Naver Builds Gigawatt-Scale AI Factories With NVIDIA

NVIDIA announced on June 7 that South Korea's Naver will deploy NVIDIA technology to build AI factories at gigawatt scale, starting with an expansion of Naver's GAK Sejong data center (NVIDIA; Reuters). NVIDIA's press release and Korea Herald state the buildout starts at 55 megawatts in the first half of 2027, expands to 100 megawatts later in 2027 and 200 megawatts by 2028, with gigawatt-scale capacity as the longer-term goal depending on power availability and procurement (NVIDIA; Korea Herald; DatacenterDynamics). The announcement links the deployment to NVIDIA's DSX AI factory platform and the vendor's full-stack AI software and models (NVIDIA).

Per NVIDIA's announcement, the project will use the DSX integrated platform to deliver an end-to-end AI factory blueprint for high-density, NVIDIA-accelerated computing optimized for training, post-training and inference (NVIDIA). Jensen Huang said, "Useful AI has arrived, and demand for AI factories is extraordinary," and NVIDIA and Naver chairman Haejin Lee framed the buildout as a sovereign foundation to serve Korean enterprises, manufacturers, government organizations and AI cloud customers (NVIDIA).

NVIDIA's release and Korea Herald report the collaboration spans infrastructure, models and physical-AI tooling: Naver is described as the first Korean company to join the NVIDIA Nemotron Coalition, plans HyperCLOVA X models built on NVIDIA Nemotron, and is developing a Seoul World Model from its urban street-view data on top of NVIDIA Cosmos world foundation models (NVIDIA; Korea Herald; Chosun). NVIDIA and Naver framed this as a sovereign-AI and ecosystem play to serve Korean industries and global AI cloud customers (NVIDIA; Korea Herald).

Companies building infrastructure at this scale are aiming to reduce token production costs, increase throughput for large-model training, and support latency- and compute-intensive physical-AI workloads such as robotics and agents. Observers frequently describe vendor-provided blueprints like DSX as a way to standardize deployment patterns and accelerate cloud-provider and enterprise rollouts.

Gigawatt-scale language for AI factories denotes a materially different class of infrastructure than traditional hyperscale data centers because it explicitly targets extreme power density, sustained GPU throughput, and co-design of software and hardware. For the regional market, announcements that tie a major GPU vendor's stack to a local cloud provider matter for latency, regulatory sovereignty, and local model optimization. Industry reporting notes power availability as a recurring constraint for such expansions; NVIDIA's regional executives also flagged power as a limiting factor in public comments (KoreaHerald).

Observers should track three measurable indicators: announced power procurement or utility agreements tied to GAK Sejong and any overseas expansions, the disclosed phased rollout (55 megawatts in the first half of 2027, 100 megawatts later in 2027, and 200 megawatts by 2028), and public technical disclosures about node counts, GPU families used, and integration of Nemotron or other model stacks. Reporting so far does not disclose specific GPU counts or total investment size (Korea Herald; Reuters; NVIDIA).