← Back to feed

SK Telecom (NYSE: SKM) announced July 5 that it will build AI data center capacity totaling 15 gigawatts, centered on a GW-scale Ulsan campus with 5GW to be activated in stages from 2029. The project pulls in the full SK Group stack — compute, operations, site infrastructure — with SKT coordinating overall. No financing total or capex figure was disclosed.

The framing SKT chose is deliberate: the company describes the buildout as “Korea’s third national infrastructure revolution,” placing it in sequence with the Gyeongbu Expressway (1970) and the country’s high-speed internet rollout in the early 2000s. Both prior examples were government-coordinated industrial investments that restructured the economy around new infrastructure. SKT is asking investors and policymakers to read the 15GW commitment the same way.

The AI G3 Play

The strategic anchor is Seoul’s “AI G3” programme — a government target to position Korea alongside the US and China as one of three leading AI powers globally. The framing treats AI compute as foundational infrastructure, equivalent to electricity or bandwidth, that must be nationally owned and scaled. SKT’s buildout is the delivery mechanism.

That framing matters for how the capital will flow. A commercially-driven data center build optimises for tenants, IRR, and utilisation curves. A national infrastructure build optimises for capacity, sovereignty, and geopolitical positioning — and tends to tolerate lower near-term utilisation in exchange for strategic readiness. SKT’s announcement is structured as the latter.

The Ulsan anchor is geographically deliberate. Korea’s industrial heartland has existing heavy-grid infrastructure from its automotive and shipbuilding legacy, along with logistics supply chains capable of handling large construction projects. It is also inland, not concentrated in Seoul’s congested data center corridor.

The SK Group Structural Advantage

SK Hynix — the SK Group semiconductor arm — holds 60-70% of NVIDIA’s Vera Rubin HBM4 allocation and is one of three companies globally that dominate high-bandwidth memory supply. The same conglomerate now plans to build the data centers that need HBM.

That vertical stack is unusual. Most hyperscalers buy memory from suppliers; most memory suppliers do not build hyperscale data centers. SK Group would be doing both simultaneously, inside the same corporate structure, at the scale of national infrastructure.

Whether this becomes a coordination advantage or a conflict of interest depends on how the buildout is managed. The more SK Telecom’s data centers succeed in attracting frontier model inference workloads, the more demand flows to SK Hynix. The more SK Hynix supply tightens, the more expensive the data centers become to build. The conglomerate structure resolves that tension internally — but the pricing and allocation decisions will matter to external tenants.

Scale Context

Nine global cloud giants are projected to spend $830 billion on AI infrastructure across all of 2026. A single national telecom committing to 15GW is a meaningful data point, but the timeline is long: 5GW by 2029 is the first checkpoint, with the full 15GW target not given a hard date. The early phases will test whether sovereign AI infrastructure commitments translate into actual compute capacity or remain aspirational.

SK Telecom said it is reviewing power supply, siting, and operations as core inputs. Given Korea’s energy mix and the scale of demand, power supply is the primary constraint. The Ulsan campus will need reliable access to grid capacity that competes with the city’s existing industrial load — the same political economy that has blocked or delayed data center expansion in other jurisdictions.

The 15GW announcement positions Korea alongside Japan ($6.2B for a domestic AI foundation model consortium), France (AION’s $10B bid for a 1GW EU gigafactory), and India’s YottaCloud and CtrlS buildouts as the cohort of non-US, non-China governments treating AI infrastructure as a sovereign priority. The total capital mobilised across that cohort is now measurable in the hundreds of billions.