> See the full index of documentation pages at https://docs.zettabyte.com/llms.txt.

# GPU instances

> Reserve GPU capacity to run notebooks, training, and inference.

Source: https://docs.zettabyte.com/docs/run-on-gpus/gpu-instances

Last updated: 2026-07-21

The **Instance** page is where you provision GPU compute, and it is the page you land on after signing in. A resource you launch here is what the notebook, task, and deployment wizards draw from. See Notebooks and IDEs, Run a training job, and Deploy a model.

The page has two tabs, **Instance** and **Resource Queue**, plus a **Launch Instance** button. Whether you need to launch anything depends on your tenant status; whitelisted tenants may already have assigned resources. See Sign in and accounts.

## Launch an instance

Select **Launch Instance** and set:

1. `Resource Type`: `Standard GPU Instance`.
2. `Computing Config`: choose a `Resource Name` from the dropdown (for example `NVIDIA-A10-PCIE-4GB` or `NVIDIA-H200-SXM-140GB`). `CPU` and memory fill in automatically for the selected specification.
3. `Local Storage`: a free local disk allocation is included.
4. `Count`: the number of instances, up to the maximum shown.
5. `Purchase Cycle`: `On-Demand`, billed hourly.
6. Select **Confirm** to purchase, or **Join The Queue** if the resource is not immediately available.

A summary panel shows the final configuration, including the shared volume mounted at `/share` (see Storage and PVCs) and the estimated compute cost. A launched `On-Demand` resource appears in the **Instance** tab with an active status and becomes selectable in the notebook, task, and deployment wizards.

## Billing and lifecycle

> [!WARNING]
> **Allocated capacity bills whether or not it is busy**
> Billing begins the moment a resource is allocated to you and continues for as long as you hold it, independent of whether a task is running. Allocated but **idle capacity still bills**. There is currently no pause; to stop charges you must release the resource.

Billing is per allocation, hourly, `On-Demand`. See How billing works for the wallet and charge model.

## Releasing an instance

Release a resource when you are done to stop billing. On the **Instance** tab, open the row menu for the resource and select **Terminate**. Releasing returns the capacity to the pool.

> [!WARNING]
> **Save your work before releasing**
> Anything not saved to persistent storage (a `PVC` or `S3`) is **lost when a resource is released**. Write checkpoints, weights, and code you want to keep to a mounted `PVC` first. See Storage and PVCs.

## When capacity is short

If the `Computing Config` you want is not available, select **Join The Queue** instead of **Confirm**. See The resource queue for how the queue works and how allocation is decided.

## Manage instances

The **Instance** tab lists your purchased resources, with these columns: `Resource Name`, `Resource Status`, `Resource Usage`, `GPU Count`, `Purchase Cycle`, `Creation Time` and `End Time`, and `Unit Price`. The row menu offers **View Details** and **Terminate**.

## Notes and limits

- Billing is per allocation, hourly, `On-Demand`. See How billing works.
- There is no pause. Release to stop charges.
- Region: only zCLOUD's US, Taiwan, and Singapore regions are offered. Any region label shown will be one of these.

## Pricing

zCLOUD does not publish resource prices in these docs. For current rates and longer commitments, talk to an expert.