NAVER’s 200MW AI factory: inside the $10 billion plan

NAVER, NVIDIA and Brookfield plan to expand an initial 55-megawatt AI factory to 200MW by 2028 at NAVER’s GAK Sejong data center in South Korea. They put the project at $10 billion and describe a longer route to 1GW. Those are targets, not operating assets: NVIDIA’s $1 billion investment is conditional, Brookfield’s offer of up to $9 billion is nonbinding, and the detailed power and commissioning schedule remains undisclosed (NVIDIA’s July 24 announcement).

55MW

Initial planned deploymentFirst-half 2027 target

200MW

Target capacity by 2028About 100,000 GPUs in NAVER’s estimate

1GW

Long-term directionNo completion date disclosed

$10bn

Proposed project sizeIncludes conditional and nonbinding finance

What did the three companies announce?

The physical project is a multi-tenant NVIDIA DSX deployment at GAK Sejong, NAVER’s hyperscale data center in Sejong. A June roadmap called for the first 55MW to begin operating in the first half of 2027, reach 100MW of overseas infrastructure during that year, and grow to 200MW in 2028. The July release put the 200MW expansion at the Korean site, added Brookfield as capital partner and framed the facility as sovereign AI infrastructure.

NAVER says 200MW is roughly equivalent to 100,000 NVIDIA GPUs. The facility is expected to include Blackwell and Vera Rubin platforms and sell production-scale computing to AI companies in Korea and the United States. The companies also tied it to model work: NAVER is adapting HyperCLOVA X around NVIDIA’s Nemotron 3 Ultra open models and building a Seoul world model with NVIDIA Cosmos (NAVER’s July 25 release).

Announced item	Current wording	Why it is not an operating result
55MW	Initial buildout; first-half 2027 operating target	Grid, hardware delivery and customer start dates are not detailed
200MW	Expansion planned by 2028	Phased construction and live capacity are not disclosed
About 100,000 GPUs	NAVER’s scale estimate	Product generation and system configuration can change the count
1GW	Long-term intention	Site, financing, power and completion date are missing
$10 billion	Proposed financing envelope	Conditional investment and a nonbinding term sheet are included

“AI factory” does not mean a semiconductor fabrication plant. It is NVIDIA’s term for a data-center system that combines accelerators, networking, storage, power, cooling and software to produce model training and inference output. The facility runs GPUs; it does not manufacture them.

What does 200MW tell us—and what does it hide?

Megawatts measure power, not FLOPS, tokens or revenue. If a 200MW facility drew its nameplate capacity for all 8,760 hours of a year, simple arithmetic gives 1,752GWh, or 1.752TWh. The equivalent ceiling is 0.482TWh at 55MW and 8.76TWh at 1GW. These are not consumption forecasts. A phased opening, utilization, maintenance, cooling load and power usage effectiveness, or PUE, will change the result.

Initial plan55MW

2028 target200MW

Long-term path1,000MW1GW

Source: NAVER and NVIDIA announcements, June–July 2026

Capacity stages in NAVER’s AI factory plan

Initial plan	55MW
2028 target	200MW
Long-term path	1000MW

Moving from 55MW to 200MW is a 3.64-fold increase in nameplate capacity. Reaching 1GW requires another fivefold expansion. “Gigawatt-scale” therefore describes a direction; the 200MW phase is one-fifth of that destination.

The 100,000-GPU comparison should not become a universal conversion ratio. Blackwell and Vera Rubin products have different power and rack designs. CPUs, HBM, network switches, storage, pumps, uninterruptible power supplies and electrical losses also draw from the facility envelope. Better software can increase tokens per megawatt without adding GPUs. The capacity tells us the size of the container. Commercial performance depends on throughput and revenue per megawatt.

This is similar to the memory bottleneck described in Merci News’s guide to why high-end AI depends on Korean HBM. Adding accelerators without enough memory bandwidth, network fabric or power does not produce proportional model output.

How is the $10 billion supposed to be financed?

NVIDIA’s release gives three pieces. NVIDIA plans to invest $1 billion in NAVER. Brookfield has entered a nonbinding term sheet to fund up to $9 billion as the exclusive capital partner. NAVER will fund remaining amounts needed for the project.

The closing condition is more important than the headline addition. NVIDIA’s investment requires customary closing conditions and NAVER to finalize at least $9 billion in committed financing separate from NVIDIA’s investment. The structure is not NVIDIA wiring $1 billion first and automatically unlocking the rest. NAVER has to secure the other commitment.

Capital source	Announced amount	Current status	Key condition
NVIDIA	Planned $1 billion	Subject to closing conditions	NAVER must secure at least $9 billion separately
Brookfield	Up to $9 billion	Nonbinding term sheet	Final diligence, contracts and capital structure
NAVER	Remaining amount	Funding intention	Depends on Brookfield’s final contribution

If Brookfield provides the full $9 billion and NVIDIA closes $1 billion, the arithmetic reaches the project envelope. “Up to” leaves room below that number, which is where NAVER’s remaining amount becomes material.

The scale also explains why infrastructure finance is involved. The same joint release says NAVER generated KRW 12.04 trillion, or $8.18 billion, in 2025 revenue. A $10 billion project is larger than that annual revenue figure. Revenue and project finance are different measures, but the comparison rules out reading this as a routine capital expenditure funded from one year of operating cash.

What does NVIDIA DSX add to the business?

NVIDIA describes DSX as a codesigned stack spanning chips, systems, software, facilities and partner technology. DSX MaxLPS is meant to maximize token throughput per megawatt; DSX OS manages lifecycle, health, resilience and multi-tenant operations. Those are vendor claims. NAVER will still need to show performance and cost on paying workloads.

The multi-tenant part matters. This is not only a private cluster for HyperCLOVA X. NAVER wants to sell training and inference capacity to startups, enterprises, public organizations and international AI cloud customers. Its search, advertising, cloud and model operations give it experience running large services; the new bet is that this experience can become an infrastructure product.

GPU ownership alone does not make the product profitable. Returns will depend on reserved-capacity contracts, hourly or token pricing, utilization, electricity, depreciation and financing costs. Accelerators can arrive before customers, and a new generation can cut the value of installed hardware faster than a conventional building depreciates.

Scale can work in the other direction. A shared pool can spread networking, cooling and operations across customers and replace failed equipment without taking an entire customer cluster offline. The decisive disclosure would be contracted capacity and revenue per megawatt, not the gross GPU count.

The self-hosted open-model guide explains the smaller version of this tradeoff: control improves when an organization owns more of the stack, but hardware utilization and operating labor determine whether that control is economical.

How can this be “sovereign” with foreign chips and capital?

Sovereign AI does not necessarily mean every component was designed and financed inside one country. In this project it points to data and compute located under Korean legal jurisdiction, operated by a Korean company, with local control over access and customer policy. Banks, hospitals and government agencies may value a large domestic option when data-transfer rules matter.

The dependencies remain obvious. NVIDIA supplies the accelerator platform, DSX software and several model components. Canada-headquartered Brookfield supplies infrastructure capital. A definition based on complete technological self-sufficiency would make the label contradictory.

A more accurate description is a globally supplied stack under Korean operational control. That can increase data sovereignty while leaving exposure to NVIDIA pricing, allocation, export controls and software licensing. Sovereignty is a spectrum of control points, not a binary country-of-origin sticker.

There is also mutual dependence. NVIDIA systems use Korean HBM, while Korean cloud providers need NVIDIA accelerators. Neither side is simply a buyer or supplier. The project turns that hardware relationship into a local service business, but it does not remove the supply chain.

Where could power and cooling block the plan?

A 200MW AI facility needs more than server-room outlets. It requires high-voltage grid connection, substations, redundant distribution, backup power, cooling and fiber. AI load also runs around the clock, so it cannot always shift away from peak hours like a flexible industrial process.

Hardware is useless if grid connection arrives late. Transmission upgrades and substations require permits, equipment lead time and community agreement. To make the 2028 target credible, NAVER needs to disclose connection milestones alongside construction progress. The power source and contract structure will also determine exposure to electricity prices and emissions.

Cooling is another missing line. Dense Blackwell and Vera Rubin racks can increase liquid-cooling requirements. Water consumption, heat rejection and seasonal efficiency affect operating cost and local impact. NVIDIA’s tokens-per-megawatt pitch addresses computing output, but facility-wide PUE and water usage effectiveness, or WUE, are needed for a fair comparison.

Local benefits deserve the same precision. Construction, property-related revenue and operations jobs can help Sejong, but permanent employment does not rise in a one-to-one ratio with megawatts. Automated data centers can use vast power with fewer ongoing jobs than a factory. A credible local bargain should state grid and water costs alongside jobs and tax revenue.

Should the announcement be believed?

The plan is more specific than a slogan. It names the operator, site, platform, capital partner, 2028 target and long-term capacity. Yet NVIDIA repeatedly uses proposed for the expansion and planned for investment. It explicitly calls Brookfield’s term sheet nonbinding and includes forward-looking-statement warnings.

Missing details include the GPU mix, final equipment price, grid agreement, PUE, WUE, construction phases, anchor tenants, reservation rate and service pricing. The release does not split the $10 billion among land, buildings, power, cooling, accelerators and financing costs. It also does not clearly define whether 200MW is IT load or total facility load.

“Nonbinding” does not make the project meaningless. Large infrastructure commonly moves from term sheet through diligence, project finance, long-term supply contracts and closing. The June 55MW plan gaining a named capital partner and larger financing structure in July is progress. The next proof must be contractual and physical.

What evidence should come next?

The financing checklist comes first. Brookfield’s term sheet should become a final investment or credit agreement; NVIDIA’s closing condition should be met; NAVER should show its remaining exposure in board decisions and financial disclosures. The difference between “up to $9 billion” and the amount that closes is material.

Power and construction follow. Look for a grid-impact review, power-supply agreement, substation completion, cooling design, first rack delivery and commissioned megawatts. Initial 55MW operating in the first half of 2027 would be the first hard test. In 2028, reports should separate built, energized and commercially available capacity.

Customers decide the economics. Long-term reservations, available GPU types, pricing, service-level agreements and utilization will show whether 200MW is a productive asset or an expensive inventory position. NAVER should also separate capacity used by its own models from capacity sold outside the group.

The clean execution sequence is target capacity → committed finance → grid connection → hardware delivery → paid utilization. The July announcement sits between the first two stages. It can become one of Asia’s larger AI infrastructure projects, but the megawatt headline has not yet crossed the grid meter.

FAQ

When will NAVER’s 200MW AI factory open?

NAVER’s earlier plan targeted initial 55MW operation in the first half of 2027. The July announcement aims for 200MW by 2028, but a detailed commissioning schedule has not been published.

How many GPUs fit in 200MW?

NAVER describes the plan as roughly 100,000 NVIDIA GPUs. That is a company estimate, not a fixed conversion, because accelerator models, servers, networking, cooling and redundancy all consume power.

Are NVIDIA and Brookfield’s $10 billion investments final?

No. NVIDIA’s planned $1 billion is conditional on NAVER securing at least $9 billion of separate committed financing. Brookfield has a nonbinding term sheet for up to $9 billion.

Does sovereign AI mean the Korean government owns the factory?

No. NAVER would operate the privately financed infrastructure. Sovereign refers mainly to local data, operational control and legal jurisdiction, not state ownership or complete independence from foreign technology.