Spot Bare Metal GPU

Spot Bare Metal GPU clusters are discounted GPU servers that utilize unused capacity at reduced pricing. They provide the same hardware specifications and functionality as standard Bare Metal GPU clusters, with one key difference: they can be reclaimed with 24 hours’ notice.

Spot vs On-demand

Aspect	On-demand (Bare Metal GPU)	Spot (Spot Bare Metal GPU)
Pricing	Standard rates	Discounted rates
Availability	Guaranteed until deleted	Can be reclaimed with 24 hours’ notice
Use case	Production workloads, critical applications	Cost-sensitive workloads that tolerate interruption
Capacity source	Dedicated capacity	Unused/excess capacity

When to use Spot clusters

Spot clusters are ideal for interruptible workloads, such as batch processing, experiments, testing, and development. They should not be used for production inference, time-critical tasks, long-running jobs without checkpoints, or any workload where unexpected reclamation could have serious consequences.

Availability

Spot Bare Metal GPU availability depends on region and current stock. When available, a Spot Bare Metal GPU option appears alongside the standard Bare Metal GPU in the cluster type selector:

If only Bare Metal GPU appears in the selector, Spot is not currently available in that region. In some regions, Spot may appear but show “Out of Stock”—this indicates the option exists, but no capacity is currently available.

Reclamation process

Spot clusters can be reclaimed when Gcore needs the capacity for on-demand workloads or other operational requirements. The reclamation process follows a fixed timeline:

1. An email notification is sent to the account owner.
2. A 24-hour window begins to save data, transfer workloads, and prepare for cluster deletion.
3. The cluster is deleted. Data on local storage is erased immediately.

The notice period is fixed and starts when the email is sent. After 24 hours, the cluster is deleted automatically.

Data preservation

When a Spot cluster is deleted, data is handled as follows:

Resource	What happens
Local NVMe storage	Erased immediately
File shares	Not affected (independent resource)
Object storage	Not affected (independent resource)

To protect critical data, use file shares for datasets, checkpoints, and model weights. Save outputs and backups to object storage. Implement regular checkpointing in training scripts every 1-4 hours. When a reclamation notice is received, prioritize transferring any data not already on persistent storage.

Pricing and billing

Spot clusters are billed at a discounted rate compared to standard Bare Metal GPU. The exact discount varies by region and GPU model. The flavor selection card displays both hourly and monthly rates.

Billing is per entire node (all GPUs on the server), calculated per minute, and aggregated hourly. Billing stops when the cluster is deleted, whether by user action or reclamation.

A minimum account balance is required before provisioning. If the balance is insufficient, provisioning will fail. For details, see GPU Cloud billing.

Creating a Spot cluster

The creation process is identical to standard Bare Metal GPU clusters, with one additional step: acknowledging the Spot terms.

1. Navigate to GPU Cloud > GPU Clusters > Bare Metal GPU Clusters.
2. Click Create Cluster.
3. Select a region where Spot is available.

4. In GPU Cluster type, select Spot Bare Metal GPU. A warning banner displays the terms and conditions.
5. Select a GPU flavor and configure network, SSH key, and cluster name.
6. Click Create Cluster.

For detailed configuration, see Create a Bare Metal GPU cluster.