Summary: South Korea plans to invest roughly $648.2 billion in AI data centers and build 18.4GW of new capacity, according to BigGo Finance. National ambitions like this are eventually settled in watts, and for the operators who run the buildings, a gigawatt target is a massive load to source and manage.

South Korea has put a number on the AI race, and it is enormous. The government and its industrial partners plan to invest about $648.2 billion in AI data centers, building 18.4 gigawatts of new capacity. A figure that large is an operational problem. Every gigawatt in that plan eventually arrives on somebody’s floor as AI data center power, and as the heat that comes with it.

From a national plan to a live electrical load

18 gigawatts is not insignificant. It is closer to the output of a fleet of large power stations, committed to a single use case: feeding accelerated computing. The International Energy Agency has been tracking this shift for a while, and its analysis of data centre electricity demand makes the trajectory apparent. AI is now a primary driver of national power planning.

The headline however describes the power needed but it does not describe the thing an operator actually manages, which is thousands of racks. Each one drawing current and rejecting heat, wired through breakers and busways that were sized months or years before the GPUs showed up. The gap between the announcement and the running floor is where the real work lives, and it gets wider the faster capacity is built.

Why gigawatts are the wrong unit on the floor

The gigawatt is the planner’s unit. The operator’s units are the rack, the circuit, and the inlet temperature. A capacity figure tells you nothing about where AI data center power concentrates. A handful of dense training racks can draw more than a whole row of legacy equipment, and the average across the hall hides exactly the racks most likely to trip a breaker or overrun a cooling zone.

This is where stranded capacity creeps in.

Stranded capacity is the gap between what you provisioned and what you can actually use.

When power and cooling are provisioned from nameplate ratings and design assumptions rather than from measured draw, operators leave headroom everywhere to stay safe. That headroom is capacity nobody is selling or computing on. In a market racing to stand up 18.4GW, losing a slice of every megawatt to guesswork is an expensive way to feel comfortable.

The mistakes operators make as AI compute capacity lands

The errors are consistent, and they are not about competence. They are about visibility.

The first is provisioning from the nameplate. A server’s rating is a worst-case ceiling, not its real draw, so a rack budgeted by nameplate is almost always budgeted for load it will never pull. The second is treating power and cooling as separate line items owned by separate teams, when they are two views of the same watt. The third is discovering the actual distribution of load only when something fails, a tripped circuit or a hot aisle that climbs out of the ASHRAE envelope during a heat event.

None of these show up on the capital plan. All of them show up in the operations budget, and they get worse as density climbs from the 15kW racks the industry grew up on toward the 30kW and beyond that AI training pushes, with the densest deployments well past that. At those densities, running a facility on design-time assumptions instead of live readings is managing by worst case, and paying for the privilege.

Power and cooling are the same constraint

Every watt a server draws leaves the rack as heat. There is no exception. A hall provisioned for a gigawatt of IT load is, by definition, a hall that has to move close to a gigawatt of heat back out, and the denser the racks, the less margin there is to get that wrong.

That is why the useful measurement is never power alone. It is power in and heat out, read at the same place and the same time. The temperature difference between what a rack takes in and what it exhausts, the inlet-to-outlet delta, tells you whether cooling is actually keeping up with the electrical load or just appearing to. Watch one number without the other and you are flying half blind. Watch both, per rack, and the gigawatt target on the national plan finally becomes something you can operate against.

How AKCP turns AI data center power into a number you can see

So how do you operate against a plan this size? You make the load visible where it lives, circuit by circuit and rack by rack. Most floors today are working from design assumptions instead of measurements. That is a monitoring problem, and it is the one AKCP is built to solve.

Start with the power itself. A contactless current meter clamps around the conductors already feeding a rack and reads the draw with no rewire, no PDU swap, and no maintenance window. That matters when the racks are dense and already in production, because you were never going to schedule an outage just to learn how much they pull. Power Train maps the distribution path from the mainline down to the individual outlet, with per-outlet current and kWh, so a high-density AI rack stops being a guess and becomes a measured load you can trend. Feed those readings up and AKCP computes real-time PUE as a first-class virtual sensor, graphed and alertable like any other reading, so the efficiency figure you report is a live measurement, not a number you reconstruct once a quarter.

Then the heat, because it is the same watt on the way out. The Thermal Map sensor reports per-rack inlet and outlet temperatures with ASHRAE scaling and color heat maps, so hot spots and stratification surface as they form instead of after a thermal alarm. sensorCFD runs a CFD simulation fed by live sensor data and returns an AI-assisted thermal report you can act on, so you are optimizing against how the room behaves right now, not how it was drawn. And because cooling units, generators, UPS, and busway meters speak Modbus, SNMP, or MQTT, AKCP ingests them as virtual sensors and lands power and thermal telemetry together in Quicklime DCIM, on one dashboard, for the whole site.

For the colocation operators who will absorb much of this new capacity, the platform is multi-tenant and stays on-prem and self-hosted, so each tenant sees its own power and thermal data while the site keeps control of the system. And where new AI racks move to liquid cooling, spot and rope leak detection plus flow sensing raise an alarm on a drip or a loss of circulation before it reaches the hardware.

We don’t just tell you where you have a problem, we tell you how to fix it. Whether the load on your floor is one rack or a slice of an 18.4 gigawatt national plan, the number that matters is the one you can measure.

What it means for operators planning AI capacity

South Korea’s headline will be matched or beaten by other countries, and the pattern underneath it is the same everywhere. The money is committed in gigawatts, the deadlines are aggressive, and the load lands on operators who are expected to run it at high utilization without tripping anything. That combination punishes guesswork.

The practical move is not to argue with the number. It is to make sure that when your share of it arrives, you are measuring real draw instead of trusting nameplates, watching power and cooling as one system instead of two budgets, and reading it per rack so the dense outliers cannot hide inside an average. AI data center power is going to keep scaling faster than most floors were designed for. The operators who stay ahead of it will be the ones who turned a gigawatt-sized plan into a live, per-rack measurement they can defend.