NVIDIA and AMD GPUs, by the hour or by the cluster.
Launch a single GPU in a VM, reserve a rack, or take a bare-metal cluster on InfiniBand. Prices are published, and reserved capacity is about 25% below on-demand.
Current generation, both vendors
Choose by memory, interconnect and price, not by which vendor your cloud happens to stock.
| GPU | Memory per GPU | On-demand / GPU-hr | Reserved 12 mo / GPU-hr |
|---|---|---|---|
| NVIDIA H200 SXM | 141 GB HBM3e | $5.50 | $4.10 |
| NVIDIA HGX B200 | 180 GB HBM3e | $7.49 | $5.60 |
| NVIDIA HGX B300 | 288 GB HBM3e | $9.25 | $6.90 |
| AMD Instinct MI300X | 192 GB HBM3 | $2.99 | $2.29 |
| AMD Instinct MI355X | 288 GB HBM3e | $5.29 | $3.99 |
Prices per GPU-hour in USD, billed per second for on-demand. Market reference, Oct 2026: H200 median $4.92 (Nebius $5.40, CoreWeave about $6.31); B300 median $9.42 (Nebius $9.50, AWS $17.80). Sources: savrn GPU index, providers' pricing pages.
Four consumption models
On-demand
1 to 8 GPUs in a VM, per-second billing, no commitment.
Reserved
1 to 36 months. Lower rate, guaranteed capacity, same API.
Bare metal
Full 8-GPU nodes with root access, for custom stacks and maximum performance.
Clusters
16 to 1,024+ GPUs on a non-blocking InfiniBand or RoCE fabric, with Slurm or Kubernetes.
The parts of a cluster that decide time-to-train
- Non-blocking fabric400 Gb/s per GPU on InfiniBand NDR or RoCE, rail-optimised for collective operations.
- Parallel storageHigh-throughput shared file system for checkpoints, plus S3-compatible object storage.
- Health-checked nodesBurn-in before hand-over, continuous GPU and link checks, automatic replacement of failed nodes.
- Slurm or KubernetesManaged Slurm, managed Kubernetes, or both on the same reservation.
- Data centersTier III US data centers; site details on request.
Need 64 GPUs next month?
Tell us the model, size and start date. We will reply with a reserved quote and a delivery date.