Naver's 200MW AI Factory Prices Sovereignty at $50M/MW

Naver, Nvidia, and Brookfield plan to expand a Korean AI factory from 55 megawatts to 200 megawatts by 2028, backed by at least $10 billion of proposed outside financing. That creates a rough $50 million capital envelope per megawatt—a warning that sovereign compute is a financing and utilization problem before it is a flag on a server rack.

The official Nvidia announcement says Nvidia plans to invest $1 billion in Naver, while Brookfield signed a nonbinding term sheet for up to $9 billion. Naver would fund remaining amounts. Nvidia’s investment depends on customary closing conditions and at least $9 billion of separate committed project financing, so operators should read this as a proposed capital stack, not ten billion dollars already poured into concrete.

A 3.64× expansion needs customers, not slogans

The planned build takes GAK Sejong from 55MW to 200MW, a 3.64× capacity expansion. Divide the announced $10 billion Nvidia-plus-Brookfield envelope by 200MW and the result is at least $50 million per megawatt. That is not the facility’s disclosed construction cost: the capital also includes Nvidia’s equity investment in Naver, financing may support broader infrastructure, and Naver will contribute additional funds. It is a useful upper-level ratio for understanding the scale of capital being organized around each unit of planned capacity.

Naver’s longer ambition is 1 gigawatt. At the same crude capital intensity, that would imply $50 billion, which is precisely why linear extrapolation is dangerous and useful. Larger deployments can spread fixed costs, hardware generations change, and financing structures differ; nevertheless, the exercise exposes the gulf between announcing sovereign capability and keeping it economically occupied. Brookfield already runs an AI infrastructure program targeting $100 billion, showing that Sejong sits inside a portfolio-scale financing strategy rather than a one-off national project.

The proposed factory will use Nvidia DSX with Blackwell and Vera Rubin systems. The Nvidia DSX platform combines facility, compute, networking, and lifecycle software; DSX MaxLPS is designed to maximize tokens per megawatt, while DSX OS manages multi-tenant infrastructure. Naver plans to serve Korean and US model builders, enterprises, industries, and government. The multi-tenant detail matters: sovereignty cannot justify idle accelerators, so the facility needs enough external demand to smooth utilization across training bursts and steady inference.

This is not simply another hyperscaler region. Naver is tying compute to models and applications: HyperCLOVA X will build on Nemotron 3 Ultra; the company joined Nvidia’s Nemotron Coalition; an AI agent platform is planned for the second half of 2026; and a Seoul world model will use proprietary street-view and spatial data. The stack attempts to keep Korean data, language capability, cloud operations, and application distribution in one commercial loop. The demand thesis includes exactly the kind of verified domain workloads described in Nvidia’s new engineering-agent stack, where model calls pull high-value simulations and solvers onto the same accelerated platform.

That makes Solar Open 2’s two-H200 deployment footprint strategically relevant. Sovereign capacity only compounds when models are efficient enough for domestic companies to consume it. A 200MW factory filled with oversized general models is dependence wearing local colocation credentials.

Buy the jurisdiction only when it changes the risk

Korean AI companies, public agencies, and regulated enterprises should negotiate capacity if data residency, Korean-language performance, or guaranteed regional supply changes a real constraint. They should not pay a sovereignty premium for workloads already served cheaply and compliantly by global clouds. The decision model needs four columns: effective token cost, utilization commitment, data/control benefit, and portability.

The cost could include reserved capacity, minimum spend, egress, migration, DSX-specific operations, and model adaptation. No public rate card exists. That absence is the central procurement gap: $50 million per megawatt describes capital assembly, not what a startup will pay per million tokens. Buyers should demand transparent on-demand and committed-use pricing before treating national infrastructure as a competitive cloud.

Naver’s plan also gains a research pipeline. A separate Nvidia-KAIST laboratory carries a stated $300 million collaboration, including $50 million of annual compute over five years and funding for at least ten researchers each year. Divide $250 million of compute by a minimum of 50 annual researcher slots and the nominal envelope reaches $5 million of compute per funded researcher-slot. That ratio is not a personal grant, but it shows how aggressively Korea is coupling talent to the same Nvidia stack.

For operators, this vertical integration can shorten the path from local research to production capacity. It can also narrow technical choice. A domestic model trained on Nemotron, deployed through DSX, and commercialized on Naver Cloud is sovereign by geography while remaining concentrated by architecture. Procurement should distinguish data jurisdiction from supplier diversity.

The thesis breaks if Brookfield’s nonbinding $9 billion does not close, Nvidia’s conditional investment stalls, grid or construction milestones slip, or tenant demand fails to fill the 3.64× expansion. It also weakens if customers cannot move models and data out without punitive costs. Sovereign infrastructure can reduce geopolitical dependency while increasing vendor concentration in Nvidia hardware and Naver’s operating stack.

That concentration echoes the $500-billion-plus Nvidia and SK Hynix memory framework: control increasingly comes from coordinating chips, memory, power, financing, and software rather than owning one component. It also sits inside Alphabet’s $811 billion commitment stack, where long-dated capital promises are remaking cloud economics before utilization is proven.

Evidence that changes the verdict includes binding financing, construction milestones, power-delivery contracts, tenant reservations, published token economics, and portability terms. Until then, operators should secure options rather than lock their stack: benchmark Korean workloads, estimate sustained demand, and negotiate a small committed tranche with expansion rights.

The proposed $10 billion stack buys Korea an option on compute independence. The derived $50 million per megawatt ratio shows why that option must produce more than patriotic capacity. Sovereignty earns its premium only when it lowers regulatory risk, improves local model performance, or guarantees supply—and when the servers stay busy enough to make those benefits affordable.