SK Group AI Data Center Plan Turns Ulsan Into a Gigawatt-Scale Test

SK Group says its Ulsan expansion is approaching 900 megawatts, which would lift the SK Group AI data center project toward 1 gigawatt. That is roughly ten times the capacity of the initial facility now under construction. The statement signals rapid commercial progress, but it does not yet amount to a completed expansion agreement.

Chairman Chey Tae-won disclosed the progress after the 2026 Ulsan Forum on September 11. He said negotiations with global technology companies were moving quickly and that SK expected to announce more details soon. The update shifts the story from a single AWS-linked campus toward a much larger infrastructure hub serving multiple customers and computing platforms.

The scale is the headline, but Chey also supplied the conflict. He warned that the AI investment cycle needs viable business models that generate returns and support continued spending. SK is therefore pursuing enormous computing capacity while its own chairman questions whether the industry has established the revenue needed to sustain that capacity.

The SK Group AI Data Center Is Moving Beyond Its Original Plan

SK’s latest disclosure turns a defined construction project into a much larger capacity commitment that still needs named customers and binding agreements.

The original Ulsan development centers on a 100-megawatt AI facility being built by SK Telecom and other SK affiliates. AWS plans to operate an AI Zone at the site, offering cloud and AI services through infrastructure located in South Korea. Operations are scheduled to begin during the second half of 2027.

SK and AWS announced their partnership in June 2025. The companies described it as a 15-year strategic arrangement combining SK’s construction, telecommunications, semiconductor, and energy operations with AWS cloud infrastructure. SK would lead construction, while AWS would establish the dedicated AI environment.

An AI Zone is localized infrastructure designed to provide AWS computing and AI services closer to customers. The Ulsan AI Zone is expected to include dedicated servers, high-speed networking, and access to services such as Amazon Bedrock and SageMaker. Local processing can also help organizations address latency and data-residency requirements.

The initial campus already represented a major step beyond SK Telecom’s conventional colocation business. AI facilities require denser racks, more complex cooling, and networks designed to connect large groups of accelerators. Those characteristics make capacity difficult to measure solely through building area or server counts.

Power capacity offers a more useful measure. A 100-megawatt facility is already large enough to support substantial AI training and inference workloads. Moving toward 1 gigawatt places the proposed Ulsan cluster in a different category, closer to a regional computing zone than a single enterprise data center.

Chey said Ulsan had come close to securing the additional 900 megawatts. Reporting on the Ulsan expansion plan described discussions with global technology companies as advanced. However, SK had not publicly identified every prospective tenant or detailed how the additional capacity would be divided.

That distinction matters. Capacity under discussion is not the same as capacity contracted, connected to the grid, financed, constructed, and filled with working systems. Each step introduces a separate schedule and execution risk.

The immediate change is therefore narrower than the headline number suggests. SK has reported strong enough interest to discuss a tenfold expansion, but the company still owes the market a detailed development plan. The next announcement must show which parts of that ambition have moved from negotiation into commitment.

Why Ulsan Can Support a Much Larger AI Campus

Ulsan gives SK something many data center developers lack: nearby industrial land, energy operations, telecommunications assets, and an existing corporate customer base.

The southeastern city is one of South Korea’s largest manufacturing centers. Petrochemical, automotive, shipbuilding, and energy facilities surround the proposed data center location. That industrial concentration gives SK a potential market for computing services beyond training general-purpose language models.

Manufacturers increasingly use AI for visual inspection, equipment monitoring, process control, design simulation, and robotics. These workloads depend on operational data collected inside factories. They also require predictable latency, reliable connectivity, and clear rules governing sensitive production information.

A nearby computing campus can address some of those needs. It reduces the distance between industrial data sources and high-performance computing infrastructure. It also creates opportunities to connect cloud systems with private networks and on-premises equipment.

SK controls assets across the required stack. SK Telecom and SK Broadband provide network and data center experience. SK hynix supplies high-bandwidth memory, which feeds data to AI accelerators. SK Gas and other affiliates contribute energy expertise, while SK AX develops enterprise technology services.

This structure does not guarantee attractive economics. It does, however, give SK more internal coordination options than a developer that must assemble every component through unrelated suppliers. The group can plan land, connectivity, energy, cooling, memory, and customer services as parts of one infrastructure program.

Ulsan also sits outside the Seoul metropolitan area, where competition for suitable sites and grid connections has intensified. Distributing data centers across other regions can reduce concentration, although it also requires sufficient long-distance network capacity and local technical talent.

Local officials argue that Ulsan has substantial electricity-generation capacity. They expect the region’s power self-sufficiency to increase further as additional generation enters service. Still, annual electricity supply and real-time data center availability are different measures.

A gigawatt-scale campus needs dependable power every hour, not simply a regional surplus calculated over a year. Developers must secure transmission capacity, substations, backup systems, and generation that can respond to changing loads. AI computing can create steep demand swings as training and inference jobs start or stop.

Cooling is another constraint. SK’s original design combines air and liquid cooling, with liquid systems removing heat from high-density computing equipment more efficiently. Expansion would require much more cooling infrastructure, water planning, heat rejection, and equipment maintenance.

The location’s industrial history may help with complex infrastructure projects. Ulsan already supports facilities with strict reliability and safety requirements. However, scaling an AI campus from 100 megawatts toward 1 gigawatt remains a major construction and grid-coordination challenge.

Ulsan’s advantage is therefore not cheap space alone. It is the prospect of combining industrial demand with the physical systems needed to supply AI computing. SK must now prove that those pieces can be assembled on the timeline customers expect.

The Real Opponent Is Demand That Fails to Become Revenue

SK is not primarily racing another Korean data center operator; it is racing the gap between announced AI demand and durable customer revenue.

Chey made that tension explicit at the Ulsan Forum. He said the industry needs better business models and commercialization so invested capital can generate returns. He also warned that the AI cycle could deflate if companies fail to create that economic loop.

That concern is unusually direct for the leader of a group planning a vast infrastructure expansion. SK benefits when demand for accelerators, memory, networks, energy, and data center capacity rises. Its exposure also means a demand slowdown could affect several businesses simultaneously.

The initial AWS commitment provides an important foundation. AWS has a large base of enterprise customers and can aggregate workloads across many organizations. A dedicated local zone may attract Korean companies that want AWS services while keeping certain workloads in the country.

Yet AWS alone does not explain a 1-gigawatt campus. The original partnership covers the first phase, while the proposed expansion appears tied to broader discussions with global technology companies. SK needs to disclose whether those talks involve leases, cloud partnerships, accelerator capacity, or less binding expressions of interest.

These contract forms carry different risks. A long-term capacity agreement can support financing and construction. A memorandum can establish strategic intent without guaranteeing utilization. A customer reservation may also depend on power delivery dates, hardware availability, or performance milestones.

Utilization determines whether the campus becomes an operating business or an expensive inventory of electrical capacity. Empty buildings produce little value. Partially filled facilities can struggle because power, cooling, staffing, and financing costs continue before the campus reaches efficient occupancy.

Demand quality matters as much as demand volume. Frontier-model developers can consume enormous computing resources, but their requirements change with each hardware generation. Enterprise inference may prove steadier, although individual workloads are smaller and more sensitive to price.

Manufacturing AI offers another route. SK wants Ulsan to become a hub where local companies can apply AI to industrial operations. That demand could be more durable because it attaches computing to production processes. It may also develop slowly as companies prepare data, validate systems, and integrate models with existing equipment.

Chey highlighted this dependency when he argued that manufacturers need large data platforms. A factory cannot obtain dependable AI results from fragmented, inconsistent, or poorly governed operational records. Infrastructure capacity does not solve that organizational problem by itself.

For enterprise teams, this is where knowledge blending becomes relevant. Combining internal documents, operational context, and structured information helps users obtain grounded outputs. However, the data must remain usable, permissioned, and connected to the work that creates value.

SK’s bet is that demand from cloud providers, model developers, sovereign AI programs, and manufacturers will overlap. If one segment slows, another could keep the infrastructure occupied. The risk is that these customers all depend on the same uncertain assumption that AI spending will produce sustainable returns.

That makes commercialization the primary opponent. The campus can secure land, chips, and power, yet still underperform if customers cannot turn computing into services worth buying. Chey’s warning acknowledges that capacity growth and economic value are not interchangeable.

AWS Is the Anchor, but Nvidia Broadens the Strategy

The proposed expansion shows SK moving from one cloud partnership toward a multi-platform infrastructure strategy.

AWS remains central to the Ulsan project. Its presence supplies a recognizable cloud tenant, an international distribution channel, and a suite of AI services. The partnership also establishes a 2027 operating target against which SK’s construction progress can be judged.

However, SK has since announced a separate infrastructure partnership with Nvidia. That initiative includes plans for SK Telecom to develop 2 gigawatts of AI factory capacity using Nvidia’s DSX platform and Vera Rubin systems. SK hynix would supply HBM4, a new generation of high-bandwidth memory designed for accelerated computing.

An AI factory is Nvidia’s term for infrastructure that converts data and electricity into model training or inference output. The concept emphasizes complete computing systems rather than isolated chips. It includes accelerators, processors, networking, storage, software, power, and cooling.

The Nvidia infrastructure plan gives SK another path to attract workloads. AWS can bring cloud customers, while Nvidia systems can serve organizations seeking dedicated accelerated computing. These routes can coexist if SK manages commercial and technical boundaries carefully.

This broader approach also changes the competitive field. SK is no longer presenting Ulsan only as a facility serving one hyperscale cloud provider. It wants to become an infrastructure operator that can host different platforms and supply customers across South Korea and the Asia-Pacific region.

The strategy draws on SK hynix’s position in AI memory. Accelerators rely on HBM to access large volumes of data quickly. Connecting memory supply, computing systems, and data center operations could help SK coordinate hardware deployment when components remain scarce.

Vertical coordination can also create conflicts. Cloud providers prefer control over their architectures and procurement. Hardware platforms evolve rapidly, making facilities designed for one generation less suitable for another. SK must avoid treating access to affiliated suppliers as proof that every customer will choose its infrastructure.

The expansion also sits within SK Telecom’s much larger roadmap. The company has outlined plans to activate 5 gigawatts of capacity in stages beginning in 2029, then expand toward 15 gigawatts by 2035. It has created a dedicated business, SK Hyper, to lead data center development.

The national buildout roadmap makes Ulsan the first visible test of a much broader thesis. If SK can secure customers, deliver power, and operate the campus efficiently, it gains a template for other regions. Delays would weaken confidence in the larger plan.

Competitors will not wait for that proof. Other Korean technology and infrastructure groups are pursuing regional data centers, while global cloud providers continue expanding their own capacity. Countries across Asia are also trying to attract AI infrastructure through energy access, incentives, and sovereign computing programs.

SK’s advantage is its combination of industrial customers and hardware exposure. Its disadvantage is the complexity of coordinating many affiliates, partners, financing structures, and construction phases. A multi-platform strategy expands the addressable market, but it also increases execution demands.

Power and Construction Will Decide Whether the Scale Is Real

A 1-gigawatt announcement becomes meaningful only when SK can show secured electricity, grid connections, phased construction, and operating customers.

Electricity is the hardest constraint because it cannot be added as quickly as servers. New transmission lines and substations require planning, equipment, permits, and coordination with utilities. Delays in any part of that chain can leave completed data halls waiting for energization.

The problem extends beyond total supply. AI clusters concentrate demand in specific places and can alter their load quickly. Grid operators must balance that demand while maintaining frequency, reliability, and adequate reserves for other customers.

The International Energy Agency expects Korean electricity demand to rise from 2026 through 2030. Semiconductor manufacturing, AI data centers, and wider electrification are the principal growth drivers. That combination means data centers will compete for grid investment with factories and other expanding loads.

The IEA’s East Asia analysis also emphasizes the effect of data centers on grid planning and operations. Local generation can help, but the usable capacity depends on transmission conditions, generation schedules, and the technical requirements of each facility.

Ulsan’s energy base makes it a credible site. SK already operates gas and industrial energy businesses in the region, while local authorities expect new generation to increase supply. The campus may also use energy-management systems and storage to reduce stress during peak periods.

Those possibilities remain different from a disclosed power-delivery contract. SK has not publicly provided a complete schedule showing when each additional block will receive electricity. It has also not detailed the generation mix or how emissions will change as capacity grows.

Cooling must scale alongside electricity. High-density AI racks place more heat inside a smaller area than many traditional servers. Direct liquid cooling can carry heat away efficiently, but it adds pumps, heat exchangers, piping, monitoring, and maintenance requirements.

Hardware cycles create another timing risk. The initial campus is expected to open in 2027, when customers will be deploying a new generation of AI systems. Expansion phases must accommodate higher rack densities without requiring costly redesigns before they begin operating.

Construction capacity could become a bottleneck as multiple Korean projects move forward. Data centers need specialized electrical equipment, backup power systems, cooling components, and trained contractors. Lead times can lengthen when many developers order similar equipment simultaneously.

SK’s corporate reach can reduce some procurement risk, but it cannot eliminate industrywide shortages. The group will need disciplined phasing so that it does not build far ahead of power availability or committed demand.

Environmental questions also deserve scrutiny. A campus of this scale affects land use, water systems, local air quality when on-site generation runs, and the regional electricity mix. Those effects depend on design choices that have not yet been fully disclosed.

The current claim should therefore be read as progress in assembling demand, not proof of completed infrastructure. SK’s next update needs to separate contracted capacity from prospective capacity and identify the power available to each phase.

Without those details, 1 gigawatt remains a destination. With them, Ulsan would become one of the clearest examples of how AI infrastructure is moving from individual facilities toward energy-scale industrial development.

Three Signals Will Show Whether Ulsan Can Deliver

Customer commitments, power milestones, and first-phase operations will determine whether SK’s expansion is a real market response or an ambitious capacity pipeline.

The first signal is a named expansion partner with a defined commitment. SK says discussions with global technology companies are nearing an announcement. Readers should look for contract duration, reserved capacity, project phases, and conditions attached to the agreement.

A binding long-term customer would strengthen the case that the additional 900 megawatts responds to real demand. A general partnership announcement without capacity terms would offer weaker evidence. It could still support future development, but it would not resolve the utilization question.

The second signal is an electricity and construction schedule. SK should explain how much power is secured, when grid connections will become available, and how the campus will add capacity in stages. The company should also identify which phases belong to the AWS project and which support other customers.

Detailed milestones would make the 1-gigawatt target easier to evaluate. They would also reveal whether Ulsan’s industrial energy advantages translate into faster delivery. A vague schedule or repeated delays would weaken the argument that regional sites can avoid the constraints affecting established data center markets.

The third signal is the first 100-megawatt phase entering operation during the second half of 2027. That opening will test construction execution, cooling performance, network readiness, and AWS customer adoption. It will also show whether SK can coordinate its affiliates around one operating campus.

Successful commissioning would support the larger SK Group AI data center strategy. It would provide real utilization and performance data for subsequent phases. Delayed operations or limited customer activity would raise questions about the rest of the roadmap.

Readers should also distinguish capacity from useful AI output. Gigawatts measure infrastructure scale, not model quality, business adoption, or productivity. The economic result depends on what customers run, how often they use it, and whether those workloads create enough value to justify continued spending.

For developers, the Ulsan project could expand access to locally hosted accelerators and cloud services. For enterprise buyers, it could offer more options for sensitive workloads and industrial AI. Knowledge workers will feel the effect indirectly through the availability, latency, and operating costs of AI services.

Teams evaluating those services should track evidence instead of headline capacity. Preserve announcements, contract changes, technical documentation, and deployment results in a searchable AI knowledge base. That record makes it easier to compare promises with delivered infrastructure.

The central question is now measurable: can SK turn reported demand into contracts, electricity, operating systems, and recurring revenue? Watch the next partner announcement, the power schedule, and the 2027 opening. Together, those signals will show whether Ulsan is becoming a genuine gigawatt AI hub or remains a much larger plan on paper.