Faster Distributed Training
Powered by a non-blocking InfiniBand fabric for maximum GPU utilization
Accelerated AI Infrastructure
High-performance GPU infrastructure built for AI training, inference, and large-scale machine learning workloads.
GPUs online today
16,384
InfiniBand per node
3.2 Tb/s
Uptime SLA
99.99%
To first boot
<60s
Deploy GPU clusters in seconds
On-demand and reserved H100, H200, and B200 — rail-optimized, liquid-cooled, and InfiniBand-available in Seoul.
Built for maximum GPU performance
We optimize networking, cooling, storage, and compute as a single system, so your GPUs spend more time training models and less time waiting on bottlenecks.
Optimized for distributed training
Collective operations pre-profiled per topology, so all-reduce hits line rate from day one.
Use your preferred tools
Slurm, Kubernetes, or raw SSH on bare metal. We hand you the cluster — you run it your way.
Reference node · B200
GPUs8× B200 SXM
GPU memory1,536 GB
NVLink domain1.8 TB/s
Network8× 400G XDR
CPU2× Intel Xeon
System RAM2 TB DDR5
Local NVMe30 TB
CoolingDirect-to-chip
Everything around the GPUs
Storage optimized for GPU throughput, and a software environment that's ready before your first SSH.
Storage that keeps up
Parallel filesystem — VAST / WEKA tiers delivering 1 TB/s+ aggregate read for dataset streaming.
Local NVMe scratch — up to 60 TB per node for checkpoints and shard caches.
S3-compatible object — durable dataset and artifact storage in-region, no egress between tiers.
GPUDirect Storage — DMA straight from NVMe to GPU memory, skipping the CPU bounce buffer.
Ready-to-train images
CUDA 12.x + drivers pre-installed and version-pinned to the GPU generation.
NCCL pre-tuned per topology, with the right environment variables already set.
Frameworks baked in — PyTorch, JAX, Megatron-LM, DeepSpeed, and vLLM.
Bring your own container — any OCI image runs on bare metal, no lock-in.
Built for every AI workload
From a 4,096-GPU pre-training run to a single inference replica — one platform, one bill, one network.
Pre-training
Build and train foundation models at scale on dedicated GPU clusters.
Fine-tuning & RLHF
Customize existing models faster with flexible, on-demand GPU capacity.
Inference
Serve AI applications with low latency, automatic scaling, and pay-as-you-go pricing.
Everything you need to know
Frequently asked questions
On-demand nodes boot in under 60 seconds for standard shapes. Large reserved clusters of 1,024 GPUs and up are typically staged within a business day once the reservation is confirmed.
Yes. Every node lands on non-blocking NDR/XDR InfiniBand with NCCL pre-tuned per topology, so all-reduce hits line rate without manual config. Slurm and Kubernetes operators are available, or you can run raw torchrun.
Parallel filesystem tiers deliver 1 TB/s+ aggregate read, with up to 60 TB of local NVMe scratch per node and GPUDirect Storage for DMA straight into GPU memory. Object storage in the same region keeps datasets and artifacts close.
Always. DanaIX is sovereign by design — data, weights, and checkpoints stay in the region you select, on single-tenant bare metal, with private networking and bring-your-own keys.
Spot is billed per second only while running. You get a two-minute interruption notice to checkpoint, and jobs auto-resume when capacity frees up if you enable it.
Spin up a cluster, or reserve a region.
No procurement calls, no week-long waits.