Naver builds AI computing cluster using 4,000 Nvidia GPUs

Naver has built South Korea’s largest AI computing cluster using 4,000 Nvidia B200 GPUs.

The company said the new “B200 4K cluster” will support the development of its proprietary foundation models and the broader application of AI technologies.

Naver said the cluster offers computing performance comparable to the world’s top 500 supercomputers.

The cluster is designed for large-scale parallel processing and high-speed communications, and Naver expects it will speed up AI model development by about 12x.

Naver also aims to expand work on omni models that process text, images, video, and audio simultaneously.

🔗 Source: Yonhap

🧠 Food for thought

Implications, context, and why it matters.

Nvidia DGX B200–based clusters and training vs inference for Naver’s 4,000‑GPU build

Naver says its build matches the “top 500 supercomputers.” Benchmarks for the DGX B200, an 8‑GPU Blackwell server, favor inference over training 1.
A typical 8‑GPU node like DGX B200 links eight B200 GPUs with fifth‑generation Nvidia NVLink at up to 1.8 TB/s per‑GPU link 2 3. A 4,000‑GPU build would use about 500 nodes that talk over InfiniBand or Ethernet 2.
Multi‑node setups add communication overhead for large, synchronous training. A DGX B200 posts about 144 petaFLOPS FP4 inference throughput, roughly double its 72 petaFLOPS FP8 training rate 1.
The company cites “about 12x” faster model development without splitting the phases. Blackwell uses FP4 for inference and FP8 for training, though Naver left the basis unclear 1.

Cloud AI operators can stand out by pairing B200 access with domain data

Naver Cloud Platform (the company’s cloud computing arm) runs 23 data centers in 10 countries with over 225 services 4. Smaller providers can rent B200 GPUs paired with datasets for healthcare imaging, financial time series, or manufacturing sensors for enterprise fine‑tuning.
The Daedong partnership in agriculture is one example. Proprietary Korean dialect voice data with farming logs 5 delivers defensible solutions that plain hardware cannot copy.
Operators without Naver’s scale can secure exclusive data deals with industry groups or government agencies before rolling out B200 access. Naver’s work with Korea Hydro & Nuclear Power (the state‑owned nuclear and hydroelectric utility) offers one template 6. That can win early ground in sectors where foundation models need localized training data.