Which option is right for you?
Three ways to get GPUs, one software stack. Move between them without rewriting code.
| GPU CloudOn-demand | GPU ClustersReserved | Custom ClustersOwned | |
|---|---|---|---|
| Best for | Experiments, fine-tuning, inference, bursty workloads | Distributed training and production inference at scale | Long-term, high-utilization, data-sovereign workloads |
| GPUs | 1 to 8 per instance, unlimited instances | 16 to 1,024+ in one InfiniBand fabric | 8 to 4,096+, sized to your plan |
| Commitment | None, per-minute billing | 2 weeks to multi-year | Hardware purchase or lease, 1 to 5 year operations |
| Time to start | Under a minute | 2 to 9 days | 4 to 24 weeks |
| Pricing model | Per GPU-hour, on demand | Per GPU-hour, flat rate by size and term | Hardware at cost plus design fee and per-GPU operations |
| Networking | 200 Gb/s Ethernet, NVLink inside SXM nodes | Non-blocking Quantum-2 InfiniBand, 400 Gb/s per GPU | Your choice: InfiniBand, Spectrum-X, up to 800 Gb/s |
| Storage | Local NVMe, persistent filesystems, object storage | Parallel file system included, object storage | WEKA, VAST, Lustre or DDN, sized to you |
| Orchestration | SSH, Docker, dashboard, API, CLI | Managed Slurm or Kubernetes, or bare metal | Any: we install and operate what you choose |
| Tenancy | Dedicated GPUs, shared facility | Single-tenant fabric and nodes | Fully owned, single-tenant, on-prem or private cage |
| Ownership | GREENAIBOX | GREENAIBOX | You |
| Operated by | GREENAIBOX | GREENAIBOX, with named solutions engineer | GREENAIBOX under an operations contract, or your team |
| Support | Email and chat, 24/7 for incidents | 24/7 engineer on call, 99.9% SLA | 24/7 with on-site response options |
| Managed Training available | Yes | Yes | Yes, on your cluster |
| Launch instance | Reserve capacity | Request a design |
A quick way to decide
Utilization over the next 12 months is the number that matters most. Here is how our solutions engineers read it.
Stay on demand
You pay only for what you use and can move to new GPU generations the day they launch. Ideal while you are still finding the workload.
GPU CloudReserve a cluster
Flat rates run 15 to 30 percent below on-demand on terms up to a year, 40 percent or more on multi-year commitments, and you still have an exit at the end of the term. Most training teams live here.
GPU ClustersOwn the hardware
Payback on a custom cluster typically lands at 18 to 26 months. Add regulatory or residency requirements and the case gets stronger.
Custom ClustersNot sure? Book a capacity planning session.
Still deciding? Talk it through with an engineer.
A 30-minute call with a solutions engineer, no sales script. We'll model your costs across all three options.