Korea’s sovereign AI factory is really a $10B financing story

The headline number out of Korea this week is 200 megawatts. It is the wrong number to watch. NAVER, NVIDIA and Brookfield said they will more than triple the NVIDIA DSX AI factory at NAVER’s GAK Sejong data center from 55 megawatts to 200 megawatts by 2028, on the way to a stated ambition of a full gigawatt of sovereign AI capacity. That is a lot of power. But the figure that actually explains the deal sits in the financing line, not the electrical one: NVIDIA is investing about $1 billion into NAVER, and Brookfield has signed a nonbinding term sheet to put up as much as $9 billion more. Strip away the branding and a national AI factory looks less like a computing announcement and more like a capital structure with GPUs attached.

Article Brief

Key Takeaways

5 Points30s Read

1. 01What’s new–NAVER, NVIDIA and Brookfield plan to expand the NVIDIA DSX AI factory at NAVER’s GAK Sejong data center from 55 megawatts to 200 megawatts by 2028, targeting a longer-term goal of one gigawatt.
2. 02The money–NVIDIA is investing roughly $1 billion in NAVER, and Brookfield has a nonbinding term sheet to fund up to $9 billion; NAVER covers the rest. The capital, not the megawatts, is the story.
3. 03Why it matters–It shows how sovereign AI actually gets paid for: global infrastructure capital and a chipmaker’s equity, not a government budget line.
4. 04The catch–The factory runs on NVIDIA’s Vera Rubin and Blackwell silicon, so it is sovereign in location and control but dependent in hardware.
5. 05Who should care–AI startups eyeing the promised compute pool, investors pricing NVIDIA’s sell-chips-and-take-equity flywheel, and policymakers weighing what sovereignty actually buys.

What the three companies actually committed to

The build sits at NAVER’s GAK Sejong hyperscale data center in Sejong, South Korea, and runs on the NVIDIA DSX platform the three companies detailed this week. Capacity is meant to climb from today’s 55 megawatts to 200 megawatts by 2028, using NVIDIA’s Vera Rubin and Blackwell systems, with a stated aim of reaching one gigawatt over time. The pitch is that the expanded factory becomes a dedicated resource pool for emerging Korean AI companies, compute, software and support in one place, rather than a single tenant’s private cluster.

Two caveats belong up front, because the announcement leans on numbers that are not all the same kind of number. The 200-megawatt target carries a date, 2028; the one-gigawatt figure is a direction of travel, not a signed contract. And the largest dollar figure in the release is the softest: Brookfield’s up-to-$9-billion commitment is a nonbinding term sheet, which means it is an intention to finance, not financing. NVIDIA’s roughly $1 billion investment into NAVER is the firmer piece, and NAVER has said it will fund whatever the two partners do not.

The capital stack is the real product

What makes this deal worth reading closely is who is writing the checks and why. Brookfield is not a technology company; it is an asset manager that runs roughly $100 billion in AI-related infrastructure and holds about $12 billion of assets in Korea alone, according to the joint statement’s financial disclosures. When a firm like that funds a data center, it is treating compute the way it treats a toll road, a port or a power plant, a long-duration asset that spits out contracted cash flows. That is a quiet but important reframing: AI capacity is being financed by the same machinery that finances physical infrastructure, on similar time horizons and return expectations.

NVIDIA’s contribution is the more interesting one. A roughly $1 billion equity stake in NAVER is small next to the chip orders the factory implies, and that is the point. NVIDIA has repeatedly taken minority positions in the infrastructure partners that deploy its chips, turning a customer relationship into a flywheel: it sells the silicon, holds equity in the buyer, and books the sovereign-AI revenue on top. Each national factory it seeds makes the next one easier to justify. Layer that onto the broader capex wave now underwriting a striking share of US growth, and the NAVER deal reads as one more node in a spending structure that has learned to fund itself.

There is a real risk buried in that neat structure, and it is worth naming rather than waving away. A toll road holds its value for decades; a hall full of GPUs does not. Infrastructure investors underwrite assets on fifteen- and twenty-year horizons, but the Blackwell and Vera Rubin systems going into GAK Sejong will be two or three generations old well before then, and the economics of an AI factory depend on keeping it full of paying workloads the whole way. That is why the startup compute pool is not just good public relations; it is the demand story the financing needs. If Korean AI demand arrives on schedule, the toll-road analogy roughly holds. If it lags, Brookfield is financing a depreciating asset dressed as durable infrastructure, and the nonbinding nature of its term sheet starts to look less like caution and more like foresight.

Sovereign in name, NVIDIA in the racks

The word doing the heaviest lifting in the announcement is sovereign. In some dimensions it is earned: the data sits in Korea, the factory is operated by a Korean company, and domestic startups get first call on the capacity. In one dimension it plainly is not: the racks are filled with NVIDIA’s Vera Rubin and Blackwell hardware, and NVIDIA owns a slice of the operator. Sovereignty over where your AI runs is not the same as sovereignty over what it runs on.

The contrast is sharp against the other gigawatt story of the month. Z.ai brought a comparable gigawatt of compute online on Chinese-made accelerators with no NVIDIA inside, a harder and riskier version of sovereignty forced by export controls. Korea’s model is the opposite trade: buy the best available silicon, keep control of the site and the data, and accept hardware dependence as the cost. Neither is obviously right, but they are genuinely different bets, and the NAVER deal shows which one most governments will pick when NVIDIA chips are still on the table. It also sharpens a question the raw megawatt figure hides, what a megawatt of Vera Rubin capacity actually converts into in usable tokens, because a factory’s real output is measured there, not in its power draw.

The part the megawatt figure hides

A 200-megawatt factory is not just a procurement question; it is an energy commitment, and the jump to a gigawatt is an even larger one. Two hundred megawatts of continuous draw is roughly the appetite of a small city, and it has to be delivered as firm, around-the-clock power to a single site in Sejong. That is where sovereign AI ambitions most often meet physics. Grid interconnection queues, transmission upgrades and the sheer availability of baseload power have become the binding constraint on data-center expansion worldwide, frequently more so than chip supply. Korea has a dense, capable grid, which is part of why the site works at all, but a five-fold ramp to a gigawatt implies power procurement and grid coordination that no term sheet settles on its own.

This is the quiet reason the 2028 date matters more than the gigawatt headline. Electricity infrastructure runs on permitting and construction timelines that do not compress the way software roadmaps do, and a factory that cannot be energized on schedule is capacity on paper only. When a national AI project stalls, it usually stalls here, at the substation, not the server, and the megawatt figure everyone quotes is precisely the variable most exposed to that risk.

Why NVIDIA keeps signing national deals

For NVIDIA, sovereign AI has stopped being a side story. It has quietly become one of the company’s fastest-growing revenue lines, and management has flagged sovereign demand as a meaningful contributor to the segment that houses its AI clouds, industrial and enterprise business, a group that generated $37.4 billion in a single recent quarter, up 74% year over year. A government or national champion that commits to a gigawatt roadmap is close to the ideal customer: large, sticky, politically motivated to keep spending, and unlikely to shop the order to a rival. Every NAVER-style agreement also normalizes the template, which is why the pattern keeps repeating across countries.

What it means for the people downstream

For Korean AI startups, the promise is real but deferred. A dedicated national compute pool is exactly what a young lab short on capital wants, access without owning the hardware. The open questions are price and timing: a pool announced for 2028 does little for a team trying to train a model this year, and available capacity that is priced like a hyperscaler helps less than it sounds. The support-and-software layer NVIDIA bundles in matters here as much as the raw chips.

For investors, the tell is whether the softest number hardens. The gap between a nonbinding term sheet and drawn financing is where deals like this succeed or quietly shrink, so Brookfield’s next move is the one to watch, not NAVER’s press cadence. And NVIDIA’s equity-plus-silicon structure is worth pricing directly: it is a durable way to compound sovereign demand, but it also concentrates the AI buildout’s fortunes in a single vendor’s roadmap. A stake in the operator, the chips in the racks and the software layer on top means NVIDIA captures value at three points in the same deal, excellent for NVIDIA, and a concentration risk for anyone betting the whole sovereign-AI category on one company’s cadence.

For policymakers elsewhere, the NAVER model is now the reference design, and it clarifies a choice more than it removes one. It buys control over where AI runs and whose data it touches, and it hands domestic startups a compute base they could not build alone. It does not buy independence from the company that makes the accelerators, and it quietly imports that company as a shareholder. Governments drawing up their own AI-factory plans are really deciding how much hardware dependence they will trade for speed, the same fork Korea and China have now answered in opposite directions. That is a strategic decision wearing an infrastructure announcement’s clothes, and it deserves to be argued as one.

What to watch next

Three markers will show whether this is a buildout or a banner. First, the 2028 milestone: 200 megawatts is a concrete, dateable claim, so progress against it is checkable. Second, the conversion of Brookfield’s up-to-$9-billion term sheet into signed financing, which turns intention into infrastructure. Third, whether the promised startup compute pool actually lowers the cost of training for Korean labs, or simply relocates NVIDIA-priced capacity onto home soil. Get those three right and Korea has a genuine sovereign AI base. Miss them and 200 megawatts stays what it is today, a number in a press release.