AI Runtime is Generally Available

AI Runtime (AIR) is now generally available, embedding fully managed GPU infrastructure directly into Databricks to eliminate the need for separate GPU stacks. The service addresses bottlenecks such as capacity planning, node maintenance, and networking that traditionally delay model training. By consolidating GPU infrastructure with enterprise data governance, AIR reduces operational overhead and minimizes risks associated with fragmented access controls and data copies.

AIR simplifies GPU workload deployment through multiple interfaces, including a single CLI command for initial workloads and natural language prompts for migrating from Kubernetes or SLURM clusters. Native integration with Lakeflow ensures GPU workloads align with upstream and downstream data pipelines. Production deployment requires minimal YAML configuration, with built-in priority scheduling and monitoring tools like Genie ZeroOps to optimize resource use and reduce operational tasks.

The platform accelerates distributed training by enabling environments to spin up in seconds, even for large-scale RDMA-enabled clusters with dozens of GPUs. Pre-built environments and optimized caching reduce setup time, while compatibility with major frameworks like Ray allows seamless code migration. Enhanced observability in MLflow provides real-time log streaming and debugging tools tailored for multinode training, further reducing delays in model iteration.

AIR extends Databricks' data governance controls to GPU workloads through Unity Catalog, enforcing consistent permissions across data access, checkpoints, and logs. VM-level isolation, PrivateLink support, and serverless egress controls strengthen security compared to conventional GPU isolation methods. Cost tracking is integrated into Databricks billing, enabling workload-level usage analysis to guide future GPU budget decisions.