Databricks introduces serverless AI Runtime with NVIDIA GPUs

Databricks introduced AI Runtime as a serverless NVIDIA GPU offering for training and fine-tuning.[2] The company’s blog frames the service as a way to make it easier to customize open models with enterprise data.[1] Databricks says its serverless GPUs can train deep-learning models and fine-tune open-source generative AI models without users managing infrastructure.[2] Its documentation describes the service as fully managed GPU compute for training, fine-tuning and other deep-learning workloads.[3] For teams evaluating it, the central change is access to managed compute for these workloads; the supplied information does not specify a price or performance benchmark.

Databricks says AI Runtime provides on-demand A10 and H100 GPUs.[3] Users can access the GPUs through the Databricks CLI, notebooks, Jobs and Ray.[3] Documentation also lists support for Ray Core, Ray Data, Ray Train and Ray Tune.[3] These access paths give teams options for working with GPUs through different interfaces and Ray components, rather than requiring one documented entry point. The evidence does not specify GPU capacity, regional availability or pricing, so teams should confirm those details for their intended workloads.

Databricks provides pre-built environments for training and fine-tuning workloads, and says users can bring their own Docker container.[3] AI Runtime includes MLflow experiment tracking, along with logging and debugging agents as built-in observability features.[3] Databricks also says the service can schedule GPU workloads and that users can run workloads in CI/CD pipelines.[3] Together, these features address environment setup, experiment records and workflow integration. Teams considering a move should check whether the provided environments and observability features fit their existing training and deployment processes.

AI Runtime can stream data from Unity Catalog volumes using fault-tolerant loaders, which Databricks says are built for high GPU utilization.[3] That connects data loading with GPU workloads in the documented offering, though the evidence does not provide comparative utilization figures. Unite.AI reports that Databricks announced AI Runtime’s general availability on October 21, 2026.[4] The materials provided here do not include further availability terms or a price. For a deployment decision, teams can use the listed hardware, access methods and data workflow as evaluation points, then verify commercial and operational details with Databricks.