Here is the whole story in one sentence: the customer buying a database is now, more often than not, a program.

Neon says roughly 80% of the databases it provisions are created by AI agents, not people. Supabase says more than 60% of its new projects are launched by AI tools, on 600% year-over-year growth. Once you internalize that number, the two strangest acquisitions of 2025 stop looking strange. Databricks paid about $1B for Neon in May. Snowflake answered three weeks later with roughly $250M for Crunchy Data. Two analytics companies, a combined $1.25B, both spent on the least glamorous thing in the stack: transactional Postgres.

They weren't buying a database. They were buying the front door that agents walk through to create one.

Why a warehouse suddenly needs an OLTP database#

Databricks and Snowflake are built for analytics — big, read-heavy, batch queries over historical data (OLAP). That is the wrong shape for an agent. An agent writing memory, checkpointing a long-running task, or storing a user's state needs the opposite: small, fast, transactional reads and writes (OLTP), on demand, with sub-second provisioning. That is exactly what Postgres does and warehouses don't.

So when agents became the fastest-growing category of database customer, the warehouses had a gap where their operational layer should be. Buying it was faster than building it. Databricks took Neon — serverless Postgres that spins up in under half a second, scales to zero, and branches copy-on-write like Git. Snowflake took Crunchy Data, the enterprise-grade, Kubernetes-and-on-prem Postgres, aimed at the regulated buyer rather than the vibe-coder.

The moat was never the SQL. Postgres is Postgres. The moat is the provisioning API — the thing an agent calls to conjure a database in 500 milliseconds — and the branching model underneath it.

The $188B round is the receipt#

It would be easy to file the 2025 deals under "acquihire" and move on. The July 2026 money says otherwise. Databricks' Neon technology, rebranded Lakebase and folded into the lakehouse, reached general availability on February 9, 2026 and — by Databricks' own account — is adopting more than twice as fast as its data-warehouse product, the business the company was built on.

Then, on July 16, 2026, Databricks signed a strategic round at a $188 billion valuation — roughly $3B, led by Coatue. Read the use-of-funds and Lakebase is right there next to Genie (its agent coworker) and the Unity AI Gateway. The operational database is no longer a bolt-on. It is one of the three things a $188B company is raising money to accelerate. That is what a vindicated bet looks like on a term sheet. (It's also the through-line behind the number founders read in this week's Wire.)

Supabase is the independent pole#

Not everyone got bought. Supabase — the Postgres-plus-Auth-plus-Storage backend that became the default target for AI-generated apps out of Bolt, Lovable, Cursor, and Claude Code — is the third pole, and the only one still independent. Its growth mirrors Neon's: agents, not humans, are creating the databases. That independence is a feature if you don't want your operational data living inside a warehouse vendor's gravity well, and a question mark if you're betting on who's still standing in three years.

What founders should actually do#

Three moves, in order.

  1. Assume your agents are already creating databases faster than you're tracking, and govern it now. If a coding agent can call a provisioning API, it will — and an ungoverned sprawl of forgotten Postgres instances is both a cost line and an attack surface. Put a budget and an inventory on it before it's a hundred instances.
  2. Pick your Postgres by where agent state belongs, not by the logo. If your agents' state should sit next to your analytics and governance, that's Lakebase. If you want cheap, independent instances that scale to zero between bursts, that's Neon. If the database rides with your app backend and you want auth and storage in the box, that's Supabase. (We break the choice down in Lakebase vs Neon vs Supabase for AI agents, and the broader field in Neon vs Supabase vs Turso.)
  3. Don't over-index on the SQL — index on the provisioning and branching. The reason these companies were worth $1.25B is that agents live and die by how fast they can get an isolated, disposable copy of a database. Copy-on-write branching and sub-second cold starts are the features that matter in an agent loop; the query planner is table stakes.

The larger pattern is the one worth remembering. For a decade the database was where humans put data. In 2026 it's increasingly where programs put state — and the companies that saw it first spent a billion dollars to stand at the door.