Five Years of Power Bills Up Front? Samsung and SK Hynix Just Told KEPCO No

On Monday, September 14, 2026, one of the more unusual financing requests in the semiconductor industry’s history ended in a quiet rejection. Samsung Electronics and SK Hynix — the two largest industrial electricity customers in South Korea — formally declined a proposal from Korea Electric Power Corp. (KEPCO) to prepay a combined 25 trillion won (about $18.4 billion) of electricity bills, roughly five years’ worth, to help the state utility finance grid upgrades for the country’s next-generation chip clusters, industry sources and internal documents reviewed by Korean media confirmed.

The answer, delivered after weeks of internal review, was no. And the reasoning behind that no says a great deal about where the AI boom now stands: even the companies selling the most picks and shovels of the AI gold rush are no longer willing to bet their balance sheets that the boom lasts five more years.

What KEPCO asked for

The proposal, first reported by Bloomberg in early September, was straightforward in structure if startling in scale. KEPCO asked Samsung Electronics to prepay about 20 trillion won ($14.7 billion) and SK Hynix about 5 trillion won ($3.7 billion) — figures calculated from each company’s 2025 electricity bills and their expected consumption over the following half-decade. In exchange, the utility offered an interest rate above the yield on two-year Korean government bonds, with accrued interest credited back against the chipmakers’ electricity bills every six months.

The funds were earmarked primarily for national grid investment, with a particular focus on the transmission and substations needed to serve the two mega-scale semiconductor clusters now under development: Yongin, south of Seoul, where Samsung is leading construction, and the Honam region in the country’s southwest, the site of the planned $500 billion chip hub announced with presidential backing in June 2026. Both clusters are the centerpiece of Korea’s attempt to defend its position in advanced memory — the HBM stacks that sit beside every flagship AI accelerator Nvidia ships — against Chinese and American expansion.

For KEPCO, the appeal of the arrangement was financial survival. The utility’s total liabilities stood at 210.7 trillion won as of the end of June, and its daily interest expenses run to roughly 11.5 billion won — about $8.5 million every day, before a single watt is generated. A temporary regulatory measure that currently allows KEPCO to issue bonds worth up to five times its combined capital and reserves expires at the end of 2027. Without it, the utility’s ability to roll over its debt narrows dramatically. A 25 trillion won prepayment from its richest customers would have bought years of breathing room and reduced dependence on new bond issuance.

Why the chipmakers said no

The chipmakers’ rejection, reported by The Investor, the Korea Herald, Chosun Ilbo and Seoul Economic Daily, rested on two arguments.

The first is the semiconductor cycle. The industry has enjoyed a strong upswing since the second half of 2025, with AI memory demand in particular soaking up supply. But a five-year prepayment is a five-year forecast, and internal reviews at both companies reportedly concluded that committing that much cash upfront was impossible to justify amid uncertainty over the longer-term chip cycle. The very executives being courted to underwrite Korea’s grid expansion are the ones watching Washington’s AI “pacing” debate — where lawmakers and even some lab CEOs have spent this week arguing for slowing frontier AI development — and Asian AI stocks selling off on the news. If AI capex cools, electricity demand forecasts cool with it, and a prepaid bill becomes an interest-free loan to a money-losing utility.

The second argument is financial flexibility. Samsung and SK Hynix are simultaneously executing some of the largest capital expenditure programs in their histories — SK Hynix’s board alone recently approved 54.3 trillion won ($38 billion) for two new memory fabs. Every trillion won parked at KEPCO is a trillion won not available for fab construction, HBM R&D, or shareholder returns. The interest sweetener, pegged just above two-year government bond yields, was never going to compensate for that opportunity cost.

The power wall behind the request

To understand why KEPCO was desperate enough to ask, look at the demand curve bearing down on the Korean grid. According to the government’s own tenth long-term electricity plan, peak power demand could reach 158.4 to 165 gigawatts by 2040, with the Yongin semiconductor cluster alone expected to need 14.7 GW by 2041. Wood Mackenzie warned in July that the KRW 4,700 trillion AI and semiconductor mega-project faces a 2.3 GW power shortfall under the current trajectory. Reuters reported in July that power demand from the four planned fabs in the Honam region could equal 70 to 80 percent of Seoul’s current annual electricity consumption.

This is not an abstract future problem. Korea has already begun filtering “fake demand” out of grid connection applications — semiconductor projects were mostly validated due to strong policy backing, but data center applications have been heavily discounted amid widespread exaggeration of power needs. The country has even softened its decarbonization stance, approving a liquefied natural gas cogeneration plant for the southwestern chip hub to keep the project’s timeline intact.

The parallel to the United States is hard to miss. American utilities and their regulators spent 2025 and 2026 wrestling with hyperscalers over who pays for grid expansion — with Microsoft’s Brad Smith publicly arguing Big Tech must “pay our way” for data center infrastructure. What KEPCO attempted is the most aggressive version of that logic attempted anywhere: not special tariffs or connection fees, but half a decade of bills collected in advance. Korea’s chipmakers just demonstrated where the private sector’s limit lies.

What happens next

The immediate consequence is that KEPCO must return to the bond market — and eventually to the government — for its grid financing. Heavier bond issuance is now the base case, as Seoul Economic Daily noted, and with the 2027 regulatory expiry approaching, pressure will grow on the Ministry of Finance and Economy and the National Assembly to either recapitalize the utility, extend the special bond-issuance allowance, or restructure its debt. Each option carries political cost in a legislature where energy policy is contested.

The longer-term consequence is a renegotiation of the social contract between the Korean state, its utility, and its chaebol chip champions. The June 2026 summit at which President Lee Jae Myung, Samsung Chairman Lee Jae-yong, and SK Group Chairman Chey Tae-won joined hands to announce the $500 billion clusters was a photo of shared commitment. The prepayment rejection is a reminder that commitment has limits: the chipmakers will build the fabs, but they will not finance the grid, at least not on these terms. Expect the next proposal to be smaller, shorter-dated, or structured as equity in grid subsidiaries rather than prepaid bills — or for Seoul to simply absorb the cost into the national balance sheet, as Washington has begun doing through federal-lands data center leases and Energy Department programs.

For the AI industry globally, the episode is a data point in the great repricing of infrastructure risk. The assumption throughout the boom has been that compute demand is infinite and everyone in the supply chain should lever up accordingly. This week, the two most important memory makers on earth — whose HBM is in every serious AI training cluster — looked at a five-year demand forecast and declined to underwrite it. That is not a vote against AI. It is a vote for uncertainty. And in the power-starved 2020s, uncertainty about who funds the grid is the one commodity nobody can afford to stockpile.