AI Runtime (GPU serverless)
AI Runtime (GPU serverless) is a Databricks compute / bi capability, introduced June 2025.
Foundation Model Training Mosaic AI Model Training Foundation Model Fine-tuning AI Runtime (GPU serverless)
GPU support for Databricks serverless compute - on-demand A10, H100, and B300 accelerators for deep-learning training and fine-tuning, with no clusters to provision.
- It puts on-demand GPUs behind serverless compute, so a deep-learning job gets accelerators without anyone sizing a cluster - the catch the docs admit is that on-demand also means periods where GPU capacity in your region is constrained or unavailable.
- It first showed up in June 2025 under the plainer name 'Serverless GPU compute.'
Limitations: Only A10, H100, and B300 accelerators; no compliance security profile support for FedRAMP High or DoD IL5. Connection attempts time out after 15 minutes for interactive notebooks and 24 hours for jobs; connected notebook sessions and notebook jobs run for up to two days and CLI jobs for up to 14 days, so long training needs checkpointing. For scheduled jobs, Environments-panel dependencies and auto-recovery for incompatible packages aren't supported; GPU capacity can be constrained, and cross-region GPUs during peak demand can incur egress costs.
Open in REbricked →- Category
- Compute / BI
- Also known as
- AI Runtime, Serverless GPU compute
- Verified
- 2026-09-27