The confusion comes from AI Runtime having two components with different preview states.
Single-node tasks are already open and do not require any toggle. You can just open a notebook, click the Connect dropdown, select Serverless GPU, choose your accelerator (A10 or H100), and apply.
Multi-GPU distributed training is different. It requires a workspace admin to enable it through the Previews page. If you do not see the toggle, it is likely not available for your region or plan yet.
If "Serverless GPU" does not appear in the Connect menu at all, I would check the following:
- Unity Catalog is enabled in the workspace
- Serverless compute is enabled under Admin Console > Compute
- The workspace region supports Serverless GPU
- Your account tier includes access to this feature
If all of the above is correctly configured and the option is still missing, I would recommend reaching out to your Databricks account team. Public Preview features are often rolled out progressively and may still require allowlisting for some accounts.