South Korea's AI Data-Center Buildout Is Outrunning Its Power Grid
KT Cloud laid out a two-track expansion plan this week that doubles as a tacit admission: the company cannot build AI data centers fast enough on its own. One track has KT Cloud self-funding new facilities in stages, starting with a site in Bucheon next year and following with projects in Gaebong, Yongin, Ansan, and Gunsan. The second hands part of the buildout to outside developers through an expanded leasing model, letting KT Cloud book capacity it doesn't have to construct itself. The target across both tracks is roughly 8x its current footprint: about 125 megawatts today, scaling to 1 gigawatt of AI-ready capacity by 2031.
Power equipment, not chips, is the binding constraint
What stands out is where KT Cloud is actually spending its partnership energy. Rather than racing to lock up GPU allocations, the company has spent recent months signing deals with power-and-cooling specialists: a September memorandum of understanding with LS Electric covering extra-high-voltage transformers, gas-insulated switchgear, and switchboards for its modular data center sites, and an October 28 strategic partnership with Vertiv focused on co-designing power, liquid cooling, and modular systems before construction starts. KT Cloud says it is also reviewing whether to retrofit existing facilities for AI workloads at all, and has concluded that power delivery and cooling capacity — not floor space — are what make many older buildings unsuitable for today's high-density GPU racks.
That sequencing is telling. Transformers, gas-insulated switchgear, and UPS systems have manufacturing lead times that can stretch over a year, and modular, factory-built designs exist specifically to compress the construction schedule around those bottlenecks rather than around chip delivery. KT Cloud's data center division head has promised to match "the supply speed and scale the market expects," but the company is effectively racing its own power-equipment supply chain, not Nvidia's.
The national grid is the harder ceiling
KT Cloud's company-level scramble sits inside a much larger structural problem that South Korean grid experts have been flagging independently for months. Electrical engineering researchers have warned that clustering new AI data centers in the already-dense Seoul metropolitan area — where more than 70% of the country's data centers already sit — risks grid overload and transmission bottlenecks, and have urged policymakers to push new builds toward the regions and ease direct power-purchase rules so developers can source electricity without routing everything through the capital grid.
Wood Mackenzie's analysis of the government's flagship semiconductor-and-AI cluster plan for the Honam region puts a number on that mismatch: peak electricity demand there is projected to roughly double, from about 11.8 gigawatts today to 19.1 gigawatts once the cluster is built out, against a forecast regional generating capacity of only 16.8 gigawatts by 2030 — a shortfall of roughly 2.3 gigawatts. The firm's proposed fixes read like a to-do list for a grid under real time pressure: extending the life of aging nuclear units to preserve capacity that would otherwise retire, prioritizing domestically built gas turbines because imported ones carry five-to-six-year lead times, and reinforcing the already-congested Yeongnam-Honam transmission corridor with non-wired measures to relieve the bottleneck sooner.
A buildout racing a plan that isn't finalized yet
The uncomfortable overlap is timing. KT Cloud's 1-gigawatt target and its first new site both land before 2030 — the same year Wood Mackenzie expects Korea's regional power shortfall to open up, and before the government's 12th Basic Plan for Electricity Supply and Demand, which is supposed to clarify how much new generation and transmission actually gets built, has been finalized. Seoul has responded with policy tools — a Distributed Energy Act, grid-impact assessments, and a pledge to mobilize up to 30 trillion won in public-private AI investment — but researchers close to the issue have been explicit that permitting support alone doesn't solve a problem that is fundamentally about power supply, land, and cooling capacity arriving on the same schedule as demand.
For a market watching compute supply as the main AI-industry chokepoint, Korea's case is a useful corrective: the gating resource increasingly isn't the accelerator, it's the transformer, the switchgear order, and the transmission corridor behind it. KT Cloud's bet that it can hit 1 gigawatt by 2031 is really a bet that South Korea's grid planning moves on the same clock as its own construction schedule — and the independent analyses of that grid suggest the two clocks aren't yet synchronized.