Naver Cloud Runs 35,000 AMD EPYC Units as AMD Pushes Hybrid AI

Naver Cloud disclosed it deployed about 35,000 AMD EPYC processors last year, up from 13,000 in 2023, at AMD's Advancing AI World Tour Seoul 2026 event on Nov. 4 in Samseong-dong. The Korean operator runs about 250,000 servers across 13 domestic data centers. AMD framed the event around "hybrid AI," a strategy that combines cloud and on-device computing for agentic AI workloads.

What did AMD announce?

The company introduced three product lines: the 6th-Gen EPYC (Venice) server processors, the Ryzen AI Max+ Pro 400 series, and the Instinct MI350P GPU. Shipments of the 6th-Gen EPYC began in the third quarter. Meta was an initial customer, and AMD plans to expand supply from the fourth quarter.

The EPYC line splits into four families:

• SP7: high-performance computing and multithreaded processing
• SP8: enterprise, balancing performance and power efficiency
• VeniceX: HPC and AI training
• Verano: cloud service providers (companies selling compute on demand) and Helios systems

Ritesh Rohit Patel, Asia-Pacific director of AMD's data center division, said: "Nobody can say exactly what workloads will emerge three years from now. We need to design products that can adapt to a changing environment." He noted that generative AI chatbots drew GPU-centric attention in 2022, but agentic AI requires CPUs and GPUs to work in tandem across stages from user request to execution, alongside networking, storage, memory, and software. That is a portfolio pitch, not a benchmark.

Do the consolidation claims hold?

Patel said consolidating 1,000 Intel Xeon-based two-socket servers into 82 EPYC servers reduces power, space, and management costs. AMD did not publish the test conditions. For existing data centers, AMD positions the Instinct MI350P as an add-in: the chip scales up inference workloads through PCI Express slots (the standard expansion interface) rather than wholesale infrastructure replacement.

Naver Cloud offered its own numbers. Park Young-hee, a director at the company, said Intel Xeon cost burdens and limited core-count expansion drove the switch. The company's internal assessment found EPYC platforms cut rack space by about 20-25% versus Intel, freed space that hosted additional servers, and consumed about 700W versus about 800W for the Intel platform. Naver Cloud plans to bring the 144-core EPYC 9005 series into its public cloud in the second half of this year. Vendor-run comparisons deserve scrutiny; these figures came from the customer, which strengthens them.

Per Naver Cloud's disclosed purchases: 13,000 units in 2023, 18,000 in 2024, and 36,000 in 2025. The 35,000 figure cited on stage refers to last year's adoption; the table shows cumulative annual purchases climbing steadily.

Why local AI?

AMD also pushed on-device AI, citing security, privacy, lower latency, and usability where networks are poor. Alexey Navolokin, AMD's Asia-Pacific lead, said AI PCs save the average office worker about nine hours per week and frequent business travelers up to 300 hours per year. Those productivity figures rest on AMD's own accounting.

He cited Goldman Sachs projections that token consumption will grow 24-fold by 2030, driving up data center power demand. His counter: performance once requiring 120-billion-parameter models is now achievable with 9-billion-parameter models that run on laptops. The Ryzen AI Max+ Pro 400 (Gorgon Halo), launched at the end of September, runs models up to 300 billion parameters at 4-bit quantization (compressing model weights to 4 bits each) using up to 192GB of unified memory. Claimed capability, not independent verification.

Clebby, a Korean AI solution company, showed "Uniform," a document engine on the Ryzen AI Max+ 395. Images and tables route to Cova, its in-house OCR engine of about 10 billion parameters. CEO Lee Hwan-ho said public institutions cannot move documents to external clouds, yet self-built servers carry power, cooling, and space burdens. Clebby is running proofs of concept with offices of education and design firms, and plans an on-premise appliance pairing Ryzen AI Max hardware with Uniform software.

What remains unresolved?

The event read as an AMD bid to recover data center GPU share lost to NVIDIA while pressing Intel in server CPUs and AI PCs. The Naver Cloud case shows the price-and-efficiency offensive works with Korean operators. If agentic AI spreads, system integration across CPU, GPU, and memory, plus the software ecosystem, will decide the outcome. Neither variable was demonstrated on stage.