1. What was actually announced

On July 24, 2026, NAVER, NVIDIA and Brookfield announced a proposed expansion of Korea’s sovereign AI infrastructure from an initial 55MW to 200MW by 2028. The buildout is planned for NAVER’s GAK Sejong hyperscale data center using NVIDIA DSX, with compute intended for Korean and U.S. companies developing models, agents and services.

The milestone extends NAVER’s June roadmap: 55MW operating in the first half of 2027, overseas capacity reaching 100MW within that year, and 200MW in 2028. NAVER’s longer-term 1GW path is an ambition without a published close date. The 200MW stage is the more concrete target, but it is still forward-looking.

2. Financing decoded: an announcement is not cash deployed

NAVER’s July 25 release describes a $10 billion project funded by up to $9 billion from Brookfield, $1 billion from NVIDIA and any remaining amount from NAVER. The global NVIDIA release is more cautious: Brookfield entered a nonbinding term sheet to fund up to $9 billion, while NVIDIA’s planned investment is subject to customary closing conditions and NAVER finalizing at least $9 billion of committed financing separate from NVIDIA’s investment.

That distinction is material. “Up to” is not the same as fully committed or funded, and a nonbinding term sheet is not a closed financing agreement. The reviewed sources do not disclose draw schedules, cost of capital, asset ownership or contracted demand. We therefore classify the $9 billion and $1 billion as announced, planned or conditional—not completed investment.

  • Confirmed: the announcement, site, 200MW target and planned technical architecture.
  • Conditional: NVIDIA’s close and NAVER securing the stated committed financing.
  • Nonbinding in the global release: Brookfield’s term sheet for up to $9 billion.
  • Undisclosed: actual cash draws, debt/equity mix, financing cost and customer contracts.

3. Why $10 billion is plausible—without treating it as a final cost

Brookfield’s broad infrastructure outlook says a hyperscale data center can cost more than $10 million per megawatt for the facility and more than $30 million per megawatt for compute, depending on semiconductors. Applying those broad thresholds to 200MW points above $8 billion. That explains the order of magnitude; it is not a NAVER project estimate and may exclude land, energy works, financing and phased procurement.

Megawatts are not a fixed GPU count. NAVER’s Korean release describes roughly 100,000 NVIDIA GPUs, but the realized number depends on platform mix, rack density, cooling, networking and facility overhead. Commercially, tokens per megawatt, utilization and energy cost matter more than accelerator count alone.

4. The factory is a cloud product, not only a building

The planned stack combines Blackwell and Vera Rubin systems with DSX and multi-tenant operations. That turns the project from internal infrastructure into sellable compute capacity. The meaningful operating metrics will be time to connect, availability, utilization, inference cost and durable customer contracts.

Models and applications are part of the thesis. NAVER says it is advancing HyperCLOVA X using Nemotron 3 Ultra, plans a Korean agent platform in the second half of 2026 and is developing a Seoul World Model with spatial data. These are announced or in-development programs, not completed products until availability, terms and measured results are published.

5. Sovereignty is operational, not absolute

Local operation by a Korean company can strengthen control over processing location, access, language and national workloads. Yet the stack still depends on NVIDIA platforms, global capital and international chip, memory and energy supply chains. Sovereignty here is distributed across location, operations, data and contracts; it is not technological autarky.

The IEA reported that data-center electricity use rose 17% in 2025 and warned that transformers, turbines, chips and grid connections are constraining builds. A concentrated 200MW load is major industrial infrastructure. Power agreements, connection milestones, cooling and efficiency deserve the same scrutiny as the GPU platform.

6. What other sovereign-AI builders should copy—and avoid

The useful pattern is a local cloud operator, a platform provider and an infrastructure-capital partner, with expansion tied to multi-tenant demand. The mistake would be copying the headline capacity without matching grid readiness, customer contracts, local-model work and operating talent.

For Gulf, Asian or European projects, track four ledgers separately: power that is deliverable, capital that is binding, equipment with delivery slots and usage under contract. Local hosting alone does not create an ecosystem. The infrastructure must support domestic languages, evaluation, startups and skills if sovereignty is to compound economically.

7. The investor and policy dashboard

The next high-quality signal is not another partnership announcement. It is conversion of Brookfield’s term sheet into binding agreements and disclosure of NVIDIA’s close. Then come power and construction permits, equipment orders, the 55MW operating date and named customers or capacity contracts.

For NAVER shareholders, capital spending and financing must be separated from prospective revenue. Idle capacity earns nothing; a long-duration customer contract may matter more than another capacity target. For government, the test is whether the factory produces local models, companies and skills—not only a large facility.