Samsung's $648 billion bet on South Korea buys the buildings. It doesn't buy the megawatts.

The largest corporate investment ever pledged by a South Korean company is a number with a lot of zeros and one quiet omission. Samsung has committed 1,000 trillion won — about $648 billion — over the next decade, set out this week to the country's presidential office, to build chip fabrication plants, AI data centers, batteries and displays across South Korea. It is a serious figure, and most of the coverage will stop at it. The more useful number is the one Samsung did not put in the announcement, because it is the one that decides whether any of this works: the cost of a kilowatt-hour delivered to a fab in Yongin, and whether the wire to carry it arrives before the building does.

You can see the constraint inside the plan itself, if you read why the money is going where it is going. A large share of the new capacity is headed for the country's southwest rather than the capital region, and the stated reason, per the reporting, is blunt: there is no more room, power or water around Seoul to expand into. That is not a footnote. That is a company with $648 billion to spend telling you that the binding limit on a semiconductor build is no longer the cleanroom or the lithography tool. It is the megawatt and the cubic metre of cooling water. The capital filled up on electricity and water before it filled up on land.

What the pledge buys

Take the plan at face value first. Of the 1,000 trillion won, reporting points to roughly 300 trillion won aimed at chip factories in the southwest, a tranche on the order of 60 trillion won toward fabs in the Yongin cluster south of Seoul, and more than 350 trillion won for AI infrastructure, data centers included. The rest spreads across batteries and displays. The government is framing it alongside parallel commitments as a set of mega-projects in semiconductors, AI data centers and robotics, presented to President Lee Jae Myung's office as a way to push advanced manufacturing out beyond the capital.

Samsung is not moving alone. SK Hynix, the other half of Korea's memory duopoly, has just filed to raise as much as $29.65 billion through a US listing, with the proceeds earmarked for advanced-memory fabs and the extreme-ultraviolet tools that print them — about 31 trillion won for the first Yongin fab alone. The presidential office's policy chief, Kim Yong-beom, said this week that both companies' cluster plans are in final discussions, and that the timelines are being pulled forward because AI demand is growing faster than anyone planned. SK Hynix is reportedly advancing the completion of one Yongin plant from 2044 to 2034. A decade has been deleted from the schedule. The schedule for the buildings, that is. Not the schedule for the power.

The number that isn't in the press release

Here is the arithmetic the won total walks past. South Korea's total data-center power capacity sat at roughly 1,960 megawatts in 2025 and is projected to more than triple, to about 6,320 megawatts, by 2030. National electricity demand is marching toward 1,000 terawatt-hours — close to the entire annual consumption of Japan, a country with more than twice the people. A fab is among the most power-hungry pieces of industrial equipment ever built; an AI data center is, in plain terms, a contract for electricity wearing a building. You can finance the building in an afternoon with a pledge like this one. The electricity has to be generated, every hour, for twenty years, and that is a different kind of promise.

It is a harder promise in Korea than almost anywhere, for a reason of geography that no investment sum can edit. The Korean grid is an island. The peninsula has no transmission interconnection to any neighbour — nothing flows in from China or Japan the way power crosses borders in Europe — so every kilowatt-hour consumed in the country has to be generated inside it. And it is generated, overwhelmingly, from fuel the country does not have: as of 2023, around 24 percent of Korea's energy came from coal, about 37 percent from oil and roughly 19 percent from gas, nearly all of it imported. Renewables remain a small slice, with the country lagging years behind peers on its share. Put those two facts together — an electrical island, burning imported fuel — and you get a structurally expensive kilowatt-hour with no cheap neighbour to borrow from when demand spikes. That cost, not the headline won figure, is the real input price on every chip and every token these sites will produce.

An AI data center is a contract for electricity wearing a building. You can finance the building in an afternoon. The electricity is a twenty-year promise.

The wire, not the building

Even if the power can be generated, it has to be delivered, and delivery is where Korea's plan is already slipping. As of late 2025, more than half of the country's planned transmission and substation projects were running behind schedule. That is the unglamorous machinery — the lines, the towers, the substations — that connects a power station to a fab, and it is the part that cannot be compressed by writing a larger cheque. A substation takes years and a right-of-way and the consent of the people who live along the route, and across the world those people increasingly say no. Build time on a data center is now shorter than connect time to the grid, which means the constraint has quietly moved from the concrete to the copper.

The policy response so far illustrates the gap rather than closing it. Korea passed a special law designating AI data centers as national strategic facilities, which speeds permits and signals intent. But lawmakers stripped out the provision that would have let these sites sign direct power-purchase agreements for their own gas-fired supply — the one clause in the bill that actually addressed where the electricity comes from. What survived was the part that helps you build faster. What was removed was the part that helps you power what you built. That is the whole problem in miniature: the easy half of the policy passed, and the half that costs something got deferred.

The one line item that helps

There is a piece of Samsung's plan that genuinely speaks to the constraint, and it is the least glamorous one: batteries. Grid-scale storage does not make a single new kilowatt-hour, and it is important to be precise about that — it cannot fix a generation deficit. What it does is move energy across hours, soaking up cheap supply and releasing it into the expensive peaks, which is exactly the problem a data-center load with a flat, relentless draw creates for a grid that strains at certain hours. At the system level, storage is now landing in the rough neighbourhood of a couple of hundred dollars per kilowatt-hour of capacity, cheap enough that it has crossed from demonstration into deployment. It is the boring, useful part of a flashy announcement. It will not power a fab. It will help a strained grid carry one without falling over at 7 p.m., and on an island grid that is worth more than it sounds.

Who actually pays

South Korea is a wealthy country that can summon $648 billion of private capital and a cooperative government and still run straight into a wall made of megawatts and transmission lines. That is the instructive part, and it travels. The same arithmetic — what does a delivered kilowatt-hour cost, and can the wire reach the site — is the one that decides whether a data center ever gets built in Lagos, or Jakarta, or anywhere the AI map currently leaves blank. The difference is that those places cannot paper over a power deficit with a pledge this size; the constraint there is simply more naked. Korea's plan is a preview of the bottleneck the whole industry is walking into, run by one of the few players rich enough to hit it early and in public.

So treat the $648 billion as what it is: a precise measure of how badly Samsung wants the fabs and the data centers, and almost no measure at all of whether the country can run them. The figure that will decide that is being set somewhere far less photogenic than a presidential signing — in the queue for grid connections, in the price of imported gas, in the substations that are already late. Watch the cost per kilowatt-hour and the date the wire arrives. The won was never the hard part. It rarely is.