NAVER to scale sovereign AI infrastructure with NVIDIA DSX, starting at 55 MW

NAVER and NVIDIA said NAVER will expand its “sovereign” AI infrastructure by building NVIDIA DSX-based AI factories, beginning with capacity growth at 55 megawatts and with plans to move toward gigawatt-scale build-outs. The announcement frames the project as infrastructure for production-grade AI, including AI agents and physical AI, rather than one-off model experiments.

NVIDIA described the initiative as a response to surging demand for useful, deployable AI. NVIDIA’s CEO Jensen Huang said NAVER is building AI factory infrastructure intended to serve its companies, developers, and industries, adding that DSX helps Korea scale sovereign intelligence for what NVIDIA calls the agentic era. NAVER’s chairman and founder, Haejin Lee, said NAVER is aiming to take customers from AI experimentation to production-scale AI factories powered by models, agents, and real-world services.

Central to the collaboration is NVIDIA DSX, which NVIDIA positions as an end-to-end, codesigned stack for AI factories that spans chips, systems, software, facilities, and partner technologies. NVIDIA said DSX is designed to reduce “token cost,” a practical metric that ties how many AI tokens (units of text) can be produced per cost or per unit of power, while also speeding “time to first production,” the point at which infrastructure becomes usable for real workloads.

NAVER will pair DSX with its data-center and GPU-cluster operating expertise. NVIDIA said DSX includes DSX MaxLPS software, which is intended to achieve the lowest token cost by maximizing token throughput per megawatt, and DSX OS, which NVIDIA describes as an operating layer for lifecycle management, consistent runtime operations, automated health and resiliency features, and multi-tenant AI factory management.

Beyond infrastructure, the companies tied the build-out to NAVER’s model development and agent plans. NAVER said it is advancing next-generation HyperCLOVA X models by fine-tuning NVIDIA’s Nemotron 3 Ultra open model with NAVER’s proprietary data and training expertise. NVIDIA also said NAVER is the first Korean company to participate in the NVIDIA Nemotron Coalition, contributing to open model development across pretraining, post-training, and reinforcement learning. NAVER said it plans to launch an AI Agent Platform in Korea in the second half of the year, powered by NVIDIA NemoClaw blueprints.

NAVER also said it is developing a Seoul World Model using its proprietary urban street-view and spatial modeling technology, building it on NVIDIA Cosmos world foundation models. NVIDIA described AI factories as increasingly critical for training, post-training, and inference, emphasizing that this infrastructure approach supports high-density, reliable operations designed to serve not only models but also the downstream systems that integrate those models into services.

What is not disclosed in the announcement are the specific hardware configurations (for example, which exact NVIDIA accelerated systems will be installed), the full multi-year project timeline beyond the starting 55 MW step, and any commercial terms or pricing. The release uses forward-looking language about scale, performance, availability, and expected benefits, without providing financial impact or near-term revenue indicators. The next checkpoints will be visible as NAVER expands the GAK Sejong facility, then rolls out production services tied to agentic and physical-AI ambitions, and as additional DSX-backed deployments are announced.