Databricks: Unable to Attach Notebook to Serverless GPU Compute — AI Runtime Not Available Error

Mars-5703 0 Reputation points
2026-11-13T16:28:37.52+00:00

Problem description

Attaching a notebook to serverless GPU compute (AI Runtime) in our Azure Databricks workspace fails for all users. The UI displays a generic attachment error, while the underlying HTTP 400 response states:

BAD_REQUEST: AI Runtime is not available. To access this feature, please contact sales for assistance.

Environment

  • Azure Databricks workspace in East US.
  • Notebook-based workflow requiring serverless GPU compute.
  • Provisioned directly through Azure, with no existing Databricks sales relationship.
  • Standard serverless CPU compute works successfully.

What I've already tried

  • Created a new notebook using standard serverless CPU compute and confirmed successful execution.
  • Confirmed that serverless GPU attachment fails across users.
  • Inspected browser network requests and identified the underlying HTTP 400 error.
  • Enabled advanced diagnostics for Microsoft Support.
  • Provided subscription information, workspace URL, notebook ID, and configuration screenshots through the support case.

Microsoft Support suggested a workspace-level availability or enablement restriction but has not confirmed the specific cause or provided enablement steps. The support workflow has now directed me to Microsoft Learn Q&A.

Current status

Serverless GPU access remains blocked. I need confirmation of:

  1. Whether serverless GPU / AI Runtime is supported for our workspace's region and configuration.
  2. The specific restriction causing the request to be rejected.
  3. The steps required to enable access and the team responsible for doing so.
  4. How an Azure customer without an existing Databricks sales contact can obtain the required enablement.

Standard serverless CPU has already been validated. Using CPU or a regular GPU cluster does not resolve the request for serverless GPU access.

Azure Databricks
Azure Databricks

An Apache Spark-based analytics platform optimized for Azure.

0 comments No comments

2 answers

Sort by: Newest
  1. Mariusz Bialek 0 Reputation points
    2026-11-19T16:35:58.3766667+00:00

    Hi, thank you for responding.

    Yes, both settings are already enabled. The screenshot provided shows “AI Runtime” and “AI Runtime multi-node H100” under Generally available features, both marked GA and On.

    Despite this, attaching serverless GPU fails with:

    “BAD_REQUEST: AI Runtime is not available. To access this feature, please contact sales for assistance.”

    Standard serverless CPU works successfully, and the GPU failure affects multiple users. Configuration screenshots and advanced diagnostic access have already been provided through Microsoft Support.

    Is there an additional entitlement or rollout restriction separate from these switches?

    The issue remains unresolved; enabling the settings is a step we have already completed.

    Was this answer helpful?

    0 comments No comments

  2. Rukshan edirisinghe 2,930 Reputation points
    2026-11-15T03:57:04.2766667+00:00

    Hi @Mars-5703

    Serverless GPU (AI Runtime) is still in Public Preview on Azure Databricks, and it stays blocked until a workspace admin turns the preview on. East US is one of the supported regions, so the region isn't the issue. Since serverless CPU works fine, a preview that hasn't been switched on is the most likely reason for that "contact sales" response.

    1. Sign in to the workspace as a workspace admin.
    2. Click your username in the top bar and select Previews.
    3. Find the AI Runtime preview and switch it on.
    4. Reopen the notebook, go to the compute selector and attach to serverless GPU again.

    If you don't see the toggle there, an account admin can check the Previews page in the account console, since some previews are managed at the account level.

    Do you see the AI Runtime toggle on the Previews page?

    If my response resolved your issue, please consider accepting it as the answer. If the issue persists, let me know and I'll be happy to continue helping.

    Reference:

    Was this answer helpful?

    0 comments No comments

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.