Real fabric
NVLink/NVSwitch in-node, InfiniBand rail-optimised between nodes, NCCL pre-tuned. All-reduce numbers ship with the quote.
Products · clusters
GPU clusters for distributed training, priced like the sheet
Multi-node pods of 8× SXM machines stitched with InfiniBand — 16 to 512 GPUs — at the same public per-GPU prices as single instances. A 64× H100 SXM pod at reserved rates: $65.98/hour, fabric included.
No cluster premium
Per-GPU price × GPU count. Fabric, head nodes and scratch storage are included — compare that line with any hyperscaler quote.
Your scheduler
Slurm installed and configured, or bare SSH and your own orchestration. Root on every node either way.
Days, not quarters
Standard pods provision in 1–3 business days. No committee, no "capacity review call" — a ticket and a quote.
Standard pod configurations
Built from 8× SXM nodes; any multiple works. Reserved rates shown — burst pods at on-demand when capacity allows.
| Pod | GPUs | VRAM total | Interconnect | $/hour (reserved) |
|---|---|---|---|---|
| 2 × 8× H100 SXM | 16 | 1.3 TB | NVLink + IB HDR | $16.50 |
| 8 × 8× H100 SXM | 64 | 5.1 TB | NVLink + IB NDR | $65.98 |
| 8 × 8× H200 | 64 | 9.0 TB | NVLink + IB NDR | $127.17 |
| 8 × 8× B200 | 64 | 12.3 TB | NVLink 5 + IB NDR | $174.85 |
| 64 × 8× H100 SXM | 512 | 41 TB | Rail-optimised NDR | $528 |
Per-GPU maths: reserved rate (public sheet) × GPU count. Nothing else is added.
What teams run on pods
- Pre-training & continued pre-training — FSDP/DeepSpeed/Megatron across nodes.
- Large fine-tunes — 70B+ full-parameter runs that outgrow one machine.
- RLHF / RLVR pipelines — actor, critic and reward fleets on one fabric.
- Batch scale-outs — thousand-GPU-hour render or simulation campaigns.
Single-node 8× machines (no fabric needed) deploy instantly from the console like any instance — clusters are for when one chassis stops being enough.
$ sinfo
PARTITION AVAIL NODES STATE NODELIST
gpu* up 8 idle pgx-[01-08]
$ sbatch --nodes 8 --gpus-per-node 8 train.sh
✓ Submitted batch job 1129
# NCCL all-reduce 256MB: 47 GB/s bus BW (rail-optimised NDR)GPU clusters: FAQ
Start with a console ticket: model, GPU count, region preference, term. A quote with topology comes back within a business day.
What interconnect do clusters use?
Inside a node: NVLink/NVSwitch on SXM platforms. Between nodes: InfiniBand NDR/HDR (400/200 Gb/s per port, rail-optimised) on flagship pods; 100 GbE RoCE on smaller A100/L40S pods. Topology details ship with every quote.
How fast can I get a cluster?
Standard blocks (2–8 nodes of 8× H100/H200) are usually provisioned in 1–3 business days from a console ticket. Larger or unusual topologies are quoted with a delivery date — typically under two weeks.
How are clusters priced?
Per GPU-hour from the same public sheet — reserved rates for committed terms, on-demand for burst blocks when available. A 64× H100 SXM pod at reserved rates runs $66/hour all-in; InfiniBand fabric and management nodes are included, not line items.
What software do they come with?
Ubuntu with the NVIDIA driver stack, CUDA, NCCL tuned for the fabric, Slurm or plain SSH per your preference, and shared NVMe scratch. Bring your own images — clusters are yours at root.