creates another constraint. AI servers concentrate more heat than traditional cloud equipment. Operators increasingly use liquid cooling, which moves heat through coolant closer to processors instead of relying only on room-scale air conditioning.

The technology can improve cooling efficiency, but it adds specialized piping, maintenance, and equipment requirements. The selected server architecture also affects electrical distribution, floor loading, and heat-rejection design.

SK must make many of these decisions before knowing the exact mix of future processors. AI hardware changes quickly, while data center electrical systems operate for decades.

Customer concentration poses a parallel risk. AWS anchors the original Ulsan development through a 15-year partnership, giving the first phase a credible source of demand. The wider expansion needs more commitments.

Chey said SK plans to disclose additional partnerships. Until those agreements appear, readers cannot assess how much of the 900 MW scope is reserved, contracted, or speculative.

An anchor tenant is especially important for project financing. A long-term customer contract can support loans by providing predictable future revenue. Planned capacity without contracted demand carries more exposure to changing cloud requirements.

SK’s structure reflects that challenge. In July 2026, SK Telecom created SK Hyper, a wholly owned subsidiary focused on data center development. The subsidiary will handle site acquisition, substations, customer recruitment, and commercialization.

SK Telecom approved capital contributions of up to 750 billion won through 2030. That commitment funds important groundwork, but it is far below the total cost of SK’s broader infrastructure vision.

The company acknowledges the gap. Its 15 GW roadmap says a typical 1 GW AI data center could require approximately 70 trillion won. SK expects to use partner investment, long-term contracts, and project financing alongside its own capital.

Those numbers reveal why customers matter as much as electricity. Reaching 900 MW cannot rest on an assumption that general AI demand will eventually absorb every server.

SK must identify who will use the capacity, which processors they require, how long they will commit, and what returns the contracts support. Each answer influences financing and construction.

AWS reduces uncertainty for the initial site. It does not automatically underwrite every later phase or SK’s regional expansion.

The pressure therefore falls back on SK. It must announce credible global partners, prove industrial demand, and obtain grid access while construction costs remain high.

Chey’s AI Bubble Warning Changes the Meaning of 900 MW

The strongest critique of SK’s expansion came from its own chairman: infrastructure spending cannot continue unless AI produces sustainable profits.

Chey told the Ulsan Forum that AI needs better monetization models. He argued that profits must eventually support continuing investment and warned that an investment bubble could emerge if returns remain insufficient.

The warning appeared in the same remarks that disclosed the enlarged Ulsan capacity. That makes it more than a general comment about technology markets.

SK is exposed to both sides of the AI spending cycle. SK hynix supplies memory used in accelerators, while SK Telecom and SK Hyper want to develop the facilities housing those systems. Other affiliates contribute energy, construction, networking, and enterprise services.

When AI capital spending rises, those businesses can reinforce each other. Strong memory demand encourages new computing deployments. New deployments consume more memory, network capacity, and power.

The same integration also concentrates risk. Slower AI adoption could weaken demand across several parts of SK’s strategy at once.

Chey’s call for a sustainable AI model recognizes this problem. The economic test is not whether customers experiment with AI. It is whether they keep buying enough computing after early trials.

Industrial AI makes that question particularly difficult. Factory deployments require integration with equipment, operational data, safety processes, and employee decisions. A technically accurate model can still fail to create financial value if deployment costs exceed the savings.

Data quality also limits scale. Manufacturing information often sits across incompatible systems and document formats. Equipment histories can be incomplete, and access rules vary among plants and suppliers.

Enterprises need repeatable methods for collecting, governing, and retrieving that information. An engineering knowledge base can support the document layer, but it does not replace operational integration or model validation.

SK’s opportunity is to combine infrastructure with industry-specific services that address those deployment barriers. Selling electricity and server space alone leaves the project exposed to commodity pricing.

The group could offer managed environments for manufacturing models, data governance, simulation, digital twins, and secure inference. These services could create higher-value relationships than basic hosting.

However, SK has not publicly quantified how much Ulsan demand will come from industrial customers, general cloud workloads, sovereign AI programs, or overseas technology companies. Each category carries different margins and utilization patterns.

Utilization is the key bridge between megawatts and profitability. A data center earns a return when contracted computing revenue exceeds financing, energy, hardware, staffing, and maintenance costs.

High nameplate capacity does not guarantee high utilization. Operators can build electrical shells before servers arrive, install equipment before workloads mature, or reserve space that customers never fully consume.

AI hardware can also become obsolete faster than the buildings supporting it. A facility designed around one generation of accelerators may need electrical and cooling upgrades for later systems.

That does not mean Ulsan will become stranded infrastructure. AWS provides an experienced operator and a customer platform, while South Korea has strong semiconductor and manufacturing industries.

It does mean SK’s nearly 900 MW statement needs a commercial counterpart. Signed customers, phased financing, delivered power, and active computing matter more than the headline figure.

SK’s Bigger Plan Extends Far Beyond One Ulsan Campus

The Ulsan expansion is the first visible piece of a proposed national network reaching 15 GW, not an isolated real estate project.

SK Telecom announced in July that it would pursue more than 2 GW of AI data center capacity across the southeastern Gyeongsang region. Ulsan would serve as that cluster’s starting point.

The company also described another 1 GW in the southwestern Jeolla region. Together with other sites, SK aims to begin opening 5 GW of domestic capacity in stages from 2029.

Its longer-term target reaches 15 GW by 2035. SK says later construction will depend on market demand and other conditions.

That qualification is important. The 15 GW figure is a roadmap, not a portfolio of completed facilities. Sites, customers, financing, transmission infrastructure, and permits must still align.

For scale, 15 GW represents a continuous electrical load comparable to multiple large power stations. The associated investment would reach levels normally connected with national infrastructure programs.

SK’s vertically connected businesses give it advantages in pursuing such a plan. SK hynix produces high-bandwidth memory, or HBM, which helps accelerators move large volumes of data quickly. SK Telecom operates networks and data centers.

SK ecoplant contributes construction experience. SK Gas and other energy affiliates can support power development. SK AX works on enterprise technology and industrial transformation.

Competitors bring different combinations. Microsoft, Google, and AWS already operate global cloud platforms with large customer bases. They can direct internal demand toward new regions, although local constraints still affect construction.

Samsung has semiconductor, construction, and enterprise technology units. Its position gives South Korea another potential vertically connected infrastructure player, especially as the government seeks more AI capacity outside Seoul.

Telecommunications companies also see data centers as a path beyond mature connectivity businesses. They already understand networks, enterprise customers, and regulated infrastructure, but hyperscale construction introduces much greater capital exposure.

SK’s relationship with AWS partly closes the cloud-platform gap. AWS gains local infrastructure and Korean industrial connections, while SK gains an anchor tenant and access to established AI services.

The partnership also creates a dependency. If the wider Ulsan expansion involves other global technology companies, SK will need to manage competing technical requirements and commercial interests.

Different customers can request separate server designs, accelerators, security boundaries, and sustainability commitments. A multi-tenant expansion offers diversification, but it can reduce design simplicity.

SK has presented standardization and modular construction as possible responses. Prefabricated data center modules can integrate servers, power distribution, and cooling equipment before delivery to a site.

Modularity can shorten deployment schedules and repeat proven designs. It cannot bypass grid connections, customer procurement cycles, or local approvals.

The regional plan also raises questions about coordination with national energy policy. South Korea must decide where generation, transmission, and large computing loads should grow together.

Developing sites outside the capital can reduce one form of congestion. It can create new regional bottlenecks if generation and transmission projects lag behind data center construction.

The 900 MW Ulsan figure therefore matters as an early test for the entire 15 GW strategy. If SK can align power, customers, and industrial demand there, the model becomes easier to repeat elsewhere.

If Ulsan experiences delays or weak utilization, the long-term roadmap will require slower phases, different partners, or revised economics.

Three Signals Will Show Whether the Expansion Is Real

The next evidence should come from contracts and infrastructure milestones, not another increase in announced capacity.

The first signal is a detailed capacity schedule. SK needs to clarify how the nearly 900 MW figure relates to the original 100 MW facility, adjacent phases, and the wider Ulsan cluster.

A useful disclosure would separate capacity under active construction from permitted, powered, financed, and planned capacity. It would also identify commissioning dates for each phase.

If SK provides that breakdown, confidence in the expansion should increase. Continued ambiguity would weaken comparisons between Ulsan and other large data center projects.

The second signal is another anchor customer. Chey said new partnerships with global technology companies would be announced as the project progresses.

The identity and commitment level of those partners will matter. A memorandum to explore cooperation carries less financial weight than a long-term capacity contract.

A partner willing to reserve computing or data hall capacity would strengthen SK’s financing case. It would also show that demand extends beyond the original AWS AI Zone.

No additional agreement would prove the project is failing. However, a long delay would make the 900 MW expansion look more dependent on expectations than contracted demand.

The third signal is a secured power and grid schedule. Large AI clusters need substations, transmission connections, technical operating standards, and reliable generation.

South Korea’s work on AI data center grid rules indicates that policymakers recognize the issue. The next step is applying those rules to specific projects without shifting unacceptable costs or reliability risks to other users.

Confirmed connection milestones would strengthen SK’s target. Delayed transmission work or unresolved load requirements would place later phases at risk, even if customer interest remains strong.

Commercial performance will eventually become the final test. Data center revenue, signed capacity, utilization, and returns will show whether Chey’s sustainable investment model is emerging.

Those figures may take longer than three months to appear. Near-term disclosures can still reveal whether SK is moving from a vision toward executable projects.

The SK Ulsan AI data center now sits at the intersection of South Korea’s semiconductor strengths, regional industrial policy, and global demand for AI computing. Its planned scale gives SK a chance to create an infrastructure business spanning memory, energy, networks, cloud services, and manufacturing applications.

Scale also raises the cost of being wrong. Nearly 900 MW requires customers that remain committed after the current investment cycle, plus a grid capable of serving them reliably.

For developers and enterprise buyers, the immediate question is not whether more computing will become available. It is what services, processor choices, data controls, and contract terms will accompany that capacity.

Watch SK’s next partner announcement, its phase-by-phase power schedule, and the first facility’s 2027 opening target. Together, those signals will show whether Ulsan is becoming a functioning AI hub or remaining an ambitious capacity plan.