Databricks makes AI Runtime generally available for GPU training

The platform handles GPU procurement, cluster configuration, networking and health checks, offering seconds‑level launch latency and automatic scaling across dozens of GPUs. Users can start a job with a single CLI command, migrate workloads from Kubernetes or SLURM with a natural‑language prompt, and integrate with Lakeflow pipelines. Superhuman’s Kostia Omelianchuk reported moving from one node with eight A100 GPUs to 16 H100 GPUs, cutting training time from roughly seven days to 48 hours—a 3.5x speedup.

Additional features include UCVolumeDataset for higher data throughput, support for custom Docker containers (including NVIDIA NGC), and observability via MLflow and Genie tools. Customers can access usage and cost data alongside other Databricks workloads in billing tables.