Why Did Jensen Huang Eat Samgyeopsal at a Samgyeopsal Restaurant?
— The Geopolitics Behind the NVIDIA-Naver Gigawatt Alliance
On the evening of June 5, 2026, samgyeopsal was being grilled at a restaurant in Hongdae, Mapo-gu, Seoul.
On one side of the table sat Lee Hae-jin, Naver’s Global Investment Officer (GIO), and across from him, Jensen Huang, CEO of the company selling the world’s most expensive semiconductors, raised a soju glass.
The restaurant was neither a Michelin-starred establishment nor a banquet set up by the Ministry of Foreign Affairs. It was a smoky barbecue joint.
Three days later, Jensen Huang appeared at Naver’s second headquarters, ‘1784’, located in Bundang-gu, Seongnam-si, Gyeonggi-do.
He walked through corridors where robots navigated autonomously, witnessed a facial recognition payment system process transactions without wallets, and waved to developers worldwide alongside GIO Lee Hae-jin in front of the cameras of the live streaming platform ‘Chijijik’.
These scenes, stitched together, seem awkward.
Why did the head of the world’s largest AI infrastructure company visit the headquarters of a Korean portal company? Why was the announcement venue a barbecue restaurant instead of a conference hall?
The answer isn’t on the surface. Looking at the scale of the deal reveals a different picture.
What the two companies agreed upon is not a simple GPU supply contract.
It’s a plan to jointly build ultra-large AI infrastructure worldwide, starting with 55 megawatts (MW) by the first half of 2027, reaching 200 megawatts in 2028, and eventually a gigawatt (GW) in the long term.
One gigawatt is approximately four times the maximum design capacity (270 megawatts) of ‘Gak Sejong’, Naver’s hyper-scale data center, the largest in Asia. Data centers of that scale do not yet exist.
This article traces what lies between a plate of samgyeopsal and the gigawatt contract.
1. This Is Not a Machine Transaction — The History of Data Centers Becoming Factories
In the early 2010s, data centers were refrigerated warehouses housing web servers and email servers.
They were passive infrastructure where servers responded to user requests by sending back HTML. Power consumption was quiet, and cooling was sufficient with air conditioners.
The point at which that completely changed can be pinpointed to a single date: November 30, 2022. The launch date of ChatGPT.
Since the deployment of large language models (LLMs) to actual users, the identity of data centers has been fundamentally transformed.
They are now factories that take raw text data as input, process hundreds of billions of floating-point operations around the clock, and output intelligence in units of ’tokens’.
The concept Jensen Huang calls an ‘AI Factory’ encapsulates this transformation in a single word.
Data centers that have become factories come with factory problems. The most brutal constraint is power.
A single rack of existing servers consumes about 10 kilowatts (kW).
A single rack equipped with NVIDIA’s latest Blackwell accelerator cluster exceeds 120 kilowatts.
If more than 10 times the power is supplied to the same space, the heat generation also increases tenfold. Existing air conditioning methods cannot handle that heat. The servers melt.
NVIDIA solved this problem with software. Beyond selling GPUs, they created a platform that orchestrates the power grid itself through code.
That platform is called DSX (Data Center Systems eXperience).
The issue is that DSX cannot be integrated into just any data center.
To power gigawatt-scale AI factories, not only software control capabilities but also the physical infrastructure to accommodate them must precede.
Cooling systems capable of handling high-density heat, operating systems that respond instantly to power fluctuations, and decades of accumulated data center design experience.
Naver possesses these. This is half the reason Jensen Huang visited Naver in person.
2. From the Samgyeopsal Table to the Facial Recognition Terminal — The Tangible Basis of Trust
The scenes Jensen Huang witnessed at the Hongdae restaurant were not something one could hear in a technical briefing.
Lee Hae-jin, the GIO, did not take out his wallet.
He presented his face to Naver’s offline payment terminal ‘Connect’ installed at the checkout counter. The terminal’s built-in camera collected his biometric data, a local neural network performed verification, and the payment was completed. The entire process took just over one second.
The importance of this scene lies not in the payment technology itself. What’s important is the place where it worked.
A regular restaurant in the Hongdae commercial district of Mapo-gu, Seoul. Not a smart city pilot project, but a real store operated by a small business owner.
Jensen Huang saw firsthand the reality of ‘practical AI’, where AI goes beyond generating text on a monitor and is fully integrated with offline payment infrastructure.
It was evidence that there was no gap between what Naver claimed and what Naver implemented.
For a technology company, such evidence is more persuasive than any press release.
Three days later, during the visit to 1784, the two leaders went live together on ‘Vision Stage’, a virtual studio on the live streaming platform Chijijik.
It was unusual for the CEO of a global big tech company to appear on a broadcast of a Korean local streaming platform.
This scene, deviating from the rigid grammar of corporate merger reporting, symbolized the announcement of emotional trust between the two companies to the public.
Strategic technology alliances do not begin on spreadsheets.
They begin with the certainty that the other party is actually doing what they claim. Jensen Huang first gained that certainty at the samgyeopsal restaurant.
3. How to Write Power with Code — An Anatomy of NVIDIA DSX
If the physical limitation of an AI factory is power, DSX is the system that pushes those limits with software.
By dissecting this platform, composed of four layers, in order, we can see how a gigawatt factory operates.
DSX OS — This is the factory’s nervous system. In a multi-tenant cloud environment where thousands of GPU racks operate simultaneously, it automatically monitors the hardware lifecycle and executes immediate software remediation when failures occur. This means human engineers don’t need to restart servers in front of racks at 3 AM.
DSX MaxLPS — This engine increases GPU density by up to 40% within a fixed power budget. Its operating principle is as follows: AI computation workloads are not uniform. Some GPUs are running at maximum capacity for training computations, while GPUs in adjacent racks are in a standby state for inference. This creates stranded power. MaxLPS detects and recovers this idle power in real-time, redistributing it to active GPUs. This allows more GPUs to be operated within a fixed megawatt budget, maximizing the number of tokens generated per megawatt.
DSX Flex — This synchronizes power grid signals from power plants directly with the data center’s computation scheduler. During peak power hours or when renewable energy supply is momentarily cut off, it automatically adjusts the deep learning training queue in nanoseconds to alleviate local power grid overload. This marks the moment the data center transitions from being a passive consumer of the power grid to an active balancing mechanism.
DSX Exchange — This is the hub that connects the IT world (software computation signals) and the OT world (cooling pumps, HVAC valves, power buses) using the high-speed MQTT protocol. If a GPU’s temperature exceeds a critical threshold, the software increases the cooling pump RPM in real-time. This creates a loop where the physical infrastructure responds to computational load.
Naver Cloud has agreed to integrate these four layers across its entire cloud infrastructure.
This is not a simple GPU purchase but a decision to redesign the entire infrastructure operation method on the DSX architecture.
4. Why Gak Sejong is Alive — The Accumulation of 30 Years of Cooling Engineering
If DSX controls power with software,
handling the heat generated by that power is Naver’s domain. This is the second core asset Naver brings to this alliance.
In 2013, Naver completed its first data center, ‘Gak Chuncheon’, in Chuncheon, Gangwon-do.
What was unique about this building was its design, which used almost no air conditioning.
Utilizing its location on a mountainside, it adopted an AMU (Air Management Unit) system that directly drew in external cool air into the server rooms.
This cooled the servers for about 90% of the year without running air conditioners.
The Power Usage Effectiveness (PUE) was 1.09, among the top globally.
‘Gak Sejong’, completed in 2023, took that technology to the next level.
The key was the elimination of air transport ducts.
Traditional data centers push cooling air through large ducts to every corner of the server room. Ducts create friction resistance, requiring powerful fans to overcome it, and these fans consume power.
NAMU III at Gak Sejong eliminated ducts and instead used densely arranged BLDC (Brushless DC) fans in a grid structure. This moved more air with less power and without friction.
The problem is that accelerators from the Blackwell and Vera Rubin generations are pushing the physical limits of air cooling.
Handling heat exceeding 120 kW per rack with air alone leads to a explosive increase in energy costs.
Naver is preparing solutions in two directions.
One is Direct Liquid Cooling (DLC): Cooling plates are directly attached to the GPU chip surface to transfer heat via water cooling.
The other is immersion cooling: The entire server is submerged in a special, non-conductive oil.
A hybrid system combining these two methods with NAMU III’s air cooling is scheduled for completion in the server rooms in 2027.
What’s important here is that this cooling system was co-designed at the hardware level with NVIDIA’s next-generation liquid cooling specifications.
Software (DSX) and physics (Gak Sejong cooling) are designed to interact within a single architecture.
Artificial intelligence reads GPU temperature data and controls fan speed and coolant flow rate in real-time.
This is the difference between a simple GPU buyer and a strategic alliance partner.
The former buys racks, while the latter jointly designs the infrastructure that allows those racks to exist.
5. Why Activate Only 55 Billion? The Paradox of Nemotron 3 Ultra
At the top layer of the hardware alliance lies an alliance of model architectures.
Typical large language models activate their entire weight parameters when predicting a single word.
For a model with 500 billion parameters, generating a single word like ’the’ involves 500 billion operations. This is wasteful.
However, it was an unavoidable structural waste in existing architectures.
Nemotron 3 Ultra solves this differently.
It categorizes the entire 550 billion (550B) parameters into functional ‘Expert’ groups and selectively activates only 55 billion (55B) weights based on the intent and characteristics of the input sentence.
A dynamic router called LatentMoE handles this selection in nanoseconds.
It uses only 10% of the total capacity while achieving equivalent or better quality than models using the entire capacity.
Adding the Mamba layer to this.
Existing Transformer architectures require computational complexity proportional to the square of the input length when processing long sentences (O(N²)).
Processing 1 million tokens was theoretically close to impossible.
Mamba compresses the sequence mathematically and processes it in linear time (O(N)).
This makes it possible to input a 2,000-page report in a single inference loop.
The results are summarized in two numbers: 5x improvement in inference speed and 30% reduction in computation cost.
Naver is leveraging this architecture to fine-tune Korean and East Asian legal data, enhancing the foundation of HyperCLOVA X.
This is not simply adopting foreign models but establishing a collaborative relationship from the stage of fundamental architecture design.
Naver has joined NVIDIA’s latest model alliance (Nemotron Coalition) as the first Korean company.
This signifies the difference between a technology importer and a technology producer.
6. When Agents Enter the Enterprise — NemoClaw’s Security Philosophy
If Nemotron 3 Ultra is the brain,
NemoClaw is the neural shield that ensures that brain operates safely within the enterprise environment.
The biggest barrier to AI agents being deployed in actual work is not performance, but trust.
When an agent accesses internal file systems, reads accounting data, or refers to company emails, the fear of leakage of corporate intellectual property prevents adoption.
In reality, this fear is not unfounded.
Existing AI agents often did not control the scope of their permissions once granted.
NemoClaw blocks this problem at the hardware level, not the software level.
Agents always run only within the ‘OpenShell’ sandbox.
Access to directories, IP ports, and network paths that the user has not explicitly allowed is physically blocked at the kernel level (Capability Drops).
Access is not denied; the path simply doesn’t exist for the agent.
All outbound packets are filtered in real-time, and if anomalies are detected, the agent kernel immediately self-destructs (Self-healing shutdown).
This structure creates a tangible change:
The legal team can delegate the summarization of 1,000 contracts to an agent.
The agent can only read the specified contract folder, cannot see beyond it, and the summary results are returned only to the legal team.
It is structurally impossible for source code to be transmitted to an external server or for a ping to be sent to a competitor’s server.
Naver and NVIDIA plan to officially launch this platform in Korea in the latter half of this year.
It will be in a form that allows corporate users to deploy NemoClaw-based agents directly into their own environments.
Areas previously blocked by security concerns, such as autonomous coding assistance tools, financial transaction analysis agents, and automated internal legal review.
The speed at which the agent era enters enterprises may need to be remeasured from the launch of NemoClaw.
7. How to Teach a City — Cosmos and the Seoul World Model
AI that ventures into the physical world requires a different kind of intelligence.
Language models that generate text need grammar and meaning. Robots that drive on roads need laws of physics.
How much tires slip on asphalt in the rain, how long it takes for cameras to adapt after exiting a tunnel, how quickly brakes need to react to the sudden appearance of a pedestrian. This comes from experience, not data, and to accumulate experience, a reality where accidents can occur is needed.
NVIDIA’s Cosmos 3 provides a virtual replica of that reality.
Cosmos 3 is a foundation model that learns the physical laws of the real world by directly feeding hundreds of millions of hours of videos of global road environments, object handling, and physical dynamics into a neural network.
When a user inputs a text description or sensor data, Cosmos 3 calculates acceleration, deformation upon collision, light reflection, and tire-road friction coefficients physically accurately to generate a 3D virtual video.
Traditional simulations required engineers to directly inject physical formulas into code. Cosmos 3 learned this from data itself.
Naver Cloud has combined this engine with Seoul’s spatial data.
Field data accumulated since the early 2000s operating Naver Maps, over 1.2 million panoramic images captured by vehicles equipped with LiDAR sensors driving throughout Seoul, and precise terrain data with road width errors measured in centimeters. Stacking this onto Cosmos 3’s physical simulation is the ‘Seoul World Model’.
What this model enables is explained through specific scenarios.
2 AM, north of Mapo Bridge. It’s raining heavily.
The streetlights are off. A pedestrian suddenly runs into the road.
How should the autonomous logistics robot ‘ROOKIE’ handle this situation? It’s impossible to repeatedly experiment with this scenario on actual roads.
Within the Seoul World Model, this situation can be replicated infinitely to fine-tune ROOKIE’s response algorithm.
The moment the camera is overexposed by sudden backlighting at a tunnel exit,
How to maintain driving solely through spatial memory without sensors.
How to accurately calculate the friction coefficient on a road surface contaminated with oil.
All of this is repeatedly simulated thousands of times in the virtual Seoul before being implanted into ROOKIE’s control algorithm.
The bottleneck for physical AI is not data but verification. The Seoul World Model makes that verification infinitely repeatable in a virtual environment.
8. The War for Data Sovereignty — The Geopolitics of Sovereign AI
In the history of technology, those who controlled infrastructure controlled the market.
In the 19th century, when Britain controlled the undersea telegraph cable network, global financial information and diplomatic communications had to pass through London.
In the 20th century, when the US controlled core internet nodes and cloud platforms, global digital traffic came under the jurisdiction of US law.
The infrastructure of the AI era is data centers and language models. And these are dominated by Google, Microsoft, and Amazon.
There are countries uncomfortable with this structure.
The EU, which legally prohibits storing its citizens’ medical data, judicial records, and defense documents on servers under US jurisdiction.
Oil-producing countries in the Middle East, who keenly felt the risks of technological dependence after US sanctions on Iran.
Southeast Asian countries whose citizens’ language data flows entirely to foreign companies because they lack AI models that operate in their own languages.
These countries share the need for what is called ‘Sovereign AI’.
AI infrastructure that operates within their own territory with their own data, not subject to foreign laws.
The problem is where to buy it.
It cannot be bought from US big tech. They always provide services only within their own laws and policy guidelines.
It also cannot be bought from China. Political risks create dependence in a different direction.
Naver exists in that vacuum.
Naver possesses a full-stack of completeness achieved by only a few globally.
Proprietary large language model (HyperCLOVA X),
Asia’s largest low-power hyper-scale data center (Gak Sejong) that powers it,
Integrated cloud infrastructure that guarantees data sovereignty.
And most importantly, a track record of delivering these to actual government clients.
The $100 million contract to build a digital twin nation with the Ministry of Municipal and Rural Affairs of Saudi Arabia is that track record.
3D scanning of key cities including Riyadh, operation of Arabic-specific AI models, and complete separation of data within domestic servers.
This contract proves that Naver can communicate in the language of demand, not the logic of technology exporters.
From NVIDIA’s perspective, Naver’s value lies precisely here.
Jensen Huang personally going to persuade Middle Eastern governments could be counterproductive.
Because it’s a contradiction for US big tech to sell ‘Sovereign AI’. However, it’s different through Naver.
A neutral partner that has already built trust with local governments integrates NVIDIA’s hardware into local infrastructure.
This is not a sales strategy but a geopolitical strategy.
9. What Failure Taught Us — Mach-1 and the Logic of Platform Transition
In 2022, Naver began designing its own AI acceleration chip, ‘MACH-1’, with Samsung Electronics.
The motivation was clear. NVIDIA GPUs were expensive and supply was insufficient.
By using general-purpose LPDDR memory instead of HBM memory to reduce costs, and by directly implementing inference computations specialized for HyperCLOVA X in the chip circuitry, they could secure price competitiveness. About 40 key researchers were involved.
MACH-1 faltered not due to hardware technology issues, but due to timing.
In 2024, by the time the silicon design of MACH-1 was finalized, AI model architectures had moved in a completely different direction.
Mixture-of-Experts (MoE), millions of tokens of context, Transformer-Mamba composite structures.
These did not align with the algorithmic structures etched into MACH-1’s circuitry for optimization.
ASICs have fixed supported algorithms at the time of design. They cannot be updated like software.
Then there was the CUDA wall. AI researchers and developers worldwide have spent decades building optimization tools, learning frameworks, and inference libraries on NVIDIA’s CUDA. The time required for a new accelerator chip to replicate that ecosystem from scratch is years. During those years, AI models will evolve by another generation.
The innovation cycle for AI model architectures is 2-4 months.
ASIC design to mass production takes at least 18-24 months.
As long as these two numbers coexist, ASICs specialized for specific algorithms will be outdated by the time they are mass-produced.
When Samsung Electronics mentioned Naver as a joint partner at the announcement of the MACH-2 roadmap, the collaboration between the two companies effectively ended.
Naver changed its strategic direction from this failure.
Instead of hardware manufacturing, they shifted to securing control of the software stack.
Platform competition instead of silicon competition: Controlling infrastructure with DSX and NemoClaw on NVIDIA GPUs, controlling models with HyperCLOVA X, and controlling endpoints with FaceSign and Chijijik.
This platform-centric approach paradoxically makes Naver an ideal partner for NVIDIA.
Not a competitor vying for hardware, but a companion that maximizes the use of that hardware to jointly pioneer the market.
Failure created the conditions for the alliance.
10. The Turning Point: The Question Gigawatts Ask
What has been described so far — DSX, Gak Sejong, Nemotron, NemoClaw, Seoul World Model, Sovereign AI — each appears as a separate piece of technology news.
However, these are multiple branches of an answer to a single question:
When artificial intelligence becomes the core infrastructure of human society, who controls that infrastructure?
At this very moment, that control is concentrated in the hands of a few US companies.
Training data, model weights, cloud services, hardware supply chains. All of this is under US laws and US corporate policies. When governments worldwide become dependent on them, they cede a part of their sovereignty.
The Gigawatt AI Factory Alliance is an attempt to change that structure.
NVIDIA supplies the hardware, and Naver integrates and operates it within a network of trust with non-English speaking governments.
Europe gets GDPR-compliant Sovereign AI infrastructure, the Middle East gets locally operating AI in Arabic, and Southeast Asia processes its data on its own servers.
This is packaged in the language of technological democratization, but it is actually a geopolitical realignment of the AI infrastructure market.
In this realignment, Naver’s position is unique.
A third-party provider that is neither US big tech nor Chinese tech. One of the very few companies with its own models, operating its own infrastructure, and able to approach governments worldwide from a politically neutral stance.
This raises a question.
Is Naver supplying Sovereign AI a true transfer of technological sovereignty, or a more sophisticated form of technological dependence?
This question is not easily answered. How ‘Saudi’ is an AI model trained on the infrastructure Naver built?
Even if Gak Sejong is in South Korea and HyperCLOVA X is trained in Korean, can we say that the Riyadh government’s data is under complete sovereignty?
This question is important not because of technology, but because of history.
For the word ‘sovereign’ to have meaning, there must be a path to resolving dependence.
Is it sovereignty to simply change the supplier from US big tech to Naver, or is it the transitional process until local governments have the capacity to train models and operate infrastructure themselves?
The difference is not written in the contract.
11. After the Transition — How Technology Melts into Daily Life
Let’s return to the scene where Jensen Huang witnessed facial recognition payment at the Hongdae restaurant.
We have already explained the technical aspects of that scene.
However, the meaning of that scene remains even after the technical explanation.
The statement that AI will change the world has been repeated for decades.
The moment that statement truly holds meaning is when technology is fully integrated into human behavior and becomes unconscious.
That moment is when presenting one’s face at a payment terminal becomes as natural as taking out a wallet.
If FaceSign implemented that moment in a single restaurant, the Gigawatt AI Factory is an attempt to expand that moment to the daily lives of billions of people.
The scale is different, but the direction is the same. The direction in which technology is erased from human consciousness. The invisibility of machines is a signal of complete integration.
ROOKIE, repeatedly running virtual night downpours in the Seoul World Model, will someday deliver goods on the real roads of Seoul at 3 AM without humans.
An agent reviewing contracts within the NemoClaw sandbox will someday handle half of the drafting work previously done by lawyers.
While Gak Sejong’s cooling system manages GPU heat,
An AI ethics legislator in the European Parliament, thousands of kilometers away, will be reviewing draft legislation on domestic Sovereign AI infrastructure.
Few people know these are connected.
Few people know that a plate of samgyeopsal is linked to a gigawatt contract.
12. The Final Question
There is always a gap between the declaration of a technology alliance and its actual implementation.
The 1-gigawatt AI factory does not yet exist.
The target is 55 megawatts in the first half of 2027, and there is no guarantee that even that will be completed as planned.
It is also uncertain whether the Sovereign AI market will develop in the direction Naver anticipates.
It is unclear whether countries truly desire technological self-reliance or simply a cheaper supplier.
The list of uncertainties can be longer.
However, the fact that Jensen Huang came to Seoul in person is not uncertain.
He ate samgyeopsal, witnessed facial recognition payment, saw robots walking inside buildings, and appeared on a live broadcast.
The CEO of the world’s most valuable semiconductor company visiting the headquarters of a partner company in person is not routine diplomacy.
It is a signal that NVIDIA has judged Naver not as a mere customer, but as a partner in a shared destiny.
Whether that judgment is correct can be known after 2027.
Until then, there is one question we must ask.
In a world where AI infrastructure, along with power grids, becomes core national infrastructure, how much attention are we paying to who designs and who operates it?
While the gigawatt factories are being quietly built, most people are looking at something else.
References
- NVIDIA DSX Platform Technical Playbook and Reference Architecture Guide for AI Factories — Warren Barkley, Chris Alexiuk (NVIDIA Infrastructure Engineering)
- Hybrid Transformer-Mamba MoE Architecture: Nemotron 3 Ultra Technical Report — NVIDIA AI Research Lab (NVIDIA Cosmos Lab)
- Sovereign AI Infrastructure Expansion using NVIDIA DSX Platform on GAK Sejong — NAVER Cloud Technical Whitepaper (Noh Sang-min, Kim Yu-won)
- Physical AI Simulation and Spatial Intelligence: Seoul World Model Development conditioned on NVIDIA Cosmos 3 — NAVER Cloud Spatial Intelligence Center (Lee Hae-jin, Kim Yu-won)
- NemoClaw and OpenShell Sandbox Hardening User Guide for Autonomous Agent Deployment — NVIDIA Product Security Team
- Study on Gak Sejong Eco-friendly Hybrid Air Conditioning System and NAMU III Energy Efficiency Maximization — Naver Data Center Integration Center (Noh Sang-min)
- FaceSign Offline Facial Recognition Payment Terminal ‘Connect’ Field Test Report — Naver Financial Technology Development Department (Lee Hae-jin, Choi Soo-yeon)
- Lessons Learned from Samsung Mach-1 LPDDR-Only ASIC Design and Analysis of CUDA Software Dependency — Journal of the Korean AI Semiconductor Society (Jung Chae-yoon, Lee Ja-kyung)
- Sovereign AI Deployment Patterns: Middle East and European Market Analysis 2025–2026 — NVIDIA Global Partnership Report
- NAVER Saudi Arabia Ministry of Housing Digital Twin Contract: Technical Architecture and Outcomes — NAVER Global Geo-Intelligence Group
- Power Usage Effectiveness (PUE) Benchmarking: Hyperscale Data Centers in Asia-Pacific — Uptime Institute Asia Pacific Report 2025
- Agentic AI: Enterprise Deployment Security Requirements and Containment Architectures — Gartner Technology Research (2026)
- Mixture-of-Experts Models: Scaling Laws, Routing Efficiency, and Hardware Implications — DeepMind / Google Research (2025)
- Geopolitics of AI Infrastructure: Sovereign Compute and the New Technology Multipolarity — MIT Technology Review Policy Series (2026)
- NVIDIA Vera Rubin Platform: Compute Architecture and Memory Subsystem Design — NVIDIA Technical Disclosure (2025)