KONST
AI Training CalculatorContact Us
EN
AIDC site | Asia - South Korea

NVIDIA B300 - South Korea

KR-B300Live in December 2026

Location
South Korea
Availability
Live in December 2026
Cooling
DLC liquid cooling
Reference price
Contact for pricing
8Nodes
64GPU
288GB HBM3e per GPUMemory per GPUabout 18.4TB total
800G RoCEv2GPU Fabric
Specifications

Full Specifications

GPU, host, network, and site conditions in one place.

COMPUTECompute

GPU model
NVIDIA B300
GPU configuration
64 units8x NVIDIA B300 SXM (HGX B300) per node
Deployment scale
8 nodes64 GPU
Memory per GPU
288GB HBM3e per GPU
Total cluster memory
about 18.4TB total

HOSTHost

Processor
2x AMD EPYC 9555 (64C/128T) per node
System memory
2,304GB DDR5 per node
Storage
8x 3.84TB NVMe (30.72TB) per node

NETWORKNetwork

GPU interconnect
Ethernet RoCEv2 800GbEnon-blocking 1:1

FACILITYSite and Delivery

Location
South Korea
Cooling
DLC liquid cooling
Availability
Live in December 2026

WORKLOADSuitable Workloads

Purpose
AI TrainingInferenceHPC

Remaining specifications are provided when you request a quote. Source: GPU cluster specification sheet · 2026-11-16

Delivery schedule

From Signing to Go-Live

  1. 01Signing5 year term
  2. 02HandoverConfirmed with quote
  3. 03Go-liveLive in December 2026

Prices, specifications, and availability are confirmed in the signed contract and do not reflect live inventory.

Location

South Korea

South Korea
Region
Asia
Country / city
South Korea
Site code
KR-B300

Location is indicative, at country or city level.

Compare | Other sites

Sites With NVIDIA B300

Compare scale, network, cooling, and delivery time side by side.

SiteScaleGPU memoryGPU FabricCoolingAvailabilityStarting price, USD / hr / GPU
KR-B300South KoreaThis page8 nodes
64 GPU
288GB per GPU
about 18.4TB total
800G RoCEv2DLC liquid coolingLive in December 2026Contact for pricing
CZ-B300-200Czechia25 nodes
200 GPU
288GB per GPU
about 57.6TB total
On requestTo be confirmed16 weeks after signing$4.80 from
IN-B300India8 to 256 nodes
Configurable
288GB per GPU
2.3TB per node
InfiniBand NDR 400GAir Cooling8 to 10 weeks after signing$4.86 from
US-B300United States8 to 96 nodes
First batch 512 GPU
288GB per GPU
about 147TB total in first batch
InfiniBand NDR 400GAir Cooling10 to 14 weeks after signing$4.96 from
FI-B300Finland144 nodes
1,152 GPU
288GB per GPU
2.3TB per node
800G ConnectX-8DLC liquid coolingAbout 22 weeks after signing$5.30 from
US-IOWA-B300United States Iowa44 nodes
352 GPU
288GB per GPU
about 101.4TB total
800G ConnectX-8Air CoolingLive in mid-February 2027Contact for pricing
US-SEA-B300-32United States Seattle32 nodes
256 GPU
288GB per GPU
about 73.7TB total
800G ConnectX-8Air CoolingLive in February 2027Contact for pricing
US-SEA-B300-64United States Seattle64 nodes
512 GPU
288GB per GPU
about 147.5TB total
800G ConnectX-8Air CoolingLive in mid-March 2027Contact for pricing
FAQ | Common questions

Common Questions

How are the price and delivery schedule confirmed?
Prices and delivery schedules are for evaluation reference. Actual specifications, availability, and commercial terms are set by the signed contract. Prices, specifications, and availability are confirmed in the signed contract and do not reflect live inventory.
How does bare metal differ from GPU cloud?
Bare metal is a physical machine with no virtualization layer, so all performance is available, but you manage the operating system and cluster yourself. GPU cloud comes with the environment prepared and activates quickly, but resources are shared with other tenants. Training at long-term full load usually goes to bare metal.
Besides overseas GPU clouds, what should be considered when renting locally?
Three things: data residency, network latency, and service hours. When data must not leave the country, equipment location is a hard requirement. Local latency is lower than transoceanic links. The local team keeps the same working hours as the lessee, so fault resolution does not wait on time zones.
Can the GPU count be increased or reduced mid-term?
Adding GPUs usually means adding nodes, with pricing folded into the original contract and subject to capacity at the time. Removing GPUs is restricted by take-or-pay terms, so in practice the scale is usually adjusted at renewal.
Who owns the equipment, and what does the rental fee cover?
KONST keeps ownership of the equipment. The renter pays for the right to use it, which covers data center space, power, network, and operations. KONST bears the risks tied to purchase funding, residual equipment value, and generation changes.
From installation to ready for use, where does KONST's delivery scope end?
From the initial operating system installation, KONST sets up the machines, drivers, and base environment, and configures a separate network segment and a dedicated VPN. After that, the renter manages frameworks, scheduling, and data storage on its own.

Plan the Next Stage of Compute Together

Where Efficiency Leads.
Where AI Defines the Future.