Category: Policy Date: September 3, 2026 Sources: SemiAnalysis, DatacenterDynamics, Light Reading, The Korea Times, AI Weekly
Key Takeaway: Korea has committed roughly $919 billion across chips, robotics and compute, and is buying gigawatts and permitting relief rather than a single flagship campus, but its sovereign stack still runs on American accelerators.
---
When President Lee Jae-myung addressed the nation on June 29, he did not unveil a campus. He unveiled a power budget. South Korea, he said, would build 18.4 gigawatts of AI data center capacity by 2035, a figure that reads less like a corporate press release than a line item in a national grid plan. Against the 1.9GW the country had operating in December 2024, per the consultancy Savills, that is roughly a sevenfold multiplication of national capacity.
“We must secure the core elements of AI faster than any other country,” Lee said in the televised address. “Semiconductors, physical AI, and AI data centers are the triple axis for our great leap forward.”
The number attached to the full program is 1.43 quadrillion won, about $919 billion. That headline figure spans three mega projects: chip fabrication, physical AI and robotics, and the data centers themselves. The data center pillar alone is budgeted at over 1,000 trillion won, roughly $650 billion, spread across a decade. The semiconductor leg is nearly as large. Samsung and SK Hynix are slated to spend 800 trillion won ($513 billion) on four memory fabs in the Jeonnam-Gwangju Integrated Special City, plus another 81 trillion won ($52 billion) on HBM packaging plants in the Chungcheong region.
Phase one is a three-way split
The first tranche targets 8.4GW by 2029 at a combined cost of about 550 trillion won ($352 billion), divided among three domestic conglomerates. SK Group takes 5GW, GS Group 2.4GW, and Naver 1GW. Ground is scheduled to break in the first half of 2028, with phased operation from 2029. SK Group alone carries the remaining 10GW in phase two, running 2030 to 2035, though the government attached a conditional clause: that expansion depends on the results of phase one and on market demand.
Much of that repackages projects already in motion: Naver's 1GW Sejong campus, announced in November 2025, SK's 1GW build in Ulsan, and GS plans for two 1.2GW facilities in Donghae and Dangjin. SemiAnalysis has identified three active phase-one sites totaling 4.4GW.
What is genuinely new is the state scaffolding: discounted electricity rates, streamlined land and water rights, and direct support on permitting and construction. Seoul is also steering capacity out of the capital region, where Korean data centers have historically clustered, toward Chungcheong, Ulsan, Donghae and Sejong. The government calls this regional balance. It is at least as much about grid congestion.
SK Group Chairman Chey Tae-won has been the most direct about what the buildings are for, telling Maeil Business Newspaper that in the AI era, data centers are factories that produce intelligence. Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon has pushed the metaphor further, calling AI data centers token factories and describing the goal as a token economy in which their output feeds physical AI and autonomous agents.
“The AI market is rapidly shifting from learning to inference, and the inference market is an open ecosystem,” Bae said.
The tournament
Running in parallel is a stranger piece of policy: a knockout competition to pick the country's foundation model. The Independent AI Foundation Model project began with 15 consortiums, narrowed to five in August 2025 (Naver Cloud, LG AI Research, SK Telecom, NC AI and Upstage), and eliminates teams every six months, reallocating their subsidies to the survivors. The government budget is roughly $350 million, including $45 million spent licensing Korean books, news archives and broadcast video.
Round one produced an upset. NC AI finished last, and Naver was disqualified for using vision and audio encoders drawn from Alibaba's Qwen family. A supplemental round added Motif Technologies, a sub-30-person startup that had raised $17 million and had access to 768 B200s. Motif then trained the strongest model in the field, ahead of every American open-weight model on the Artificial Analysis Intelligence Index according to SemiAnalysis, at a total training compute cost of roughly $15 million.
On August 18, Motif was eliminated. Scoring weighted benchmarks at 40 percent, expert review at 35 percent and user testing at 25 percent. Motif won the first category and finished last in the other two. LG AI Research, SK Telecom and Upstage advanced. The final two teams will be named in early 2027.
Why This Matters
Three things separate this from the sovereign-AI announcements that have become routine. First, Korea is buying power, not just chips. Most national programs procure GPUs; this one procures gigawatts, land rights and tariff treatment, which is the constraint that actually binds. Second, sequencing: the compute arrives before the model is chosen, so whichever team wins in 2027 inherits infrastructure instead of waiting for it.
Third is the gap the plan cannot close with money. Korea can finance land, power, cooling and buildings. It cannot manufacture a globally competitive accelerator, and its domestic NPU designers, Rebellions and FuriosaAI, remain behind Nvidia and AMD on deployment scale and software maturity. That dependency is visible in the deal flow: SK Group has committed to 2GW of Nvidia Rubin systems inside its 5GW allocation, and Naver has signed for another 200MW. The sovereign compute stack still runs on American silicon.
There is a quieter risk in the tournament design. A program built to produce a national champion just eliminated the team that produced the best model, on criteria participants describe as opaque. SemiAnalysis notes Motif may now need to leave Korea to raise the capital and compute to continue, an unusually literal case of industrial policy defeating its own objective.
What to watch
Three dates matter. Early 2027, when the government names its final two model teams and the reallocated subsidies land. The first half of 2028, when phase-one sites break ground and the schedule stops being a slide. And the phase-two decision around 2029, where SK's additional 10GW is explicitly conditioned on demand, the clause that quietly converts a headline pledge into an option.
Korea has committed the capital and cleared the permitting, which is more than most countries chasing sovereign AI have managed. What it has not shown is that 18.4GW of Korean inference has 18.4GW of buyers, or that the tokens will be worth more than the electricity spent making them.
---
Sources: - SemiAnalysis - DatacenterDynamics - Light Reading - The Korea Times - AI Weekly
“Semiconductors, physical AI, and AI data centers are the triple axis for our great leap forward.”— Lee Jae-myung, President, Republic of Korea