GPUs by the hour
Resource costs
Minimum of 0.125 cores per container
Includes 1 TiB / mo free
Hardware
Nvidia GB300 NVL72
72 Blackwell Ultra GPUs in one NVLink domain.
$920.06 / rack / hour
- 72 GPUs per rack
- Root SSH or your own Kubernetes namespace
- Firewall rules and SSH keys from the dashboard
- Live node status and addresses
Nvidia HGX B300
Eight Blackwell Ultra GPUs per node.
$51.11 / node / hour
- 8 GPUs per node
- Root SSH or your own Kubernetes namespace
- Firewall rules and SSH keys from the dashboard
- Live node status and addresses
Nvidia HGX B200
Eight Blackwell GPUs per node.
$45.00 / node / hour
- 8 GPUs per node
- Root SSH or your own Kubernetes namespace
- Firewall rules and SSH keys from the dashboard
- Live node status and addresses
Compare features
| On-demand | Prepaid | Custom | |
|---|---|---|---|
| Billing | |||
| Rate | List | Discounted | Volume pricing |
| Term | Monthly | Prepaid term | Negotiated |
| Billing unit | GPU-hour | GPU-hour | GPU-hour |
| Hardware | |||
| HGX B200 | |||
| HGX B300 | |||
| GB300 NVL72 | |||
| Access | |||
| Root SSH | |||
| Kubernetes namespace | |||
| Firewall rules | |||
| SSH key management | |||
| Node status and addresses | |||
Frequently asked questions
What counts as billable time?
Reserved GPUs bill by the GPU-hour for as long as they are reserved to you, busy or idle. The per-second rate is the same price divided by 3,600.
How are CPU and memory metered?
CPU bills per physical core, which is 2 vCPU, with a minimum of 0.125 cores per container. Memory bills per GiB.
What is the difference between on-demand and prepaid?
On-demand bills monthly at the list rate. Prepaid pays for a term up front at a lower rate. Talk to us for a quote.
How do I access my nodes?
Root SSH, or your own Kubernetes namespace. Add SSH keys, set firewall rules and see every node and its address on the dashboard.
What do disk and network egress cost?
Persistent disk is $0.081 per GiB per month, with 1 TiB free each month. Egress is $0.036 per GiB.
Can Morph run the model for me instead?
Yes. Dedicated inference serves open models on reserved B200 and B300 capacity behind the Morph API, so you never run the servers.
Need a whole cluster?
Multi-rack GB300 and B300 builds with volume pricing on a custom term.