What is Databricks Mosaic AI

Databricks Mosaic AI is best for companies that want AI built directly on top of the data platform they already trust. The key strength here is not a flashy chat interface. It is the combination of model serving, AI functions, governance, evaluation, and private-data workflows inside Databricks.

Mosaic AI sits on top of Databricks’ long-standing position in data engineering, analytics, and machine learning platforms, strengthened by Databricks’ MosaicML acquisition. That history matters because the product is built for teams that see AI as an extension of their data stack, not as a separate consumer-style assistant.

Core offerings

  • Model Serving for real-time and batch inference.
  • AI Functions for applying models directly to data workflows.
  • Support for custom models, Databricks-hosted foundation models, and externally hosted models.
  • Governance, scaling, and route-optimised serving inside the wider Databricks platform.

Pricing

Databricks does not market Mosaic AI with one simple sticker price. Public documentation frames pricing around serverless compute, model serving SKUs, and pay-per-token access for supported foundation models. In practice, the commercial question is how your inference traffic, serving topology, and Databricks account structure translate into DBU and serving costs.

Model footprint

Databricks supports Databricks-hosted foundation models, custom models, AI agents, and external model providers through one serving layer. For many enterprise buyers, that unified control plane is more valuable than a consumer-style model picker.

Why select Databricks Mosaic AI

Mosaic AI is compelling when your company’s strategic advantage lives in private data, governed analytics, and production ML or AI systems. If Databricks is already your operating backbone, keeping AI near the data is a practical advantage.

Official sources: Databricks Model Serving, Supported foundation models, Serving cost monitoring.

Model disclosure note

Mosaic AI is a serving and governance layer over custom, Databricks-hosted, and external models. It is intentionally model-agnostic, so a buyer-facing page should track supported deployment patterns more than a fixed model roster.