Why Jensen Huang Visited Naver: The Blueprint for the Sovereign AI Alliance

On the night of June 5, 2026, the checkout counter at ‘Hyungnim Jeoyo’, a Korean BBQ restaurant near Hongdae, originally featured only a Toss facial payment terminal. That day, one more terminal—a Naver Pay ‘Npay Connect’—was added.

After the meal, Naver Chairman Lee Hae-jin stood in front of the terminal.

As he presented his face, a payment success notification appeared. According to a field report by Financial News, NVIDIA CEO Jensen Huang, who was watching from the side, let out a brief exclamation of surprise. SK Chairman Chey Tae-won and LG Chairman Koo Kwang-mo were also at the same table.

Reading through various reports, what caught my eye was the note that the terminal had been ‘installed that very day.’ There were even rumors that Naver rushed to install it upon hearing that Chairman Lee would be the one paying. It was a staged moment. And it signifies just how important that evening was to Naver.

Three days later, on June 8, Jensen Huang appeared at Naver’s second headquarters, 1784, in Bundang, and the two companies announced they would jointly build a gigawatt (GW)-scale AI factory. When you piece together the fragments between a plate of pork belly and a gigawatt contract, it becomes clear that the core of this deal isn’t just about GPUs, but a division of labor aimed at the Sovereign AI market.

The Naver-NVIDIA AI Factory Alliance: What Was Agreed?

In short, the agreement marks Naver stepping down from being a mere customer of NVIDIA to becoming a business partner.

According to ZDNet Korea, Naver will lead the site acquisition, construction, and operation of the data centers, while NVIDIA will provide GPUs and share the burden of global customer acquisition, revenue, and business risks.

The roadmap is phased.

In the first half of 2027, 55 megawatts (MW) will be activated at Naver’s hyperscale data center, ‘Gak Sejong’. The plan is to scale to 100MW within the same year, expand overseas to 200MW by 2028, and eventually reach gigawatt-scale. iNews24 reported that 1GW is roughly four times the maximum capacity of Gak Sejong and large enough to house hundreds of thousands of NVIDIA’s latest GPUs at once.

What’s more noteworthy than the numbers is the risk-sharing structure.

If it were just a transaction involving selling GPUs, NVIDIA would have no reason to shoulder the burden of customer acquisition and revenue. Both companies stated that they are targeting markets in the Middle East and Europe, beyond the Asia-Pacific region, with this infrastructure.

The fact that they chose to build the factory together implies that NVIDIA believes there are markets that can only be unlocked through Naver.

Why Power Has Become the Bottleneck in the AI Factory

As data centers transform into ‘factories,’ the first resource to run short is not GPUs, but electricity.

In the early 2010s, data centers were more like quiet warehouses running web pages and emails, where air conditioning was sufficient for cooling.

The turning point was the launch of ChatGPT on November 30, 2022. As large language models were released to actual users, data centers became facilities that ingest text as raw material and tirelessly churn out intelligence in the unit of ’tokens’.

This is what Jensen Huang calls an ‘AI Factory.’ During the announcement, he mentioned that the demand for AI factories is exploding as practical AI enters the field. Exploding demand simply means a shortage of electricity.

Factories come with factory-level constraints. A standard server rack used to consume around 10 kilowatts (kW) of power, but a rack packed with NVIDIA Blackwell accelerators exceeds 120kW. If ten times the electricity enters the same space, the heat generated increases proportionally.

Therefore, the report card of an AI factory has shifted from ‘how many GPUs’ to ‘how many tokens are extracted per megawatt.’ Power contracts are fixed, and building new power plants takes years. It has become a game of squeezing more computation out of the given electricity.

NVIDIA’s answer is the DSX platform. According to NVIDIA’s description, DSX is a design framework that bundles chips, systems, software, facilities, and partner technologies into a single stack to lower token costs and reduce time-to-market. Within it, DSX MaxLPS is software that boosts token throughput per megawatt, while DSX OS handles lifecycle management, state automation, and multi-tenant operations where multiple customers share one factory.

The principle is intuitive. AI workloads are not uniform; some GPUs run at their limit for training, while adjacent racks sit idle in inference standby. By redirecting power to the GPUs that need it, you can run more GPUs within the same megawatt budget. I covered similar bottlenecks occurring in memory in my article on the 2025 semiconductor super-cycle.

Why Gak Sejong’s Cooling Technology is the Foundation of Sovereign AI Infrastructure

If DSX treats power as code, removing the heat generated by that power is Naver’s job. This is the second asset Naver brings to the alliance.

Since its first data center, ‘Gak Chuncheon,’ opened in Chuncheon, Gangwon-do in 2013, Naver has relied on outdoor air cooling—drawing in cold air from the mountainside—instead of traditional air conditioning.

Gak Sejong, which opened in 2023, went a step further. Conventional data centers push cold air to the corners of server rooms through large ducts, which create friction, requiring fans to run harder. Fans also consume electricity.

Gak Sejong’s cooling system, NAMU III, eliminates ducts and places BLDC (Brushless DC) fans in a dense grid, moving more air with less power.

The problem is that Blackwell and the next-generation Vera Rubin accelerators are pushing the physical limits of air cooling.

When heat exceeds 100kW per rack, relying solely on air cooling causes the electricity required for cooling to skyrocket. This is why direct-to-chip liquid cooling or immersion cooling, where servers are submerged in insulating oil, has become a challenge for the entire industry.

This is why NVIDIA specifically pointed out that Gak Sejong is ‘designed for high-density NVIDIA accelerated computing environments.’ You can buy GPUs if you have the money, but a building that keeps those GPUs running without melting can only be built quickly by a company that has been designing and operating data centers for over a decade. Simple buyers purchase racks; strategic partners design the space where those racks stand.

Why the Sovereign AI Model HyperCLOVA X Rides on Nemotron 3 Ultra

On the model front, instead of developing everything alone, Naver chose to base its work on NVIDIA’s open models.

According to NVIDIA, Naver fine-tunes Nemotron 3 Ultra with its own data and training expertise to enhance HyperCLOVA X, and then uses this model as a platform for Sovereign AI business in Europe and the Middle East.

The official model name of Nemotron 3 Ultra includes ‘550B-A55B’.

This means it has 550 billion parameters in total, but only 55 billion are actually activated at once. Most language models activate all weights even when predicting a single word. This model uses a Mixture of Experts (MoE) structure, where parameters are split into several ’expert’ bundles, and only the bundles matching the input are used. By activating only 10% of the total, it can extract more tokens with the same power. This is the exact same direction as the ’tokens per megawatt’ game we saw earlier.

The Nemotron 3 series also incorporates Transformer and Mamba layers.

Transformers suffer from computation that grows quadratically as sentences get longer, whereas Mamba compresses long sequences to process them in near-linear time. This difference becomes significant in agent tasks where hundreds of pages of documents are processed at once.

Naver has also become the first Korean company to join NVIDIA’s Nemotron Coalition.

This is a position that contributes to open model development across pre-training, post-training, and reinforcement learning. There is a clear difference between companies that buy finished models and those that sit at the table to build the foundation models.

NemoClaw and the Seoul World Model: Two Paths for Sovereign AI

Building upon the infrastructure and models, the alliance extends in two directions: AI agents for internal corporate use and physical AI for roads and factories.

The bottleneck for agents is trust rather than performance.

The moment an agent reads internal files and references emails, corporations worry about internal data leakage. NVIDIA’s NemoClaw is a blueprint that keeps agents running within a sandbox, ensuring they cannot reach outside authorized folders and network paths. For instance, if the legal team tasks an agent with summarizing 1,000 contracts, the agent only accesses the designated folder and returns the results only to that team. Naver plans to launch an ‘AI Agent Platform’ based on this blueprint in Korea in the second half of 2026.

The bottleneck for physical AI is verification rather than data.

Robots on the road need physical intuition—how much a tire slips on a rainy road or how much a camera is blinded by backlight when exiting a tunnel—but these cannot be tested repeatedly on real roads. NVIDIA Cosmos is a world foundation model that has learned physical laws through vast amounts of video, and Naver is layering its accumulated urban Street View data and spatial modeling technology onto it to create the ‘Seoul World Model.’

Once this is achieved, scenarios like a pedestrian jumping into the road at the north end of Mapo Bridge during a downpour at dawn can be simulated thousands of times in a virtual Seoul. It is a testing ground for machines like the delivery robot ‘Rookie,’ which roams inside 1784, before they head out onto real streets.

The Structure of Naver Becoming a ‘Third Supplier’ in the Sovereign AI Market

Naver is special to NVIDIA because it is a rare company that possesses its own models, data centers, and cloud, while being neither a US Big Tech firm nor a Chinese company.

Looking at the history of communication infrastructure shows why this position is so valuable.

When the UK controlled the undersea telegraph cable network in the 19th century, global financial information and diplomatic dispatches had to pass through London. In the late 20th century, as US companies took over key internet nodes and cloud infrastructure, global digital traffic fell under the jurisdiction of US law. The cables of the AI era are data centers and language models, and currently, US Big Tech controls most of them.

There are countries uncomfortable with this structure. Europe is reluctant to store its citizens’ medical and judicial data on offshore servers; Middle Eastern oil-producing nations have realized the risks of technological dependency; and Southeast Asian countries lack indigenous AI, causing language data to flow entirely to foreign firms. What they all share is the need for ‘AI that runs within their own territory, using their own data, and is not bound by foreign laws’—in other words, Sovereign AI.

The problem is where to buy it. Buying from US Big Tech keeps you within the policies of a US firm, and buying from China creates a different kind of dependency. Naver occupies that gap. NVIDIA also introduced Naver as an alternative that provides infrastructure in Europe and the Middle East to meet each country’s regulations and data sovereignty standards.

There is also a track record.

In October 2023, Naver won a $100 million contract from the Saudi Ministry of Municipal and Rural Affairs and Housing to build a digital twin platform for five cities, including Riyadh.

For NVIDIA, the calculation is simple. It is a contradiction for a US company to go out and sell ‘Sovereign AI,’ but the story changes when a partner that has already built trust with local governments integrates NVIDIA hardware into local infrastructure. It may look like a sales strategy, but in reality, it is a geopolitical one.

However, a question remains that is not easily answered.

How much of the AI trained by Saudi Arabia on infrastructure provided by Naver is truly ‘Saudi’s’? Is sovereignty achieved simply by switching the supplier from US Big Tech to Naver, or is it an implementation process that continues until the local government gains the capacity to train models and operate infrastructure itself? The difference is not written in the contract.

When viewed alongside my analysis of Samsung and SK Hynix’s investment in Anthropic, it becomes clearer where Korean companies are trying to position themselves in the AI supply chain.

Why the Failure of MACH-1 Became a Condition for the Naver-NVIDIA Alliance

Naver becoming a partner rather than a competitor to NVIDIA is rooted in a past failure.

In 2022, Naver began designing an AI inference chip called ‘MACH-1’ together with Samsung Electronics.

The motivation was clear. NVIDIA GPUs were expensive and in short supply. The calculation was that using general-purpose LPDDR memory instead of expensive HBM, and embedding inference operations tailored to HyperCLOVA X into the circuit, would create price competitiveness. However, the collaboration between the two companies fizzled out, and Naver exited the chip business.

I believe timing, rather than technical capability, was the stumbling block.

During the 1-2 years it takes to design and mass-produce a chip, the model structure changed direction several times—toward Mixture of Experts, long-context, and Transformer-Mamba hybrids. Application-Specific Integrated Circuits (ASICs) have the algorithms they support hardened at the moment of design. They cannot be updated like software.

Then there is the wall called CUDA.

For over a decade, AI developers worldwide have built training frameworks and inference libraries on top of NVIDIA CUDA. By the time a new chip catches up to that ecosystem, the model generation has already shifted again.

Naver’s conclusion from this failure was to compete on platforms rather than silicon.

They use NVIDIA chips, manage infrastructure with DSX and NemoClaw, own the model with HyperCLOVA X, and control user touchpoints with FaceSign and Chzzk. A partner that does not fight over hardware but opens up the market using that hardware—this is exactly the type of partner NVIDIA welcomes most. The failure effectively created the conditions for the alliance.

Gigawatt AI Factories Do Not Yet Exist: Remaining Questions

There is always a gap between an announcement and implementation, and this alliance is no exception. 55MW in the first half of 2027 is only the first goal, and there is no guarantee that the schedule will be met.

There are more uncertainties. We do not yet know if the Sovereign AI market will grow as Naver expects, or if what governments really want is technological independence or just a slightly cheaper supplier. Whether the agent platform will pass the hurdles of corporate security teams and whether the intuition learned from the Seoul World Model will work on real roads are all yet to be verified.

What is certain is that Jensen Huang came to Seoul in person. He ate pork belly, watched a facial payment demo, and stood in front of a Chzzk live camera with Chairman Lee Hae-jin in the 1784 basement studio. For the CEO of the world’s most valuable semiconductor company to visit a partner’s headquarters and even do a broadcast is not mere ceremonial diplomacy.

When ZDNet Korea revisited that BBQ restaurant two weeks later, the Npay Connect terminal that was rushed in that day was still on the checkout counter. Whether the gigawatt contract will hold its place like that terminal will only be confirmed when the first 55 megawatts are turned on at Gak Sejong in the first half of 2027.