Databricks Blocks New Foundation Model Fine-Tuning Installs on AWS
The August 14 cutover leaves existing runs working but moves new custom-model projects toward a public-preview serverless GPU environment.
Databricks documented August 14, 2026 as the cutover after which new Foundation Model Fine-Tuning installations on AWS would be blocked, while existing runs could continue. The company directs new custom-model work toward AI Runtime, changing the starting point from its older managed fine-tuning path to a serverless GPU environment.
The shift matters to machine-learning platform teams that planned to begin a new fine-tuning project through the retiring route. It is not a retirement of every model produced through that system, and it does not require a current run to stop. It is a boundary between work already under way and the setup path available for the next job.
What the August 14 cutover changes
Databricks' retirement notice makes a narrow but important distinction: existing Foundation Model Fine-Tuning runs continue to work, while new installations are blocked after the stated date. That makes this a migration event for future starts rather than an abrupt invalidation of work already running.
The scope also matters. The source page is the Databricks on AWS documentation, so this article does not extend the same operational conclusion to Azure or Google Cloud workspaces. Teams using another cloud should check the documentation for their own deployment rather than assume that every region, account and interface moved in exactly the same way.
AI Runtime is a different operating model
Databricks presents AI Runtime as the destination for new training and fine-tuning work. Its documentation describes serverless GPU access for deep-learning workloads and marks the service as Public Preview. The documented AWS accelerator choice is limited to A10 and H100 hardware.
Those details make the change more than a product rename. A team moving from a dedicated fine-tuning route to a serverless compute environment should re-check how it provisions work, installs dependencies, tracks experiments and handles capacity. The official pages establish the direction of travel, but they do not establish that every old workflow has a one-step equivalent in the new environment.
Who needs to act now
The clearest affected group is an AWS team preparing to start a new Foundation Model Fine-Tuning project through the legacy installation path. Its immediate task is to test the intended workload in AI Runtime before promising a delivery date or assuming feature parity.
Teams with an existing run have more breathing room because Databricks says that work can continue. Even so, the retirement boundary is a reason to document reproducibility now: a future retraining cycle or replacement project may need a different environment from the one used for the original run.
Platform owners should also treat the preview label as a real planning constraint. The documentation supports the existence and broad shape of AI Runtime; it does not, by itself, prove production reliability, cost equivalence, capacity availability or migration success for a particular workload.
Evidence limits and what to watch
The available evidence is official and specific about the scheduled date, existing-run continuity and the successor environment. It is not a post-cutover operations report. Databricks has not published, in these two pages, a completion notice showing that every workspace now enforces the block, a migration success rate, a price comparison or customer outcome data.
The next useful evidence would be completed-cutover language, a detailed parity or migration guide, and a change in AI Runtime's preview status. Until then, teams should verify the behavior in their own AWS workspace and treat unsupported assumptions about cost, performance or availability as open questions.
Status and confidence
Confirmed. Databricks' official AWS documentation specifies the August 14 cutover and identifies AI Runtime as the migration path. Internal confidence is medium because the evidence is first-party, the retirement page still uses scheduled language, and there is no independent or post-cutover confirmation for every workspace.
Sources
Update note: Last reviewed 2026-08-15. We will revise this post if Databricks publishes completed-cutover evidence, migration details or a change to AI Runtime availability.
Sources
- Databricks on AWS — Foundation Model Fine-Tuning retirement — official
- Databricks on AWS — AI Runtime — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.