In January 2026, Korea’s government published a shortlist that did not include Naver. Three organisations won the right to build the country’s “national champion” foundation models. Naver, the company most foreigners know as the search engine that beat Google in Korea, was cut in the first round. Six months later the same company announced the Naver AI factory, a compute build-out large enough to reshape the national grid.
The turn was abrupt. Jensen Huang flew to San Francisco to sign a deal with Naver’s founder. Nvidia bought 4.5 percent of the company. Brookfield agreed to arrange up to $9 billion. Together, the project became the largest single piece of AI infrastructure ever attempted by a Korean private company.
The whiplash is the story. Naver lost the contest to build Korea’s sovereign AI model, and then won the far bigger job of building the machine that everyone else’s models will run on. For anyone trying to understand how Korea is actually financing its AI ambitions, that reversal explains more than any government white paper.
On July 24, 2026, Naver chairman Lee Hae-jin and Nvidia chief executive Jensen Huang signed an agreement in San Francisco. The terms surfaced in Seoul three days later, and they were unusual.
Nvidia is taking 7,241,564 newly issued shares for roughly $1 billion, or about 1.48 trillion won. That gives the chipmaker a 4.5 percent stake and third place on Naver’s shareholder register, behind only the National Pension Service and BlackRock. It is also Naver’s first capital raise in 22 years.
Investors reacted immediately. Naver shares jumped about 10 percent on the news, which is a large move for a 27-year-old company with a stable earnings base.
Yet the equity stake is the smaller half of the arrangement. Alongside it, Nvidia, Brookfield and Naver announced an infrastructure programme in which Brookfield commits up to $9 billion. Brookfield already manages roughly $100 billion across data centres, compute and power generation, so it brings the balance sheet that a Korean internet firm does not have. Naver funds the remainder.
Add the two numbers and you get about $10 billion of committed financing. For context, that is comparable to the market capitalisation of several mid-sized KOSPI companies, aimed at a business line that currently produces no revenue at all.
Foreign readers usually meet Naver as a wall of green boxes and blog posts. That impression is two decades old, and it hides an unusually early infrastructure habit.
Naver launched in 1999. By 2013 it had opened GAK Chuncheon, the first data centre in Korea built and owned by an internet company rather than a telecom carrier or a chaebol IT arm. At the time the decision looked eccentric. Rivals rented capacity; Naver poured concrete in a mountain province.
The AI pivot came next. In 2021 the company unveiled HyperCLOVA, Korea’s first hyperscale language model, with 204 billion parameters — larger than the GPT-3 that had stunned the industry a year earlier. More importantly, it was trained overwhelmingly on Korean text, at a volume no foreign lab had bothered to assemble. HyperCLOVA X, the conversational successor, arrived in August 2023.
Then GAK Sejong opened three months later. Seen in sequence, the 2026 announcement is not a swerve. Rather, it is the fourth step in a pattern Naver has repeated for thirteen years: build the physical layer first, and find the demand afterwards.
The phrase “AI factory” is Nvidia marketing language, so it deserves a plain translation. An AI factory is a data centre designed around one job: turning electricity into machine intelligence at industrial scale. Instead of hosting websites, it trains and runs models.
Naver’s version sits at GAK Sejong, its second data centre campus, about 110 kilometres south of Seoul. The scale-up plan runs in three steps. First, 55 megawatts of Nvidia infrastructure are already committed. Second, capacity expands to 200 megawatts by 2028. Finally, the stated ambition is a full gigawatt.
That last number is worth pausing on. A gigawatt is roughly the output of a large nuclear reactor. No Korean company has ever pointed that much power at a single computing workload.
The hardware roadmap covers Nvidia’s Blackwell generation and the newer Vera Rubin platform, deployed on Nvidia’s DSX stack. In addition, Naver plans to fine-tune Nvidia’s Nemotron models into its own HyperCLOVA X family, and to use Cosmos world models to build a “Seoul world model” from years of accumulated street-view data.
GAK Sejong is not a rendering. The campus opened in November 2023 after four years of planning and construction, at a cost just under $500 million.
Its specifications read like a small city. Buildings and grounds cover 290,000 square metres. At full build-out across three phases, it is designed for 270 megawatts and up to 600,000 servers. Moreover, the cooling system is homegrown: an in-house unit called NAMU pulls in natural outside air rather than relying entirely on mechanical chillers.
Other details show how long Naver has been preparing. Waste heat warms the floors and the facility’s water. A snow-melting system keeps the access roads clear in winter. Robots named GaRo and SeRO haul equipment and audit assets inside the halls, coordinated by the same ARC platform Naver uses in its robot business. The structure is engineered to survive a magnitude-7 earthquake.
In other words, the Naver AI factory is being retrofitted onto real estate that already exists, already has grid connections, and already runs. That is not a small advantage in 2026.
Now return to January. Korea’s Ministry of Science and ICT ran a competition to select the developers of independent, or “from scratch,” foundation models. The winners announced on January 15 were LG AI Research, which scored highest, along with SK Telecom and the startup Upstage. Upstage’s selection in particular was read as a signal that scale alone would not decide the outcome.
Naver Cloud did not make the cut. The reason was specific and, for the company, embarrassing. Its HyperCLOVA X SEED 32B Think model used a vision encoder derived from Alibaba’s Qwen 2.5 — Chinese open-source technology inside a project whose entire premise was technological independence. One analysis put the similarity between the two encoders above 99 percent. Officials judged that this failed the independence criterion.
The controversy ran through Korean tech media for weeks. Naver’s initial defence was that the encoder was a swappable component. Critics answered that a sovereign model built on someone else’s eyes is not sovereign. The company declined to enter the re-screening round.
Then Naver did the engineering. By March 2026, Naver Cloud had finished its own vision encoder, one trained to link images directly to Korean rather than routing through English. Performance, the company said, matched top-tier alternatives including Qwen. Crucially, it handled Korean proper nouns and cultural imagery better, because it never needed a translation layer.
So the sequence matters. Naver lost a government beauty contest in January, fixed the underlying weakness by March, and signed with Nvidia in July. The government picks who builds Korea’s models. The market, meanwhile, picked who builds Korea’s compute.
Ambition of this size shows up in financial statements long before it shows up in revenue. Naver’s second-quarter results, released in August, are the clearest place to see the trade.
Revenue reached 3.3888 trillion won, up 16.2 percent year on year — a record. Operating profit, however, slipped 0.2 percent to 520.3 billion won. Net income told a different story again, jumping 41.6 percent to 704.3 billion won, or about $495 million.
The segment detail explains the squeeze:
| Segment | Q2 2026 revenue | Change |
|---|---|---|
| Naver Platform (search, ads, services) | ₩1.9022 trillion | +12.3% |
| Global initiatives (C2C, content, enterprise) | ₩1.0159 trillion | +24.4% |
| Financial platform (Npay, Smart Store) | ₩470.7 billion | +16.0% |
| Total | ₩3.3888 trillion | +16.2% |
Two lines deserve attention. Advertising revenue grew 7.5 percent, and Naver says AI accounted for more than 60 percent of that growth — a striking claim from a company whose search franchise was supposed to be dying. Separately, C2C revenue rose 74.9 percent, driven by the resale platforms Naver has been quietly assembling overseas.
Npay processed 25.2 trillion won in the quarter, up 21 percent. That payments engine is the same one now sitting alongside Korea’s largest crypto exchange operator inside Naver Financial.
Chief executive Choi Soo-yeon framed the margin hit directly. The AI factory, she said, is “an investment in a structurally growing market, not a one-off business opportunity.” Management expects the first revenue from it in the first half of 2027. Until then, the cost lands in the income statement and the benefit does not.
Notably, Naver also retired about 4.9 million treasury shares worth roughly 1 trillion won. Buying back stock while raising 1.48 trillion won from Nvidia looks contradictory at first glance. In practice it is a signal: the company wants the dilution absorbed, not resented.
Here is where the Naver sovereign AI plan collides with physics. AI hardware is extraordinarily power-hungry, and Korea’s grid was not built for it.
A single server holding eight Nvidia B200 GPUs draws roughly 20 kilowatts once cooling is counted. Pack four of those into a rack and you are at about 80 kilowatts. A conventional data centre rack, by contrast, runs at three to five kilowatts. The difference is not incremental; it is a different category of building.
Meanwhile, more than 70 percent of Korea’s data centres sit in the capital region, where transmission capacity is already strained and local opposition to new lines is fierce. This is precisely why the grid has become the binding constraint on Korea’s AI boom, and why Naver’s decision to put its second campus in Sejong, rather than Gyeonggi Province, now looks prescient rather than merely thrifty.
The state is pushing regardless. Seoul aims to catalyse 65 trillion won of private AI investment between 2024 and 2027, and it raised the national AI budget to 10 trillion won for 2026. Even so, budgets do not build substations. Whether 200 megawatts in Sejong can actually be energised on schedule is the single largest execution risk in the entire project.
Compare the numbers internationally and the ambition looks less outlandish, though still stretching. A gigawatt would put the Naver AI factory in the same conversation as the largest American campuses now under construction, and well beyond anything operating in Japan or Southeast Asia today. Korea has one genuine advantage in that race: an industrial base that manufactures the memory these machines consume. It also has one genuine handicap, since it imports nearly all of its primary energy.
That tension defines the whole project. Korea can build the chips and the buildings faster than almost anyone. Powering them, by contrast, means confronting an energy debate the country has deferred for a decade.
Naver is not alone, although its allocation is larger than most coverage suggests.
In late October 2025, Nvidia announced that more than 260,000 of its GPUs would be deployed across South Korea. The distribution was revealing.
| Recipient | Nvidia GPUs | Purpose |
|---|---|---|
| Naver Cloud | 60,000+ | Sovereign cloud, physical AI |
| Samsung Electronics | 50,000+ | Semiconductor AI factory |
| SK Group | 50,000+ | AI factory and cloud |
| Hyundai Motor Group | 50,000 | Mobility and manufacturing |
| Ministry of Science and ICT | up to 50,000 | National sovereign cloud |
Naver Cloud received the single largest corporate allocation. Furthermore, it was named as one of three operators — with NHN Cloud and Kakao — running the government’s own sovereign cloud deployment. So while Naver lost the model contest, it never lost its position in the infrastructure layer.
This also connects to the wider semiconductor story. Every one of those GPUs needs high-bandwidth memory, which is why the HBM race between SK Hynix and Samsung sits directly upstream of Naver’s build-out, and why Korea’s domestic AI chip startups are pitching themselves as an eventual alternative for inference workloads.
Most coverage of this deal is written for equity analysts. For foreigners working in Korea, however, the practical consequences arrive sooner than the revenue does.
If you run a startup here, the relevant date is the second half of 2026, when Naver plans to open its AI agent platform. Until now, Korean founders training or serving models at scale have leased capacity from Amazon, Microsoft or Google, priced in dollars and hosted abroad. Domestic supply changes the negotiating position, even for teams that ultimately stay on a foreign cloud.
If you work in a regulated industry, the sovereign angle is not marketing. Korean financial firms, hospitals and public agencies operate under data-residency rules that make offshore processing legally awkward. Consequently, a large domestic AI facility is less a patriotic gesture than a compliance product. Naver Cloud has already signalled industry-specific models for shipbuilding and security.
If you build products in Korean, the encoder fight from January matters more than it sounded. A model that links images and text directly in Korean handles local proper nouns, signage, menus and cultural references without an English translation step. In practice, that is the difference between an AI feature that works in Seoul and one that almost works.
If you are simply living here, expect the effects to be indirect but real. Electricity policy, regional development spending and even office rents in Sejong all move when a gigawatt of demand appears on a planning map.
Investors looking at Naver in 2026 are really underwriting two businesses: a profitable internet company and an unproven infrastructure venture. Reasonable people disagree about the second.
The bull case rests on assets that capital cannot quickly replicate. Naver owns the land, holds the grid connections, and has operated large data centres since 2013. It writes its own cooling systems, runs its own robots inside the halls, and now has Nvidia as both supplier and shareholder — an alignment few customers enjoy. In addition, demand for sovereign compute is genuinely structural: Korean banks, hospitals and public agencies face data-residency rules that make foreign hyperscalers awkward. Naver also has adjacent pull, since its digital twin platform was adopted as Saudi Arabia’s national smart-city standard in August 2026, and those workloads have to run somewhere.
The bear case is equally concrete. Naver Cloud is a distant challenger in Korea’s public cloud market, where Amazon Web Services has long dominated and Samsung SDS has climbed past Microsoft. Selling raw compute is a capital-intensive, low-margin business that rewards operators with global scale, and Naver has neither. Depreciation on GPU fleets is brutal, particularly when Nvidia ships a new architecture roughly every year. Moreover, $9 billion of infrastructure financing carries interest whether or not the racks are full. Revenue does not begin until 2027, by which point Korea’s data centre boom may already be crowded with chaebol-backed competitors chasing the same enterprise customers.
There is a third reading, and it may be the most useful. The Naver AI factory is partly a hedge against Naver’s own core business. Search advertising in Korea remains lucrative, yet nobody at the company assumes it will still be lucrative in 2035. Owning compute is a way of staying relevant even if the query box eventually disappears into an assistant.
The honest verdict is that both cases are live. Naver has converted a strategic setback into a strategic position, but it has done so by taking on the balance-sheet risk of a utility while retaining the cost structure of an internet firm.
For readers tracking this rather than trading it, four checkpoints will settle most of the argument.
It is easy to file Naver under “Korea’s Google” and stop there. That framing is now badly out of date.
The company runs the country’s most-used search engine, a payments network processing 25 trillion won a quarter, a global resale portfolio, a crypto exchange, a robotics division exporting city-scale digital twins to Saudi Arabia, and shortly, one of Asia’s larger AI compute facilities. Few Western firms combine that spread. Alphabet comes closest, and even Alphabet does not own a used-goods marketplace in Spain.
What the Naver AI factory really represents is a wager about where value settles in the AI era. Naver is betting it settles in the layer underneath the models — in power contracts, land, cooling and racks — rather than in the models themselves. Having been told in January that its model was not sovereign enough, the company decided to own the ground those models stand on instead.
Whether that was vision or consolation, the concrete is already pouring in Sejong.
In 1997, years before Naver existed, a small Seoul company handed Koreans something nobody else…
Land at Incheon, open your phone, and the very first thing that breaks is the…
At 7:30 on a summer evening in July 2026, a fishing crew off Bangeojin hauled…
For centuries, the Korean fire festival season opened the same way. On the eve of…
The bowl arrives still boiling, and nobody at the table asks where the bird came…
Nobody comes to South Korea expecting a desert. Yet roughly 19 kilometres from the North…