Databricks Mosaic AI

Databricks Mosaic AI is the AI and machine learning layer built into the Databricks Data Intelligence Platform — its defining difference from a managed inference gateway is that it's built around the lakehouse: training, fine-tuning, and serving models directly against data that already lives in an organization's Databricks environment, rather than treating model access as the primary product and data as something fed in separately through an API call. That lakehouse-first design shapes almost everything about how the platform is evaluated, priced, and compared to the managed inference gateways offered by the major cloud providers.

Training on Your Own Data Is the Point, Not an Add-On

Where a platform like a managed inference gateway is optimized for calling a pre-trained model, Mosaic AI is optimized for the opposite starting point: an organization's own data, already sitting in Delta Lake tables, becomes the training set for a custom or fine-tuned model without an export-and-reimport step. Mosaic AI Training handles the fine-tuning and pre-training compute, Model Serving deploys the result behind a low-latency endpoint, and Unity Catalog governs access to both the data and the resulting models under one permission system — the same one already governing the rest of a team's Databricks workspace. MLflow, already the de facto standard for experiment tracking in much of the ML community, is built into the same workspace, so the same tool that tracks a training run also registers and versions the resulting model before it goes to a serving endpoint. That tight data-to-model loop is the platform's actual differentiator, not raw inference speed or model selection.

Pricing Structure

Pricing is entirely usage-based and doesn't reduce to a flat number any more than a comparable enterprise AI platform's does. Training and fine-tuning jobs are billed by compute time (DBUs, Databricks' own compute-unit pricing, on top of the underlying cloud infrastructure cost), Model Serving endpoints are billed by provisioned throughput or by token for pay-per-token foundation model access, and vector search and feature serving carry their own separate meters. Anyone budgeting for Mosaic AI needs to price an actual training job and serving workload, not treat it as a subscription with a single monthly figure.

Mosaic AI vs. Managed Inference Gateways

The criterion that separates Mosaic AI from AWS Bedrock or Google Vertex AI is scope, not cloud preference: those platforms are primarily model access plus a managed wrapper around calling a hosted model, while Mosaic AI is built around the data lakehouse — training and fine-tuning on an organization's own pipeline is the primary use case, not an optional extra bolted onto inference. A team that mainly needs to call Claude or GPT from an existing application fits a managed gateway; a team whose real bottleneck is training a custom model against a large, already-governed internal dataset is closer to what Mosaic AI is built for. That also puts it in a different lane from Hugging Face: Hugging Face is where a team discovers and experiments with open models before deciding where to run them, while Mosaic AI is where a team that has already chosen to build on its own data does that training and serving, inside infrastructure it already governs through Unity Catalog.

The practical decision comes down to whether the real workload is calling an existing model or training one against proprietary data that's already sitting in a Databricks lakehouse — the second case is where Mosaic AI's tight data-to-model loop pays for itself, and the first case is usually better served by a narrower managed inference gateway.