Databricks announces AI Runtime is generally available

On October 21, 2026, Databricks announced the general availability of AI Runtime (AIR), bringing fully managed GPU infrastructure directly into Databricks, allowing customers to train models in the Lakehouse where their enterprise data resides. According to Databricks, AIR offers second-level startup, RDMA GPUs, environment caching, and built-in debugging tools, and is protected by Unity Catalog governance and workload isolation, and has already supported training runs of over 100,000 GPU hours. On the same day, another report stated that hundreds of companies are already using AIR, that users can launch their first workload with a single CLI command, and that they can migrate workloads on Kubernetes or SLURM using natural language prompts. Kostia Omelianchuk of Superhuman said the team scaled training from a single node with 8 A100s to multiple nodes with 16 H100s, reducing training time from about 7 days to 48 hours.

AI-generated from public reports

1. Oct 21· 2 reports

   English translation is being prepared.

   Source:Databricks Blog