AWS Ulsan AI Data Center Is Real, but the 15 GW Headline Is a Road Map

Amazon Web Services is advancing the AWS Ulsan AI data center with SK Group, while SK Telecom targets 15 gigawatts of nationwide capacity by 2035. That distinction matters. The facility under construction in Ulsan is not itself a 15 GW campus.

The Ulsan project is the first anchor in a much larger infrastructure strategy. Its initial phase should begin operating in 2027, followed by expansion to approximately 103 megawatts by 2029. SK says it has secured a path toward nearly 900 MW around Ulsan, although most of that capacity remains prospective.

The wider target puts SK against a harder opponent than another cloud provider. It must convert a national 15 GW ambition into financed, powered, occupied facilities. AWS supplies a credible anchor customer, but it does not remove that execution test.

What the AWS Ulsan AI Data Center Actually Includes

The confirmed project is a phased AI facility measured in megawatts, while 15 GW describes SK’s broader national ambition.

SK Group and AWS announced their Ulsan partnership in June 2025. The agreement places a new AWS AI Zone inside infrastructure developed by SK Group in South Korea.

An AI Zone is dedicated AWS infrastructure designed to provide cloud and AI services closer to customers. It differs from a full AWS Region, which normally contains several separate availability zones.

SK leads construction of the data center. AWS plans to deploy and operate its AI Zone within that facility, according to the companies’ Ulsan partnership.

The agreement covers 15 years of cooperation after operations begin. It combines AWS cloud services with SK’s telecommunications, semiconductor, energy, construction, and data center operations.

Ulsan’s government describes a phased build. The first 41 MW portion is scheduled to start operating in 2027. Full completion is planned for 2029 with approximately 103 MW of capacity.

Those figures provide the clearest near-term measurement of what is being built. They also show why the 15 GW description requires context.

One gigawatt equals 1,000 megawatts. A 15 GW portfolio would therefore equal roughly 146 facilities at the Ulsan project’s planned 103 MW scale.

SK does not claim that every future site will copy Ulsan. Later campuses would probably become much larger. The comparison still illustrates the distance between the first project and the national target.

The investment associated with the Ulsan development has been reported at 7 trillion won. That spending is tied to a multiyear buildout, specialized computing equipment, and supporting infrastructure.

Ulsan officials say the completed site will contain high-performance GPUs. These processors handle the parallel calculations used for training and serving large AI models.

The location also carries industrial significance. Ulsan sits outside the Seoul metropolitan area and hosts major energy, chemicals, automotive, and manufacturing operations.

SK wants the center to serve both cloud customers and industrial AI deployments. Potential workloads include refinery optimization, factory safety, predictive maintenance, logistics, and enterprise model development.

The project is already under construction, so this is more than a preliminary memorandum. However, the physical facility remains years away from full planned capacity.

That timing changes how readers should interpret the latest coverage. AWS has joined an active Ulsan development, but SK owns the wider 15 GW expansion goal.

The distinction does not make the announcement less important. It reveals the real story: Ulsan is a test site for a far larger infrastructure thesis.

Why SK Telecom Is Targeting 15 GW Now

SK is treating AI computing capacity as industrial infrastructure, not simply an extension of its telecommunications business.

SK Telecom presented its expanded plan in July 2026. The company aims to activate an initial 5 GW in stages from 2029, then reach 15 GW by 2035.

The planned network begins with Ulsan. SK also expects to develop sites across the Yeongnam region, followed by capacity in Chungcheong and Honam.

Its 15 GW roadmap responds to a basic shift in AI economics. The limiting resource is increasingly the infrastructure surrounding the chips.

AI servers need reliable electricity, high-capacity network connections, cooling systems, substations, land, and long-term customers. Securing accelerators alone does not create usable computing capacity.

SK controls several pieces of that stack. SK Hynix supplies high-bandwidth memory, which feeds data rapidly into AI accelerators. SK Telecom and SK Broadband provide domestic network infrastructure.

Other SK businesses bring energy, engineering, construction, chemicals, and facility-management expertise. This collection gives the group a plausible reason to act as an infrastructure developer.

AWS strengthens the model because it can bring cloud demand and technical requirements. It also reduces the risk of building the first facility without a recognized operating partner.

The relationship reaches beyond a standard property lease. AWS equipment, services, and operational practices are expected to shape the Ulsan AI Zone.

For South Korea, the strategy addresses another pressure. Much of the country’s data center capacity has historically clustered around greater Seoul.

That concentration places new loads near dense populations and constrained transmission networks. Moving AI facilities toward power-producing regions can reduce some connection problems.

Ulsan reported electricity self-sufficiency of 103 percent for 2024. City officials expect its position to improve as additional nuclear and offshore wind generation develops.

Such forecasts depend on projects, policy, and grid access that remain subject to change. Still, the region starts with a stronger energy case than many congested metropolitan locations.

Manufacturing demand adds another reason to build locally. Industrial companies often need low-latency connections and controlled handling of operational data.

An onshore AWS AI Zone can help organizations keep sensitive workloads within South Korea. It can also shorten the path between manufacturing data and cloud-based AI services.

South Korea wants to become a larger AI infrastructure hub for the Asia-Pacific market. That goal connects Ulsan to national policies supporting domestic computing capacity and regional development.

The timing also reflects scarcity expectations. SK cites forecasts that data center demand will continue outpacing available supply in major markets.

However, forecasts are not customer contracts. SK must still prove that demand will materialize in the specific Korean regions where it plans to build.

The 15 GW Plan Puts Execution Ahead of Competition

SK’s main contest is between announced capacity and deliverable capacity, not between AWS and another cloud provider.

A 15 GW target creates an impressive headline because it rivals the scale of national power systems. It also bundles projects at very different levels of certainty.

The Ulsan facility has a named partner, a construction plan, and an operating schedule. Future regional clusters have fewer disclosed sites, customers, power arrangements, or financing structures.

SK has created a dedicated subsidiary called SK Hyper to manage that gap. The company will locate suitable land, secure power connections, build substations, and attract customers.

SK Telecom plans to invest 750 billion won in the subsidiary through 2030. Its SK Hyper structure separates early development work from ordinary data center operations.

That structure acknowledges what hyperscale development requires. A company must spend money on land, studies, permits, and power before a server starts producing revenue.

SK Hyper is supposed to prepare sites and transfer mature projects into operating entities. External investors and strategic partners can then finance portions of each development.

This approach limits how much capital SK Telecom must carry alone. It also introduces dependencies on lenders, infrastructure funds, utilities, and anchor tenants.

The Ulsan project provides an early template. AWS offers recognizable demand, while financial investors can support the property and infrastructure layer.

Yet the same formula cannot automatically fill 15 GW. SK would need multiple customers capable of committing to exceptionally large and sustained computing loads.

Microsoft, Google, Oracle, Meta, local cloud companies, model developers, and national computing programs all compete for chips, power, and engineering capacity. Some also build their own facilities.

AWS itself continues investing in infrastructure across several Asian markets. Ulsan must therefore earn workloads against alternative AWS locations and competing clouds.

The local manufacturing base may differentiate the site. Ulsan can combine cloud computing with operational data generated by factories, refineries, logistics networks, and vehicle production.

That demand profile is different from a campus built mainly for consumer internet services. It could support industrial model training and inference close to production systems.

SK has also discussed an AI Factory with Nvidia. The term describes infrastructure that turns electricity and data into trained models or generated outputs at industrial scale.

Nvidia’s involvement provides another route toward customers and technical integration. However, it should not be confused with the AWS AI Zone agreement.

These partnerships can coexist, but they involve different services, hardware choices, and commercial relationships. Mixing them into one commitment would exaggerate the certainty of the portfolio.

The same caution applies to SK Chairman Chey Tae-won’s statement that the group has secured nearly 900 MW around Ulsan. Secured capacity is not equivalent to installed servers.

It can refer to land, electrical availability, development rights, or projects at varying stages. Investors need more detail before treating all 900 MW as operationally committed.

The Ulsan project therefore matters as a proof point. It must show that SK can coordinate construction, cooling, networks, electricity, AWS deployment, and customer onboarding on schedule.

If that works, the project becomes a repeatable template. If delays emerge, the 15 GW goal will look more like strategic positioning than a development pipeline.

Power Is the Hard Limit Behind SK Telecom’s 15 GW Goal

Electricity supply makes Ulsan credible, but supplying 15 GW across South Korea requires far more than choosing energy-rich regions.

AI data centers create dense and relatively continuous electrical loads. They also require backup systems and cooling equipment that add to the facility’s total demand.

Ulsan offers access to generation and industrial land. SK’s energy affiliates already operate major facilities in the area, creating potential advantages for power procurement and site development.

The city also sits within Yeongnam, a region with substantial nuclear generation. That position supports SK’s argument for moving computing closer to available electricity.

However, regional generation does not guarantee a data center connection. Developers still need substations, transmission capacity, interconnection approvals, and reliable delivery schedules.

The challenge grows when capacity moves from 103 MW toward several gigawatts. Each new cluster must fit within a grid built around existing industrial and residential demand.

The International Energy Agency expects Korean electricity demand to rise from 2026 through 2030. AI data centers, semiconductor plants, and broader electrification are major drivers in its electricity outlook.

That combination creates competition inside the power system. Data centers will not be the only large projects seeking dependable, around-the-clock supply.

Semiconductor fabrication plants need substantial electricity and water. Manufacturing electrification adds more demand. New generation can take years to permit and construct.

Transmission creates another bottleneck. Power-rich regions cannot always move surplus electricity to locations where new projects want to connect.

Wood Mackenzie estimates that South Korea’s major semiconductor and AI plans face a 2.3 GW power shortfall without faster grid investment and electricity-market reform.

Its grid analysis highlights the need to reinforce connections between Yeongnam and Honam. Those upgrades involve long planning, approval, and construction cycles.

SK has discussed additional generation and electrification solutions around Ulsan. It has also raised possible energy cooperation between South Korea and Japan.

Regional cooperation could include joint purchasing, storage, renewable energy, hydrogen, or nuclear technologies. None offers a quick substitute for local grid connections.

The environmental question also remains open. AI facilities can support more efficient industrial operations, but their electricity and water requirements carry direct local costs.

Cooling design will therefore matter. The Ulsan facility plans to combine air and liquid cooling, with liquid systems removing heat from high-density AI equipment.

That hybrid approach can support denser racks than conventional enterprise data centers. It does not eliminate the need for water, pumps, heat rejection, or backup systems.

Efficiency metrics deserve close attention once the first phase operates. Power usage effectiveness measures how much total facility energy is required for computing.

A low ratio indicates less overhead from cooling and power conversion. However, it does not reveal the carbon intensity or absolute amount of electricity consumed.

SK has not yet published operating measurements because the facility remains under construction. Claims about efficiency should therefore remain projections until production data becomes available.

Power prices present a commercial risk as well. SK Telecom has identified rising demand and electricity costs as risks connected with AI data center expansion.

Long-term contracts can stabilize revenue, but they can also lock operators into unfavorable economics. Contract design must allocate changes in electricity costs between SK and its customers.

Ulsan solves part of the siting problem by moving computing toward generation. The national plan still depends on grid policy, new power sources, and detailed regional engineering.

AWS Reduces Demand Risk, but It Cannot Guarantee Returns

A respected anchor tenant validates the first site, while profitability still depends on utilization, contract terms, and disciplined expansion.

Chey has openly identified the economic challenge. He warned that AI investment can resemble a bubble when companies keep adding resources without meaningful returns.

His argument shifts attention from capacity announcements to sustainable business models. Revenue must eventually fund the next round of infrastructure.

Data center economics reward high utilization. Expensive buildings and electrical systems generate weak returns when customers reserve less capacity than expected.

AI demand can also change quickly. Improvements in processors, model architectures, and software may reduce the computing required for a given task.

The opposite outcome is possible too. More efficient models can lower costs, stimulate usage, and increase total infrastructure demand.

SK must plan assets with operating lives measured in decades while AI hardware changes within much shorter cycles. That mismatch increases design risk.

Facilities need enough flexibility to accept future server generations. They must accommodate changes in rack density, voltage, cooling, networking, and physical dimensions.

AWS helps because it understands the infrastructure requirements of large cloud deployments. Its participation gives SK a demanding customer against which to design the first site.

The 15-year agreement also offers a longer commercial horizon than a short equipment cycle. However, the public announcement does not disclose utilization guarantees or revenue commitments.

It remains unclear how much of the Ulsan facility AWS has contractually reserved. The parties have not published pricing, minimum consumption, or renewal provisions.

The phrase “anchor tenant” signals a meaningful role, but it does not tell investors whether AWS assumes capacity risk. Those details determine how much protection SK receives.

Future sites will need equally credible customers. Building capacity before demand appears would expose SK to financing costs without matching operating revenue.

Waiting for complete certainty creates another problem. Power, land, transformers, and construction teams may become unavailable when demand arrives.

That creates the central tradeoff. SK must reserve scarce infrastructure early without committing too much capital before customers sign.

Government support can improve the equation through faster reviews, regional electricity policies, and coordinated land or water planning. It cannot manufacture durable cloud demand.

Local industrial customers could provide another demand layer. Ulsan’s factories produce operational data that can support predictive maintenance, quality control, and energy optimization.

Some workloads will require on-premises systems for safety or latency reasons. Others can use dedicated cloud infrastructure housed nearby.

The commercial opportunity depends on moving beyond experimental projects. Factories must integrate models into production systems and keep using them after pilot funding ends.

That transition is difficult because industrial data often sits across incompatible equipment, old databases, engineering documents, and isolated teams.

Developers and enterprise buyers should therefore watch actual workloads, not only installed GPUs. Utilization indicates whether the facility supports productive services or idle capacity.

SK expects AI infrastructure to become a major business. The first operating phase will test whether partnerships and regional advantages translate into recurring revenue.

What to Watch Before the First 2027 Capacity Goes Live

Three signals will show whether Ulsan is becoming a repeatable infrastructure model or remaining a high-profile pilot.

The first signal is the commissioning schedule for the initial 41 MW phase. SK and Ulsan currently expect operations to begin during 2027.

Commissioning involves more than completing the exterior structure. Teams must test electrical systems, cooling, fire protection, networks, security, and server deployments under realistic loads.

An on-time opening would strengthen confidence in SK’s ability to coordinate its affiliates and AWS. A delay would compress the path toward the 103 MW target for 2029.

Readers should look for precise milestones. Useful disclosures include substation energization, completed data halls, installed cooling systems, and live AWS customer services.

A ceremonial opening without available computing capacity would provide weaker evidence. Operational workloads matter more than a ribbon-cutting date.

The second signal is documented power and customer coverage beyond the first 103 MW. SK says Ulsan can expand toward a GW-scale cluster and cites nearly 900 MW of secured capacity.

The company should eventually identify which portions have grid agreements, approved sites, anchor customers, and committed financing. Those categories should not be treated as interchangeable.

Named projects elsewhere in Yeongnam, Chungcheong, or Honam would strengthen the 15 GW thesis. Broad regional targets without sites would leave it largely aspirational.

Power disclosures are especially important. Investors should watch for interconnection approvals, new substations, generation contracts, and transmission commitments.

Customer disclosures also matter. Another long-term cloud, model developer, public-sector, or industrial tenant would show that Ulsan’s commercial formula can extend beyond AWS.

The third signal is operating economics. SK must demonstrate that the AI data center business can generate recurring returns after electricity, financing, maintenance, and hardware-related costs.

Revenue alone will not answer that question. Capacity utilization, contracted backlog, energy expense, and capital intensity offer a clearer picture.

SK Hyper’s funding arrangements will reveal how the group divides risk. Outside investment can preserve SK Telecom’s balance sheet, but investors will demand predictable returns.

The pace of expansion should follow verified demand. Rapid construction without contracts would weaken the sustainability argument that Chey himself has emphasized.

By contrast, full utilization at the first Ulsan phase would support larger commitments. It would also place pressure on other Korean infrastructure developers to secure their own power and anchor tenants.

The AWS Ulsan AI data center is therefore important for reasons beyond its initial size. It joins cloud demand, industrial geography, and energy strategy inside one measurable project.

It does not yet prove that SK can build 15 GW. That proof requires dozens of decisions involving customers, capital, electricity, land, and public policy.

For developers, the project can determine where additional Korean AI capacity becomes available. It may influence service latency, data residency, hardware access, and regional cloud choices.

Enterprise buyers should watch which AWS services launch inside the AI Zone. They should also evaluate whether those services support production requirements rather than assuming local infrastructure solves every compliance issue.

Knowledge workers may feel the effect indirectly. More regional compute can support faster enterprise AI services, but infrastructure spending only creates value when organizations use it effectively.

Teams tracking these projects need to connect contracts, technical documents, policy changes, and operational milestones. A searchable knowledge base can keep those signals tied to the original evidence.

The next meaningful headline should not simply repeat 15 GW. It should show a powered data hall, a live AWS service, a committed customer, or an approved expansion site.

Until then, treat Ulsan as a credible first build and 15 GW as a conditional road map. Watch the first 41 MW closely, because its performance will determine how believable the next 14.9 GW becomes.