Skip to main content

Databricks is a cloud platform where a company stores, analyzes, and builds AI on all of its data in one place.

Companies whose data is split across cheap raw storage and a separate, more rigid analytics system, and who want one platform instead of two.

Companies often run two separate systems, a cheap place to dump raw data and a separate, reliable system for analyzing it. Databricks removes the need to keep both in sync.

The lakehouse architecture

Companies used to run two separate systems for their data. A cheap data lake held raw files but wasn't reliable for serious reporting, and an expensive data warehouse held clean data but couldn't easily store raw files. A lakehouse is one system that does both jobs, so data never has to be copied between the two.

A retailer's website logs, sensor data, and sales records all land in one system. The same data can then feed a machine learning model and a finance report without being moved or duplicated.

It's like living in a house with one open-plan room instead of a garage and a separate formal dining room. Everything is under one roof, so nothing has to be carried from one room to the other before it's usable.

The technical details

Databricks originated and popularized the term "lakehouse." A data warehouse offers low query latency and high reliability for business intelligence. It often relies on proprietary formats that limit machine learning support, though.

A data lake stores data of any format cheaply, but lacks the validation that business intelligence reporting needs. A lakehouse pairs an optimized metadata layer with validated data stored in cloud object storage. That combination gives open, direct access to data in standard formats, alongside the indexing and reliability that both machine learning and business intelligence need.

Delta Lake

Delta Lake is what makes raw files in cloud storage behave like a proper database instead of a folder of loose files. A write either fully completes or doesn't happen at all, and the data has to match an expected structure. Every past version is also kept, so you can see what the data looked like before a change.

Two pipelines write to the same table at the same time. Without Delta Lake, that can corrupt the data. With it, both writes complete safely, or one is rejected cleanly.

The technical details

Microsoft's own Azure Databricks documentation describes Delta Lake as "an optimized storage layer that supports ACID transactions and schema enforcement," ACID meaning atomicity, consistency, isolation, and durability. As raw files are converted into Delta tables, schema enforcement checks for missing or unexpected data. Schema evolution then lets the structure change later, without rewriting the downstream logic.

Delta Lake is open source, released by Databricks under the Apache License 2.0. It was fully donated to open source in June 2022, alongside MLflow.

Unity Catalog

Unity Catalog is the single place where a company controls who can see and use its data, models, dashboards, and AI agents. Instead of setting permissions separately in each tool, administrators set rules once in Unity Catalog and they apply everywhere.

A compliance team needs to know exactly which reports and models were built from a customer table containing personal data. Unity Catalog tracks that lineage automatically, so they don't have to trace it by hand.

The technical details

Databricks' own materials describe Unity Catalog as governing data isolation boundaries and lineage tracking as data is transformed. It applies one unified model that keeps sensitive data private and secure across data, models, dashboards, and agents.

Databricks open-sourced Unity Catalog in June 2024 under the Apache License 2.0, the same license family as Delta Lake. It now has its own independent home at unitycatalog.io, separate from the databricks GitHub organization.

Problems it solves

Keeping raw data and clean data in sync, without running two systems

Companies used to run a data lake for cheap raw storage and a separate data warehouse for clean, structured analysis. Then they had to copy data between the two. Databricks combines both into a lakehouse, one system that holds both kinds of data, so nothing has to move between systems to be usable.

Getting database-grade reliability on data sitting in ordinary cloud storage

Plain files in cloud storage do not guarantee that a write finishes cleanly or that the data matches an expected structure. Delta Lake adds those guarantees to ordinary files in S3, ADLS, or GCS. A write either fully completes or does not happen at all, and every past version is kept.

Running analytics and machine learning against the same data

Keeping separate copies of data for BI reporting and for machine learning means the two can drift out of sync. Delta Lake and Unity Catalog sit underneath both, so a report and a model can read the exact same underlying tables, not two different copies of them.

Tracking who can access data, and where a report or model came from

Without a shared governance layer, an organization has to set permissions separately in every tool. Tracing which report or model was built from which data then means checking each system by hand. Unity Catalog governs access and tracks lineage automatically, in one place, across data, models, dashboards, and agents.

Competitors

Snowflake, Google BigQuery, Amazon Redshift, Microsoft Fabric

Pricing

Consumption-based (DBUs), no published starting price. From ~$0.02/DBU, per third-party sources

Business model

Private company. $134B valuation, Series L, Dec 2025

Links

Find it
Homepage, Docs, GitHub, X
License
Proprietary (platform); several underlying open-source projects use Apache License 2.0

Going deeper

What happens to data once it lands in Databricks?

Databricks describes a specific, documented pipeline for how data moves from raw files to something a business can query and build AI on.

The pipeline, in order

Data lands first in its raw, native format in the customer's own cloud object storage, such as Amazon S3, Azure Data Lake Storage, or Google Cloud Storage. Databricks does not require moving that data into a proprietary store it controls.

Delta Lake then wraps that storage with ACID transactions and schema enforcement. This converts raw files into reliable "Delta tables." Apache Spark is the compute engine that runs queries and transformations against that data, decoupled from where it's stored. Unity Catalog governs access, lineage, and audit across all of it, in one place.

Microsoft's own Azure Databricks documentation describes this staged movement, from raw to curated to business-ready, as a pattern the industry calls the "medallion architecture." The saved research for this explainer does not detail its layers further.

A peer-reviewed origin

The term "lakehouse" comes from a peer-reviewed paper Databricks' own founders published, titled "Lakehouse: A New Generation of Open Platforms that Unify Data Warehousing and Advanced Analytics." It was presented at CIDR (the Conference on Innovative Data Systems Research) in 2021, by Michael Armbrust, Ali Ghodsi, Reynold Xin, and Matei Zaharia.

Separating compute from storage this way also helps organizations avoid data egress fees. Their files stay in open formats like Delta Lake and Apache Iceberg, not a single vendor's proprietary format.

How does Databricks bill you?

Databricks charges per second of compute used, but the way that translates into a bill catches many teams off guard. Databricks' own pricing page renders its rate tables in JavaScript, not plain text. The dollar figures below come from Microsoft's Azure pricing page and the FinOps vendor CloudZero instead.

The billing unit and the two-bill structure

Databricks bills in DBUs (Databricks Units), a measure of processing power consumed per hour, charged per second of use. Cost is the number of DBUs used, multiplied by a dollar rate. That rate varies by compute type, cloud provider, region, and pricing tier.

A documented gotcha sits underneath that number. Customers receive one bill from Databricks for DBU consumption. A separate bill comes from their cloud provider (AWS, Azure, or GCP) for the underlying compute and storage infrastructure.

Teams that budget only for the DBU charge often underestimate their total spend. CloudZero puts the gap at 50 to 200 percent, since they overlook that second bill.

Tiers and discounts

Databricks sells Standard, Premium, and Enterprise tiers, though Standard is being phased out. It was sunset on AWS and GCP in October 2025. Azure is following by October 2026, leaving Premium and Enterprise as the tiers most organizations use.

Two mechanisms can lower the on-demand rate. DCUs (Databricks Commit Units) are a prepaid commitment that discounts all workloads across every cloud. Spot Instances can separately cut costs by up to 90 percent on batch and non-interactive workloads.

Databricks also offers a free 14-day trial and a limited, free Community Edition, per CloudZero's reporting.

What has Databricks open-sourced, versus just supported?

Databricks draws a distinction, in its own words, between open-source projects it originated and technologies it merely supports on its platform.

What Databricks originated

By Databricks' own account, its engineers are the original creators of Apache Spark, Delta Lake, Unity Catalog, MLflow, and Delta Sharing. MLflow manages the machine learning lifecycle, including experimentation and a model registry, and Delta Sharing is an open protocol for secure data sharing.

Delta Lake, MLflow, and Unity Catalog are not hosted under the databricks GitHub organization, though. They live in their own independent homes (delta-io, mlflow, and unitycatalog.io), reflecting their status as independently governed open-source projects rather than Databricks-controlled repositories.

What it acquired or just supports

Redash, an open-source data visualization tool, was acquired by Databricks in June 2020 rather than built in-house. Databricks also supports, without having created, Apache Iceberg (originated at Netflix) as well as widely used frameworks like TensorFlow, PyTorch, and scikit-learn.

Where did Databricks come from, and how has it grown?

Databricks was founded in 2013 by the people who built Apache Spark. It has raised money unusually fast, closing three major rounds in less than a year.

From a university lab to a company

The company grew out of the AMPLab project at the University of California, Berkeley. Its founders had built Apache Spark there, an open-source distributed computing framework.

Its founders are Ali Ghodsi, Andy Konwinski, Ion Stoica, Matei Zaharia, Patrick Wendell, Reynold Xin, and Arsalan Tavakoli-Shiraji. The company is headquartered in San Francisco. It remains privately held, with no public listing as of mid-2026.

Funding trajectory

Databricks raised a $13.9 million Series A in 2013, led by Andreessen Horowitz. Funding accelerated sharply in its more recent rounds.

A $10 billion Series J in December 2024 valued the company at $62 billion. A $1 billion Series K in August 2025 pushed that past $100 billion. A Series L of more than $4 billion in December 2025 set its valuation at $134 billion.

That was the company's third major fundraise in less than a year. As of June 2026, Databricks reported a $6.9 billion annualized revenue run rate, up 80 percent year over year. It has been free cash flow positive since 2025.

Sources

Last checked August 2026