NAVER Scales Sovereign AI to Gigawatts via NVIDIA DSX
6
sources tracking this story
Key insights
- SK Telecom is training its A.X K1 model on NVIDIA Nemotron datasets under Korea's government Sovereign AI Foundation Model Project.
- NAVER's GAK Sejong roadmap runs 55 MW now, 200 MW by 2028, then gigawatt scale, targeting Europe and Middle East as export markets.
- Jensen Huang named SK Hynix NVIDIA's largest memory partner and signaled annual purchases already in the billions will expand further.
Why this matters
Jensen Huang's Seoul visit produced signed commitments across five South Korean conglomerates, moving Korea from chip supplier to co-builder of sovereign AI infrastructure. SK Telecom's framing of telecom networks as national AI backbone, combined with NAVER's gigawatt-scale buildout targeting European and Middle Eastern export markets, gives the deals a trade-policy dimension beyond bilateral commerce. SK Hynix's confirmed status as NVIDIA's largest memory partner, with annual purchases already in the billions and expanding, signals a customer-specific sourcing relationship replacing commodity procurement. Korea's government simultaneously procured 9,704 GPUs worth 2.08 trillion won to train its own sovereign foundation models, tying state AI policy directly to the same infrastructure wave.
Summary
NAVER is expanding sovereign AI infrastructure in South Korea, starting at 55 megawatts at the GAK Sejong data center and targeting gigawatt scale on NVIDIA's DSX platform.
One of only three companies worldwide to develop a hyperscale large language model, NAVER will fine-tune NVIDIA Nemotron 3 Ultra for its HyperCLOVA X models, making it the first Korean member of the NVIDIA Nemotron Coalition.
Essentially: NAVER + NVIDIA are co-building a sovereign AI factory covering compute, model training, and world-model simulation.
- Seoul World Model in development using NAVER's proprietary urban street-view data and NVIDIA Cosmos world foundation models.
- AI Agent Platform launch targeted for the second half of 2026.
- NVIDIA DSX MaxLPS software optimizes token throughput per megawatt to make factory economics work at scale.
NVIDIA's DSX platform enters Asia with NAVER as its first major Korean sovereign AI anchor.
Potential risks and opportunities
Risks
- Gigawatt-scale energy demand in Sejong could run into South Korean grid capacity constraints or regulatory approval delays, stalling NAVER's expansion well beyond the initial 55-megawatt footprint.
- NAVER's HyperCLOVA X model roadmap now depends on continued access to NVIDIA Nemotron 3 Ultra and NVIDIA GPU supply; any tightening of US export controls on advanced semiconductors would directly disrupt NAVER's training pipeline.
- Seoul World Model training on proprietary urban street-view data may attract regulatory scrutiny or legal challenge under South Korea's Personal Information Protection Act if individuals are identifiable in the training corpus.
Opportunities
- Korean enterprises seeking sovereign, low-latency alternatives to US hyperscaler cloud for fine-tuning and inference gain a credible option through NAVER Cloud Platform running on this new DSX infrastructure.
- NVIDIA's Nemotron Coalition gains its first Korean anchor in NAVER, creating a visible pathway for other Asian AI labs and enterprises to adopt Nemotron-based fine-tuning pipelines under the coalition framework.
- Energy efficiency and high-density compute vendors stand to benefit as NAVER scales from 55 megawatts toward gigawatt capacity, given that DSX MaxLPS token throughput per megawatt optimization will become a procurement criterion at every expansion stage.
What we don't know yet
- Timeline and capital commitment for the gigawatt phase beyond the initial 55 megawatts: the announcement names the scale target but gives no schedule or financing structure for that expansion.
- Whether NAVER Cloud Platform's external enterprise customers will have access to the new AI factory capacity, or whether it remains reserved for NAVER's own HyperCLOVA X model development and the Seoul World Model.
- What data governance and privacy framework governs the Seoul World Model's urban street-view training corpus under South Korea's Personal Information Protection Act, given that street-level imagery routinely captures individuals.
What others are reporting
-
Korea JoongAng Daily Read →
On-the-ground Korean reporting with direct executive quotes and NAVER's specific capacity roadmap (55MW to 200MW by 2028, then gigawatt scale), absent from Western coverage.
SK is our largest memory partner. We are expanding our partnership to include many new markets.
-
Seoul Economic Daily Read →
Frames this as SK Group competing to build Asia's largest AI infrastructure, with a pre-announcement Taiwan meeting between SK Chairman Chey and Huang establishing group-level strategy.
Telecommunications networks are evolving into national AI infrastructure.
-
NVIDIA Blog Read →
NVIDIA's own overview tying all five conglomerate partnerships (NAVER, SK, LG, Hyundai, Doosan) into a single sovereign AI narrative; the only source naming all five together.
You have everything that it takes. We are here to partner with you.
-
NVIDIA Blog Read →
Details the LG deal's physical AI data factory using NVIDIA Cosmos for synthetic training data, EXAONE sovereign model integration, and LG Energy Solution energy collaboration.
The AI factory will provide LG Group with accelerated computing infrastructure to train, simulate, validate and deploy AI-based applications.
-
Cryptopolitan Read →
Directly confirms SK Hynix as NVIDIA's largest memory partner with annual purchases in the billions expected to increase, quantifying the supply chain dependency the other coverage treats abstractly.
SK Hynix is Nvidia's largest memory supplier and will remain its largest partner...billions of dollars worth of products.
Originally reported by nvidianews.nvidia.com
Read the original article →Original headline: NVIDIA Signs AI Factory Deals With South Korea's NAVER, SK Telecom, and Doosan — Gigawatt-Scale Sovereign Infrastructure Coming by 2027