Databricks makes AI Runtime generally available for managed GPU training

The announcement targets teams training models where their enterprise data already resides. Databricks says it handles GPU procurement, cluster configuration and networking. The service combines computing infrastructure with developer tools and data governance, reducing the infrastructure work that these teams would otherwise manage separately.

The AWS documentation qualifies access: supported geographic areas and cross-area processing requirements must be checked, and GPU availability varies by workspace. Workspaces with the compliance security profile enabled require an administrator to enable the AI Runtime and AI Runtime multi-node H100 previews. Regional capacity can be constrained or unavailable; using GPUs across regions can also incur data-transfer charges.

The quickstart begins with an enabled workspace and an installed or updated command-line tool authenticated to it. Its minimal example requests one A10 GPU and prints the device name, before introducing Python code. Submitting the workload returns a run identifier and workspace link. The optional --watch flag keeps the terminal connected and streams logs until completion; these are documented steps, not a test performed for this article.

Custom Docker containers remain a beta feature. They require administrator-enabled AI Runtime Beta Features and Artifact Registry previews, CLI version 1.19.0 or later, registry read and push permissions and running local Docker. Images must be stored in the same workspace’s registry and stay below 20 GB. Additional packages belong in the Dockerfile: the custom-image setting cannot be combined with the standard environment’s dependency or version fields.

For budget tracking, the usage guide provides a billing-table query that totals usage quantities for serverless GPU workloads. It links separately to current GPU pricing. The query’s quantity total is therefore not itself a monetary cost estimate.