Korea's 8.4GW AI Ambition Hits a Power Wall

The number that changes everything

South Korea announced in June it would build 8.4 gigawatts of AI data center capacity by 2029 — broken down as 5GW for SK Group, 2.4GW for GS Group and 1GW for Naver. The long-term plan stretches to 18.4GW. The announcement read like a bold bet on the AI arms race. It also quietly exposed the single most dangerous gap in Korea’s plan: nobody has figured out how to deliver the power, move the heat or assemble the pieces at that scale.

The problem is not that Korea lacks ambition. It is that ambition outran the infrastructure underneath it.

Rack-by-rack, the math gets brutal

AI data centers are not the same beast as traditional ones. A conventional server rack draws perhaps a few kilowatts. An AI rack now sits at roughly 150kW and will push toward 600kW within a generation — possibly 1MW per rack in the long run, according to Jeong Jeong-ho, a mechanical engineering professor at Ajou University who has tracked the trajectory closely.

That jump is not incremental. It is structural.

The prevailing power architecture inside data centers runs at 48 volts. Low voltage is familiar and relatively safe. It is also a terrible fit for the scale Korea is targeting. Pushing 100kW through a 48V rail requires more than 2,000 amps. At 600kW or 1MW per rack, the current climbs to dangerous, impractical levels — and with it come cable size, heat and transmission losses that grow faster than the power itself.

A paper posted on arXiv in July by a team including Korea University’s Lee Sang-hee flagged the same wall. The authors, working internationally, argued that the old 48V structure and low-voltage AC distribution are hitting their limit precisely because AI workloads spike both total demand and instantaneous current swings at the same time.

Their proposed fixes are not small tweaks. They include raising internal rack voltage to 400V or 800V, rewiring data centers around direct current, and connecting the facility to the grid through medium-voltage solid-state transformers — devices that use power electronics rather than heavy iron to shape and steer current. The physics is sound. The engineering is not yet proven at data-center scale, and the paper lists protection systems, grounding, fault response, long-term reliability and co-design of heat and electricity as open problems.

A data center cannot stop. Designing a new power topology around a facility that must run 24/7 is a different kind of risk. A short circuit in a high-voltage DC system behaves nothing like one in a familiar AC layout. The industry still needs to prove it can shut things down safely, avoid arc flashes, handle grounding and keep systems stable over years — not just in a lab.

Cooling follows the same arc

Electricity that enters a GPU does not mostly become compute. Most of it becomes heat. When a rack jumps from 150kW to 600kW, cooling moves from an afterthought to the binding constraint.

Traditional air cooling simply will not carry that load. The industry is shifting toward liquid cooling — circulating water or dielectric fluid close to the chip instead of blowing air across it. Liquid moves heat far more efficiently than air. But even liquid cooling has reach limits.

Kim Sung-jin, a mechanical engineering professor at KAIST, noted that several companies, including Intel and HP, are experimenting with cooling structures embedded inside the chip package itself. The idea is simple in theory and hard in practice. Integrating cooling channels across many dies on a large package introduces mechanical stress, leakage paths and manufacturing complexity that have not been solved at volume. This is not a near-term fix.

The practical consequence for Korea is direct: every rack design that arrives with higher power will require a tailored cooling solution. There is no generic plug-and-play answer.

The missing integrator

Here is the detail that makes Korea’s posture especially risky. Data centers are not assembled from independent sub-systems that any vendor can swap in. The rack form factor, the power distribution, the cooling layout and the server placement are designed together by the server manufacturer. Change one and the rest must change with it.

Jeong Jeong-ho put it plainly. Korea may be able to produce individual enabling technologies — a pump, a heat exchanger, a power module — but there is no domestic company that designs and builds the complete AI data center system, or the server-rack architecture that ties it all together. That means even if power and cooling somehow arrive on time, Korea would still be assembling someone else’s machine, on someone else’s terms.

He warned that building AI data centers in Korea in a way that serves Korean strategic interests is nearly impossible without that systems-level capability. In effect, Korea is betting on the output without controlling the integrator.

The outside grid is no easier

The power that must reach the data center sits on a separate, equally strained timeline. A study from the University of Calgary examined how rapidly expanding AI loads change more than total megawatt demand — they concentrate demand in time and place, which stresses grid flexibility and reliability in ways that simple capacity expansion does not fix.

Lee Yu-su, an energy policy professor at Sungshin Women’s University, raised a specifically Korean concern. South Korea has talked about routing renewable energy from the Honam region to power new data centers. Renewables are intermittent. AI training runs require stable, round-the-clock power because large GPU clusters cannot pause and resume without losing training state or incurring massive re-committed costs. Without firm capacity — nuclear or liquefied natural gas — paired with renewables, the math does not close.

Transmission construction adds its own drag. Building new lines across Korea involves land acquisition, community opposition, cost overruns and delays. The dates when generation capacity comes online and when data centers need power are unlikely to line up cleanly without deliberate, coordinated scheduling — and that coordination does not yet exist at the scale required.

What this means for the regional AI buildout

Korea’s 8.4GW pledge is visible. The constraints behind it are not. That gap matters for three reasons.

First, it is not unique to Korea. The same power-and-cooling chokepoint appears in Texas, Finland, Japan and elsewhere. Nations that treat AI capacity as purely a software or chip problem will pay in delayed projects and stranded investment. The bottleneck is infrastructure, and infrastructure moves slowly.

Second, Korea’s specific vulnerability is deeper than most. Many countries face a grid problem. Korea faces a grid problem and a systems-integration gap. It can import servers and cooling components, but it cannot easily export a complete data center design the way it once exported memory chips or displays. The value capture sits elsewhere.

Third, the timeline is tight. The 2029 target leaves roughly three years to validate new power architectures, prove liquid and immersion cooling at rack scale, secure transmission routes and assemble operating data centers. Every open problem on the research list — solid-state transformer reliability, DC fault protection, chip-internal cooling — needs field proof, not just simulation. None of that is overnight work.

Who wins, who loses

If Korea moves too fast on the announcement and too slowly on the grid, the losers are obvious: developers who break ground without power, tenants who lease space that cannot be energized, and regional grids that absorb demand spikes without receiving upgrades in return. The winners will be the firms that control the integrator role — the ones designing the rack, the power topology and the cooling loop as a single system — regardless of where the physical data center sits.

If Korea invests deliberately — treating power, cooling and integration as a coordinated program rather than three parallel workstreams — the upside is real. The country still has strong semiconductor and display supply chains, deep engineering talent and the capital to fund long lead-time projects. The question is whether the industrial strategy matches the scale of the problem.

The hidden story

Every headline about AI capacity targets the wrong variable. More GPUs mean nothing without more watts, more cooling and a system design that holds them together. Korea’s 8.4GW pledge is a signal of intent. The real test begins when the grid operators, the cooling engineers and the system integrators show up with something that works at scale.

Until then, the number on the page is less a plan than a wish.