Naver Cloud plans gigawatt-scale AI factories with NVIDIA technology as demand for AI services and physical AI rises

NVIDIA said Monday that NAVER, the South Korean internet conglomerate, plans to build AI factories at gigawatt scale using NVIDIA technology, aiming to meet growing global demand for AI services and for “physical AI” applications that connect AI systems to real-world environments. The report from Reuters did not provide a specific power capacity number beyond the gigawatt scale characterization, nor did it disclose investment size, partner names, or timelines for when the facilities would come online.

The gigawatt framing matters because it indicates a shift from traditional, smaller data center deployments to purpose-built compute campuses that require long lead times for planning and construction, as well as large power and cooling capacity. NVIDIA did not detail in the report which components of its stack NAVER will adopt, such as the exact GPU generation, server platforms, or data center reference architecture elements.

The announcement comes after NAVER Cloud, a major part of NAVER’s infrastructure and services business, described a separate but related partnership direction with NVIDIA on June 2. NAVER Cloud said it would work with NVIDIA to build “global AI factories” using full-stack capabilities spanning infrastructure, models, and services, and that more execution details would be unveiled in Korea during an upcoming meeting between NAVER founder and global investment officer Lee Haejin and NVIDIA CEO Jensen Huang.

In its June 2 statement, NAVER Cloud said the collaboration would involve technical work across the AI stack, including models and physical AI. NAVER Cloud plans to use NVIDIA’s open large language model, Nemotron 3 Ultra, to further advance its HyperCLOVA X system, and the two companies plan to jointly research optimization and core technologies for hyperscale language models. NAVER Cloud also highlighted its “Seoul World Model,” which it said was created using NVIDIA Cosmos and trained on 1.2 million panoramic images captured across Seoul, aiming to recreate real-world road and spatial environments. NAVER Cloud CEO Kim Yu-won characterized the effort as moving beyond a model-centric approach toward an inference-centric “AI factory era” where large infrastructure must be operated reliably.

NVIDIA’s broader “AI factory” concept emphasizes that such deployments are not just chips, but integrated systems for designing, building, and running large-scale AI workloads. In a prior NVIDIA blog post about its Omniverse DSX Blueprint, the company described a blueprint that unites design, simulation, and operations, including codesign of facility elements like power and cooling within a digital twin environment to validate performance before construction.

For South Korea, the push also reflects the region’s wider competition to secure advanced AI infrastructure and to expand the use of AI beyond chat-style applications into industrial settings. NAVER Cloud said it wants to provide AI environments tailored to customer workloads while supporting the development of regionally optimized sovereign AI models through local and government partnerships, positioning its infrastructure platform as a durable base for both general AI services and physical AI use cases.

Still, key details remain unconfirmed publicly. The Reuters report did not specify the number of gigawatts, the location(s) of the facilities in South Korea, the expected commissioning dates, the hardware configuration, or the commercial structure of the relationship between NVIDIA and NAVER. NAVER’s June 2 statement likewise did not quantify gigawatt-scale capacity, focusing instead on the partnership’s full-stack scope and near-term model and physical AI collaboration.

What to watch next is whether the Korea meeting between Lee Haejin and Jensen Huang leads to more concrete disclosures, such as phased deployment schedules, power and cooling plans, and the specific NVIDIA platform elements NAVER Cloud will standardize on to scale inference and physical AI workloads. The disclosures could also show how the companies intend to manage reliability and cost at data-center scale, which is often the differentiator once projects move from announcements to construction and operations.