South Korea broke ground on a national AI computing center in Haenam this week, a roughly $1.7 billion project led by Samsung SDS that aims to have 15,000 GPUs running by 2028, according to Aju Press.
The project had a hard birth. Two earlier tenders failed, and the government restarted the effort this time in partnership with the private sector, with Samsung SDS — the IT services arm of the Samsung group — taking the lead. The new structure is a concession to reality: the state tried to build the center itself, could not get the deals done, and now leans on a company with the engineering muscle to execute.
The center is the most concrete sign yet of a global trend: governments building AI computing capacity the way they once built roads and ports. Compute has become strategic infrastructure, and countries without enough of it fear being locked out of the AI boom. The race spans the United States, Europe, Japan and the Gulf states, each pouring public money into data centers, chips and the grid to power them.
For Korea, the gap is uncomfortable. The country dominates the memory side of AI — SK Hynix and Samsung make the high-bandwidth memory that Nvidia’s chips need — and it has been a winner of the AI hardware boom. But it lacks large-scale domestic computing capacity, and most Korean AI companies rent computing from American clouds. Researchers have complained about cost, access and the fact that Korea’s most sensitive data flows through foreign servers.
Fifteen thousand GPUs sounds large but is modest next to what American hyperscalers operate. Some data-center projects in the United States have deployed more than 100,000 GPUs in a single site. Korea’s center is better understood as a foundation: enough capacity to train models, run research and seed a domestic AI industry that analysts said will grow around it.
The project is also a matter of sovereignty, in the language of Korean policymakers. Officials have said the country cannot rely on foreign clouds for technology that underpins its economy and security, and similar arguments have driven projects from Tokyo to Brussels. The European Union’s EuroHPC program and Japan’s subsidies for domestic AI development have poured public money into computing, and Korea is now joining that club.
The choice of Samsung SDS signals the government’s preference for big domestic players over foreign vendors, a pattern in Korean infrastructure projects. Samsung SDS brings experience running data centers and enterprise clouds, and its involvement gives the project a private-sector owner with a reputation to protect. The company will build the facility, operate it and presumably sell capacity to Korean companies and researchers.
Construction is under way in Haenam, a county on Korea’s southwestern coast chosen for its land and power access. The center is scheduled to come online in stages, with the full 15,000-GPU deployment targeted before 2028. The timeline is aggressive, and the project’s history suggests the schedule could slip.
The project faces the same constraints as every AI build-out: chips and power. Korea’s electricity grid has limited headroom, and the government has had to balance AI ambitions with concerns about stability. The center’s GPU purchases will compete with everyone else’s in a market where supply is tight, and delays in chip deliveries would push the whole schedule back.
Whether government-built compute is a good use of money is a live debate. Proponents say AI capacity is a public good, and that leaving it entirely to private clouds concentrates power in a few American companies. Critics say the private market is building capacity faster than any government can, and that state projects risk obsolescence by the time they open. Korea’s center is small enough to avoid the worst of that critique and strategic enough to matter, analysts said.
The project also creates domestic demand for the components Korean companies make, and it gives Korean AI startups a place to train models without sending data abroad. The country’s large language model builders, search and messaging companies and research institutes are all potential customers, and the center’s success will be measured by how much of that demand it captures.
The center is the centerpiece of a broader push by Seoul to reduce the country’s dependence on foreign cloud providers. Korean AI companies have long trained their largest models on infrastructure run by American hyperscalers, a dependency that carries both currency risk and a data-governance question mark. The government has answered with a series of incentives: subsidies for domestic data centers, relaxed permitting for new power connections and a fund seeded with public money to co-invest in private compute projects. The Haenam facility is the first project to combine all three tools, and officials have said more sites will follow if the model works.
The economics of sovereign compute are not straightforward. A 15,000-GPU cluster represents a fraction of what the largest American labs operate, and maintaining utilization requires a steady stream of paying customers rather than a one-time construction contract. Korea’s approach has been to line up anchor tenants early, including the country’s largest search and messaging companies, which are racing to catch up in large language models against rivals backed by far larger compute budgets. If the center fills its capacity, it becomes a template for other mid-sized economies weighing the same choice: build national AI infrastructure, or keep renting it from abroad.
The groundbreaking in Haenam is a small ceremony with a large meaning. Korea, which built its economy on making chips, is now betting on the computing that consumes them. The next question is whether 15,000 GPUs will be enough — or whether, like most infrastructure bets, the number will grow.
