Databricks AI
Databricks AI (The data + AI platform for building production-grade GenAI systems) Databricks AI is the artificial-intelligence layer of the Databricks Data Intelligence Platform — a unified suite of products for building, serving and governing machine-learning and generative-AI applications on top of an organisation's own data.
The umbrella covers **Mosaic AI** (model training, fine-tuning, evaluation, serving, and Mosaic AI Gateway), the open-source **DBRX** foundation model, **AI Functions** (SQL-level access to LLMs from inside the lakehouse), **AI/BI Genie** (natural-language data exploration), **Vector Search** (managed retrieval for RAG), and **Foundation Model APIs** (pay-per-token access to Llama, Claude, Mistral and DBRX).
The platform is built on Apache Spark, Delta Lake, MLflow and Unity Catalog — all open-source projects that Databricks originated or maintains. In June 2023 Databricks acquired MosaicML for $1.3 billion to bring foundation-model training in-house, and in March 2024 released **DBRX**, a 132-billion-parameter open-source mixture-of-experts model that briefly held the top spot on several LLM evaluation leaderboards. As of 2024-25 reporting, Databricks has crossed a $3 billion annual run rate (publicly disclosed by the company) with the AI portfolio cited as its fastest-growing segment, and serves more than 10,000 customers globally.
Overview
What's inside Databricks AI
The platform is best thought of as five tightly-integrated layers stacked on top of the Databricks lakehouse:
- Mosaic AI Model Training & Fine-tuning — pre-train or fine-tune foundation models on private data without leaving the customer's cloud account, using the GPU infrastructure Databricks inherited from MosaicML.
- Mosaic AI Model Serving — single endpoint that serves OpenAI, Anthropic, Llama, DBRX or a customer-fine-tuned model with autoscaling, A/B routing, cost-attribution and governance baked in.
- Mosaic AI Vector Search — managed vector index synced from Delta tables; the recommended retrieval layer for RAG applications on Databricks.
- AI Functions in SQL — call
ai_query(),ai_classify(),ai_extract()directly from SQL inside a notebook or dashboard, with billing rolled into the lakehouse DBU charge. - AI/BI Genie — natural-language data exploration trained on the customer's schema and Unity Catalog metadata; non-technical users ask questions in English and Genie generates the SQL.
Why people choose Databricks AI
The headline advantage is that the AI lives next to the data. Customers who already store their analytical and ML data in Databricks can build a RAG application, fine-tune a model, or attach an AI function to a BI dashboard without exporting data to a third-party API or stitching together four separate vendors. Unity Catalog provides a single governance plane across the data, the models, the vector index and the serving endpoints.
DBRX, the open-source model
DBRX is a 132-billion-parameter mixture-of-experts model released by Databricks under a permissive licence in March 2024. It was designed and trained by the Mosaic AI research team and is available both as a hosted endpoint inside Foundation Model APIs and as weights downloadable from Hugging Face. DBRX is the in-house option Databricks recommends when customers want a model that runs entirely inside their own VPC.
History
Databricks AI was founded in September 2013 in San Francisco, California, U.S. by Ali Ghodsi (Co-founder & CEO), Matei Zaharia (Co-founder & CTO), Ion Stoica (Co-founder & Executive Chairman), Reynold Xin (Co-founder & Chief Architect), Andy Konwinski (Co-founder), Patrick Wendell (Co-founder) and Arsalan Tavakoli-Shiraji (Co-founder).
Databricks was founded in September 2013 by the original creators of Apache Spark at UC Berkeley's AMPLab — Ali Ghodsi, Matei Zaharia, Ion Stoica, Reynold Xin, Andy Konwinski, Patrick Wendell and Arsalan Tavakoli-Shiraji. The company commercialised Spark, then introduced Delta Lake (2019), MLflow (2018), and Unity Catalog (2021) — open-source projects that together became known as the "lakehouse" architecture.
In June 2023 Databricks acquired MosaicML, the foundation-model training start-up co-founded by Naveen Rao and Jonathan Frankle, for $1.3 billion. The MosaicML team became the core of what is now Mosaic AI, and Rao took the title VP of Generative AI inside Databricks. The acquisition gave Databricks the in-house capability to train foundation models from scratch on customer data.
In March 2024 Databricks released DBRX, a 132-billion-parameter open-source mixture-of-experts model that briefly led several public LLM benchmarks. Throughout 2024 and 2025 the company expanded the AI portfolio with Vector Search, AI Functions in SQL, Mosaic AI Gateway, and AI/BI Genie. In late 2024 a Series J round valued Databricks at roughly $62 billion, making it one of the most valuable privately held enterprise-software companies in the world.
Features
Databricks AI offers the following capabilities:
- Mosaic AI Model Serving. Single managed endpoint that serves OpenAI, Anthropic, Llama, DBRX and customer-fine-tuned models with autoscaling, A/B routing and cost attribution.
- Mosaic AI Vector Search. Managed vector index automatically synced from Delta tables — the recommended retrieval layer for RAG applications on the lakehouse.
- AI Functions in SQL. Call ai_query(), ai_classify(), ai_extract() and friends directly from SQL inside notebooks, dashboards or jobs. No separate model endpoint to manage.
- AI/BI Genie. Natural-language data exploration trained on the customer schema and Unity Catalog metadata. Non-technical users ask in English; Genie generates the SQL.
- DBRX foundation model. 132-billion-parameter open-source MoE model released by Databricks in 2024. Available as a hosted endpoint and as downloadable weights on Hugging Face.
- Foundation Model APIs. Unified pay-per-token API for Llama, Claude, Mistral, DBRX and other models — no separate vendor account required.
- Unity Catalog governance. Single governance plane spanning data, ML features, vector indexes, models and serving endpoints. Lineage, access control, audit logs across everything.
- Mosaic AI Training. Pre-train and fine-tune foundation models on private data using the GPU infrastructure Databricks inherited from MosaicML — without data leaving the customer cloud account.
Use cases
- Enterprise RAG applications (AI platform teams). Build retrieval-augmented chatbots and copilots grounded in private documents using Vector Search + Foundation Model APIs + Unity Catalog access controls.
- Foundation-model fine-tuning (ML researchers). Fine-tune Llama, DBRX or a custom base model on proprietary corpora without exporting training data outside the customer-managed cloud account.
- Natural-language analytics (Business teams). Deploy AI/BI Genie on top of the lakehouse so non-technical users can ask data questions in English and get back governed answers and visualisations.
- Inline LLM SQL workflows (Data engineers). Use ai_classify() / ai_extract() / ai_query() inside ELT pipelines to enrich, classify and structure unstructured columns at lakehouse scale.
- Regulated-industry AI (Healthcare, finance, public sector). Run GenAI workloads with HIPAA-eligible deployments, customer-managed keys and full Unity Catalog audit trails — meeting compliance requirements other AI platforms struggle to satisfy.
Pricing
Databricks AI uses a subscription pricing model, starting at Pay-as-you-go (per DBU). The product offers a 14-day free trial.
| Plan | Price | What's included |
|---|---|---|
| Free trial | $0 /14 days | Full Databricks Workspace with Mosaic AI Model Serving, Vector Search, AI Functions and Foundation Model APIs. Trial credit is consumption-based; expires after 14 days.
|
| Standard | $0.07 /DBU (Jobs Compute, AWS) | Entry-tier SKU. Pay only for compute you use, billed per Databricks Unit (DBU). DBU rates vary by workload type (Jobs, All-Purpose, SQL Warehouse, Model Serving) and cloud provider.
|
| Premium | $0.10 /DBU (Jobs Compute, AWS) | For teams needing role-based access, audit logs, and managed identity. Same DBU consumption model as Standard with a higher per-DBU rate to cover enterprise security.
|
| Enterprise | Custom /contract | Volume-discounted DBU rates, dedicated Solutions Architect, SLA-backed support, HIPAA-eligible deployments and single-tenant options.
|
Databricks AI vs Snowflake Cortex
The table below contrasts Databricks AI with Snowflake Cortex (www.snowflake.com) on a few key dimensions.
| Dimension | Databricks AI | Snowflake Cortex |
|---|---|---|
| Pricing model | Consumption-based per Databricks Unit (DBU); SKU tiers Standard / Premium / Enterprise | Consumption-based per Snowflake credit; flat per-token rate for Cortex functions |
| Foundation models | DBRX (in-house, open-source 132B MoE), Llama 3, Mistral, Claude, OpenAI via API | Snowflake Arctic (in-house 480B MoE), Llama 3, Mistral, Reka — narrower selection |
| Vector search | Mosaic AI Vector Search — managed, synced from Delta tables | Cortex Search — managed, synced from Snowflake tables |
| Inline AI in SQL | ai_query, ai_classify, ai_extract, ai_translate, ai_summarize | COMPLETE, EXTRACT_ANSWER, SENTIMENT, SUMMARIZE, TRANSLATE |
| Fine-tuning | Full fine-tuning + pre-training via Mosaic AI on customer GPUs | LoRA fine-tuning only on selected base models |
| Natural-language analytics | AI/BI Genie — schema-aware text-to-SQL with Unity Catalog governance | Snowflake Copilot — schema-aware text-to-SQL inside Snowflake worksheets |
| Governance plane | Unity Catalog — single governance across data, models, vector indexes, endpoints | Horizon Catalog — single governance across data and Cortex artefacts |
| Best for | Teams running mixed analytics + ML + GenAI on a single lakehouse | Teams whose data + governance already lives in Snowflake |
Platforms & tech
- Supported platforms: Web (Databricks Workspace), AWS, Azure, Google Cloud.
- Deployment: cloud.
- Built with: Scala, Python, TypeScript, Rust, SQL.
Integrations
Databricks AI integrates with 16 third-party tools:
Awards & recognition
- Gartner Magic Quadrant — Cloud Database Management Systems, Leader (2024)
- Forrester Wave — Data Lakehouses, Leader (2024)
- Fast Company — Most Innovative AI Companies (2025)
References
- "Databricks AI — official product page". databricks.com.
- "Mosaic AI — official product page". databricks.com.
- "Databricks Blog (DBRX + acquisitions)". databricks.com.
- "Databricks Newsroom — press releases". databricks.com.
External links
Frequently asked questions
- What is Databricks AI?
- Databricks AI is the artificial-intelligence layer of the Databricks Data Intelligence Platform. It bundles Mosaic AI (model training, fine-tuning, serving), the open-source DBRX foundation model, AI Functions for SQL, AI/BI Genie for natural-language analytics, Vector Search for RAG, and Foundation Model APIs that proxy access to Llama, Claude, Mistral and DBRX. The platform is built on top of the open-source lakehouse stack — Apache Spark, Delta Lake, MLflow and Unity Catalog — that Databricks itself originated.
- How much does Databricks AI cost?
- Databricks AI uses consumption-based pricing billed in Databricks Units (DBUs). Rates vary by SKU tier (Standard ~$0.07/DBU, Premium ~$0.10/DBU, Enterprise custom), by workload type (Jobs Compute, All-Purpose Compute, SQL Warehouse, Model Serving) and by cloud provider (AWS, Azure, GCP). There is no flat monthly fee for the AI products themselves — every component (Model Serving, Vector Search, AI Functions, Foundation Model APIs) draws from the same DBU pool. A 14-day free trial with consumption credits is available; there is no perpetual free tier.
- What is Mosaic AI inside Databricks?
- Mosaic AI is the brand for Databricks' generative-AI and ML product suite, formed after the $1.3 billion acquisition of MosaicML in June 2023. It includes Mosaic AI Model Training (pre-training and fine-tuning on private data), Mosaic AI Model Serving (managed inference endpoints), Mosaic AI Gateway (rate-limiting, cost attribution, model routing), Mosaic AI Vector Search (managed vector index) and Mosaic AI Agent Framework. Naveen Rao, the MosaicML co-founder, leads the team as VP of Generative AI inside Databricks.
- What is DBRX and how is it different from Llama or GPT?
- DBRX is a 132-billion-parameter mixture-of-experts large language model released by Databricks in March 2024 under a permissive open-source licence. It is the in-house foundation model Mosaic AI recommends when customers want a strong model that runs entirely inside their own VPC. Unlike GPT (proprietary, OpenAI hosted) DBRX weights are downloadable from Hugging Face. Unlike Llama (also open weights) DBRX uses an MoE architecture that activates only ~36B parameters per token, giving it inference economics closer to a much smaller dense model.
- What are Databricks AI Functions?
- AI Functions are SQL functions you call directly from a Databricks notebook, SQL warehouse, dashboard or job — including ai_query(), ai_classify(), ai_extract(), ai_translate(), ai_summarize() and ai_gen(). They wrap Foundation Model API calls with the row-and-column semantics SQL expects, so a data engineer can classify or summarise a column of text the same way they would call LOWER() or COUNT(). Billing rolls into the lakehouse DBU charge — no separate model endpoint to provision or pay for.
- Does Databricks have a free tier or free trial?
- There is no perpetual free tier. New accounts receive a 14-day free trial with consumption credits that cover the full platform — Workspace, Mosaic AI Model Serving, Vector Search, AI Functions and Foundation Model APIs. The trial credit is enough to build a non-trivial prototype but expires after 14 days. Customers in academic or research programmes can apply for separate Databricks for Education credits.
- How does Databricks AI compare to Snowflake Cortex?
- Databricks AI and Snowflake Cortex are the two most-compared AI platforms native to a major data warehouse / lakehouse. Databricks brings a broader portfolio (full pre-training and fine-tuning via Mosaic AI, the open-source DBRX model, AI/BI Genie) and is the better fit for teams running mixed analytics + ML + GenAI on a single platform. Snowflake Cortex is the natural choice for teams whose data and governance already sit in Snowflake — it offers LoRA fine-tuning, an in-house Arctic LLM, Cortex Search and an inline COMPLETE function. Pricing is consumption-based on both sides (DBUs vs Snowflake credits). See the comparison table above for a row-by-row breakdown.
- Can I run Databricks AI inside my own cloud account?
- Yes — Databricks runs in the customer's AWS, Azure or GCP account (control plane is multi-tenant, but the data plane and Mosaic AI compute live in the customer VPC). Customer-managed keys, private link, IP allowlists and HIPAA-eligible deployments are available on Premium and Enterprise SKUs. For organisations that need a fully isolated deployment, a single-tenant option is offered as part of Enterprise contracts.