KEPCO Signs an MOU. The Substation Doesn't Care.

KEPCO Signs an MOU. The Substation Doesn't Care.

One figure in the Korea Electric Power Corporation's April briefing tells you almost everything you need about the shape of Korea's AI power problem: five hundred and ninety-eight of eight hundred and fifty-three new industrial power-connection applications submitted through March 2026 landed in the capital region. Seventy percent of a country's declared future load, aimed at the one slice of grid that already sits at about 0.61 energy self-sufficiency and draws electricity from other regions to hold voltage.

I have been reading KEPCO's own statements this quarter with what I hope is a fair reader's attention. The company is doing two things at once, and it is worth being careful about which of them is the story.

From the utility's own communications

On 27 May 2026, KEPCO said it would deploy an AI-based grid-management system its communications expect to reduce annual electricity procurement costs by roughly 110 billion won, or about $73 million. CEO Kim Dong-cheol described the change as an overhaul of demand-forecasting and an optimization of static synchronous compensators, the STATCOMs that hold reactive power in balance on long transmission runs. One month later, KEPCO signed a memorandum of understanding with OpenAI at Seoul National University. The stated aim is joint policy and technology exchange around AI in the energy sector; a Korean utility gets access to a leading model shop, OpenAI gets a large-scale operational partner with genuinely deep telemetry.

Neither announcement is empty. Both are the kind of upgrade an incumbent grid operator should have on its roadmap. A ₩110-billion procurement saving is a real number, even for a utility with a balance sheet this stressed. Better demand forecasting and better reactive-power management lower the cost of running the network. If I were writing an operating plan for KEPCO I would want more of exactly this.

What I would resist is the reading these announcements will attract from anyone outside the control room: that Korea's AI-and-grid problem has an AI-shaped answer.

What software cannot do for a transformer

The load KEPCO now has to serve has been publicly quantified. On 29 June 2026, Korea's Ministry of Science and ICT unveiled a plan for 18.4 GW of AI data-centre capacity by 2035, with 8.4 GW targeted by 2029, backed by more than 1,000 trillion won of committed investment across SK Group, GS Group and Naver. Samsung and SK hynix separately committed roughly $880 billion across 2026–2036 to memory-fab and AI infrastructure in the Gwangju–South Jeolla corridor. Government and industry estimates put the additional generation requirement for the southwestern semiconductor cluster and AI data centres alone at roughly 24.7 GW beyond the last power-plan baseline.

Twenty-four gigawatts of new generation is a nuclear-programme-sized ask. It is not a demand-forecasting-model-sized ask. The IEA's 2025 Korea Energy Policy Review, published before this news cycle and worth reading against every announcement since, was clear that the country's stable-supply challenge for semiconductors and data centres was structural: a question of where plants and lines and substations get built, and how quickly, in a country where about 75 percent of existing data-centre capacity is concentrated in Greater Seoul.

Now put the second inconvenient number against that. KEPCO's debt has surpassed 202 trillion won — roughly $145 billion — and directly constrains how quickly it can fund grid upgrades. More than half of the utility's planned transmission-expansion projects are late. Several of the ultra-high-voltage runs from coastal generation into the capital region have slipped, in Korean reporting, from delayed to indefinitely delayed on the strength of local opposition and cost. A high-voltage line takes three to seven years to build. Transformer lead times run two to five. A hyperscale data hall goes up in eighteen to twenty-four months, and the ones being planned now assume the wires will be there when the racks turn on.

They will not be. Not for all of them. The bottleneck is metal, permitting and municipal politics; it lives outside the model weights.

What the Special Act quietly concedes

The government's answer, delivered in stages, has been to give up on KEPCO as the sole channel and route around it wherever the physics resists. In May 2026 the National Assembly passed the Special Act on the Promotion of the AI Data Center Industry. Qualifying AI data centres are now designated national strategic facilities. Permits move through the National AI Strategy Committee on an integrated track. And this is the operative clause and the one most worth reading closely: certain non-metropolitan projects are exempt from grid-impact assessments below a specified scale.

Beneath that headline there is a quieter market development. AI data-centre operators in Korea have started building the LNG-generation supply chain themselves, through direct power-purchase agreements. The Seoul Economic Daily reported in April that the average unit price KEPCO paid to LNG generators last year was 158.4 won per kWh, against a general industrial tariff of 168.9 won: a 6.2 percent gap that flips the economics for a large operator who can buy generation direct. Operators are doing exactly that where they can. Steel and petrochemical firms asked for the same PPA carve-outs a decade ago and were turned down on grid-operation principles. The AI industry has just gotten what heavy industry could not, which tells you both what has changed in Seoul's policy priorities and what the utility is now being asked to absorb.

That is the industry stating in cash the same thing the transmission delays state in months. The traditional model, in which KEPCO plans supply and delivers it, cannot be counted on to meet the new load in the new places on the new timeline. Something has to give, and what the government has decided will give is the monopoly assumption.

The nuclear pivot, and why I would not lean on it

There is a counter to this reading. On 8 July 2026, KED Global reported that Korea's presidential policy chief signalled a pro-nuclear shift to power the AI expansion. Today's Korea Times piece puts nuclear back at the centre of the conversation for the first time in a decade, quoting a minister who says new reactors are on the table if the chip cluster keeps growing.

I want to give this its due. Baseload matters. A 24/7 chip fab or training cluster is a poor match for the intermittency profile of a Korean renewables build that generated only about 10 percent of national electricity in 2024. Nuclear-plus-firmed-renewables is a genuinely defensible answer to that load shape. If you were designing a Korean 2035 grid on a blank sheet, you would put nuclear on it.

But nuclear is the right long-run answer to the wrong short-run problem. A new Korean reactor from siting to grid connection is a ten-to-fifteen-year project, and the political calendar keeps rewriting the last few years of that. Between now and 2032 the deficit will be closed with gas, with corporate PPAs, with permit-exemption sites and whatever regional generation the Special Act can conjure — not with reactors that have not yet been sited.

An old control room, remembered

In 2008 I spent a week shadowing a control-room supervisor at a European transmission operator during a summer heatwave. The SCADA rebuild had just gone live: screens crisp for the first time in a decade, alarms grouped sensibly, colour-coded severity that finally made sense at three in the morning. He said the same thing three times in five days, in a voice he must have used with a great many visiting consultants: "the screens are prettier and I can see the trouble sooner. I still cannot make copper appear where copper isn't." He did not have anything close to what KEPCO now has under the OpenAI banner. He would have taken it. He would also have told you that the plant that never got built is still not there, and no dashboard hides an absent transformer for long.

That is roughly the shape of the KEPCO situation. The visibility gets better. The materials do not.

Where this leaves the substation

KEPCO's May and June announcements will be quoted a great deal in the months to come, by policymakers who need to show motion, by vendors selling into the energy sector, by AI companies looking for large-utility partners. Some of that quotation will be accurate and some of it will be inflation. The specific claim that AI-driven operations reduce KEPCO's procurement cost by about $73 million a year is credible and worth tracking. The general claim that AI is the answer to Korea's grid-and-AI-load problem is not, because the problem sits outside the software perimeter and always did.

The substation that is not being built along the line from Yeongheung into the capital cannot be trained. The permit a small municipality has not yet granted cannot be prompted. Between the AI that runs the grid and the AI that runs on the grid, the second one has moved much faster, and Korea's next four years will be shaped by whether the first can meet it on the ground with wire and steel, not model weights.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan, an Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.