Every big company kept its data twice; making the cheap half answer moved the expensive half's budget

The product

Databricks sells the work done on a company's data without selling the place the data is kept. The files stay in the company's own cloud storage, in formats anyone can read; Databricks supplies the engines that run against them – pipelines, SQL and reporting, machine learning, and now applications and agents. It was founded in 2013 by the group that, in its own description, created Apache Spark, and by August 2026 it was past a $7 billion revenue run-rate, growing more than 80% year over year, with more than 20,000 organisations on it, among them adidas, AT&T, Bayer, Block, Mastercard, Rivian, Unilever and 70% of the Fortune 500.

How it works

Signing up takes no card: the free trial opens the platform, and some trials come with credits to spend on it. Configure Databricks against your own cloud account, as the pricing page assumes, and the cloud provider goes on billing you for the machines itself.

What sits on top has grown section by section. Data engineering handles the loading and the scheduling, for batches and for streams. The product now called Lakehouse is a serverless data warehouse for SQL analytics on that same open lake data. Unity Catalog governs who may see what, across clouds, formats and platforms. Then the newer floor: Agent Bricks for building agents, Lakebase – serverless Postgres – for the data an application reads and writes while it runs, Genie to answer questions in words, and Unity AI Gateway to route between models and cap what they cost. The homepage today opens not on the lakehouse at all but on Lakebase: "One database for AI, apps and agents."

Business model

Consumption, metered in DBUs – a Databricks Unit, the company's normalised unit of processing power. There are no named tiers. Accounts are pay-as-you-go, with no up-front cost and billing "at per second granularity", and customers who can commit to a volume sign a committed-use contract for a discount. The cloud bill is the customer's own and arrives from the cloud provider.

The shape follows from what Databricks does not own. It does not hold the asset, so it cannot charge rent on it; it can only charge for the work, and work is divisible. That lets one analyst's query and a bank's entire reporting stack be the same product at different volumes – by August 2026 more than 1,000 customers were consuming at over $1 million a year and more than 100 at over $10 million.

The spike

In January 2017 the homepage said plainly what the company was: "The Best Place to Run Apache® Spark™", with Apache Spark in the navigation bar. Two years later the release for its $250 million Series E named what it was meeting: "Explosive Growth and Global Demand for Unified Analytics". Both describe one business – the place where a company's data engineers and data scientists worked on the cheap heap of files in its cloud storage. The questions the business itself asked, the reports and the dashboards, were answered somewhere else: in a data warehouse holding a curated second copy, loaded out of the lake by pipelines somebody had to keep running.

Databricks' founders wrote down what they thought of that arrangement, in a paper at CIDR in January 2021. Data, they wrote, "is first ETLed into lakes, and then again ELTed into warehouses, creating complexity, delays, and new failure modes"; keeping the two consistent "is difficult and costly"; the warehouse copy is stale, "with new data frequently taking days to load"; and when someone wants to train a model, the warehouse vendors tell them to export files, "adding a third ETL step!" The two-tier pattern, they noted, was in use "at virtually all Fortune 500 enterprises". Their conclusion was not modest: the data warehouse "will wither in the coming years".

The turn was to stop being the cheaper half of that pair. Databricks gave up the complement's position – feeding a system it did not sell, for a budget that was not its own. The lake stayed, as it always had, the customer's. To the files already sitting in it Databricks added the two things the warehouse had been holding over them: transactions, shipped as Delta Lake in April 2019 and handed to the Linux Foundation that October, and speed. In November 2020 it put SQL Analytics on top. It meant, Ali Ghodsi told TechCrunch, that "SQL Analytics enables you to now do the data warehousing workloads directly, and concretely, the business intelligence and reporting workloads, directly on the data lake". A year later the same engine held the audited TPC-DS record at 100TB.

What the company describes is an architecture correcting itself. Read from the money, almost nothing about the product changed and everything about the buyer did: the same files, in the same storage, sold to the department that signs for reporting instead of the one that signs for data science. The hard part was not the migration, because there wasn't one. Even the word was borrowed rather than planned – "We've been saying Unified Data Analytics," Ghodsi said that month, "but no one picked that up. But 'lake house' caught on as a term."

Why it matters

Both alternatives were real and neither was free of a catch. The cloud data lake – S3, ADLS, GCS – was cheap, open and would hold anything, but, as the lakehouse post put it in January 2020, it did "not support transactions", did "not enforce data quality", and its "lack of consistency / isolation" made it "almost impossible to mix appends and reads". The warehouse – Teradata in the older generation, Redshift or Snowflake in the cloud – was fast and trustworthy and closed. So companies ran both, and the step neither of them removed was the one in between: choosing what to copy, building the load, and re-running it forever. The paper cites a Dimensional Research and Fivetran survey in which 86% of analysts were using out-of-date data and 62% reported waiting on engineering resources numerous times a month.

That the copying stopped shows up in what companies buy instead. Revenue run-rate went from over $200 million in the third quarter of 2019 to over $350 million a year later; ARR was $425 million in February 2021 and past $600 million by that August, as reported by TechCrunch on the two rounds. Data warehousing – the part that did not exist before November 2020 – passed a $1 billion revenue run-rate by December 2025 and $1.5 billion by August 2026, growing more than 100% year over year – faster, by then, than the company as a whole. Net retention has stayed above 140%: the same customers hand over more each year, which is what moving a second system's work onto the first one looks like on an invoice. By February 2026 CNBC could observe that Databricks was larger than Snowflake, which had reported $1.21 billion of revenue in its October quarter.

The playbook

When customers already keep an asset in cheap, open storage, a supplier can sell just the work done on it, metered by the second, so one analyst's query and a bank's reporting are one product at different volumes. The sign elsewhere: the asset sits in one place and a second vendor is paid to hold a working copy. Databricks has kept net retention above 140%, customers buying more work each year. The bet fails if the open store stays slower than the closed one.

The lesson

Databricks spent its first six years selling into a chain it did not control. Its customers were paying for their files twice: once to keep them, once more to keep a usable duplicate in a system that could actually answer a question. Rather than build a better version of that second system and ask everyone to move, it made the original answer – and the money that had been paying for the duplicate had nowhere else to go. The rule is available to anyone: when your customers pay to keep a thing and pay again for a duplicate that only somebody else can work with, make the thing itself work where it already lies. Framer, elsewhere on this site, reached the same rule by removing a different step.

The second decision is what made the first one acceptable. A company will only consolidate onto one supplier if leaving is cheap, and Databricks went out of its way to make leaving cheap: open file formats, Delta Lake given away to a foundation, the creators of the rival format bought and their work kept open, the governance layer open-sourced, and the data itself left in the customer's own account. Its own paper says why this matters to a buyer – an open format "means that they will always have direct access to the data without blocking on a vendor". Being easy to abandon is what persuades a cautious customer to give you everything.

Neither move is free: the first means proving yourself on an incumbent's own benchmark, the second that you own no lock and must win every renewal on merit. Which of your customers keeps a second copy of something they already hold, and who do they pay to hold it? And if one wanted to leave next month, what could they not take with them?