Databricks acquires Electric to bring Postgres to AI agent workflows
Databricks just bought the company that puts a full Postgres inside WebAssembly, because AI agents do not want one shared database. They want thousands of tiny ones they can spin up, branch, and throw away.
What Databricks acquired
Per The New Stack, Databricks is acquiring Electric, the startup behind PGlite (Postgres compiled to WASM) and the Electric sync engine. The Electric team will join Neon, the serverless Postgres company Databricks bought last year and the base of its Lakebase service. Deal terms were not disclosed.
PGlite is a complete Postgres that runs in the browser, in Node, or inside the sandboxes agents use to execute code. It can load extensions dynamically, including pgvector. Weekly downloads jumped from about 1 million to 13 million over the last year.
The sync engine is the other half of the bet. It keeps a central Postgres in near-real-time sync with browser tabs, mobile apps, or agents. Neon’s team frames the hard parts as conflict resolution, partial replication, and reconnection logic. Those are the pieces teams rarely want to rebuild.
Open source stays open. Founders James Arthur and Valter Balegas said previously released OSS (the sync engine, PGlite, Durable Streams, TanStack DB) remains open source. Electric Cloud is winding down. Hosted users will need to self-host or move.
Agents already treat databases like git branches
This purchase tracks how Lakebase is actually used. Databricks engineers Ippokratis Pandis, Nikita Shamgunov, and Reynold Xin have written that agents create roughly four times more databases than human users on Lakebase. The average project carries about 10 database branches. Some projects run more than 500 branch iterations deep. For some Lakebase application types, average database compute lives under 10 seconds.
That is not a human SaaS pattern. Traditional apps share one long-lived database among many clients. Agents branch databases the way they branch code: spin a sandbox, instantiate PGlite, build and test, then discard the instance or sync the result into a Lakebase branch. Because Lakebase separates storage from compute and stores data in Postgres page formats on object storage, that branch is a cheap copy-on-write metadata move.
Neon’s team put the economic point bluntly: when coding agents drive creation cost toward zero, most new apps are small. A server with a fixed minimum compute floor fights that abundance. A WASM Postgres that can live for seconds does not.
Two halves of the same idea
PGlite did not start at Electric. It began as an experiment by Neon co-founder Stas Kelvich, who compiled Postgres to WASM to see if it could run client-side. Electric turned that work into a production project. Databricks frames the deal as reuniting those two halves. The company has also been stacking database bets: Neon, then transactional startup Mooncake, now Electric.
Disposable Postgres is the new unit of work
If you ship coding agents, eval harnesses, or multiplayer data apps, the scarce resource is no longer “a Postgres.” It is cheap, disposable Postgres instances plus a sync path back to something durable. Databricks is wiring that stack under Lakebase instead of letting every agent framework invent its own toy DB. Builders who leaned on Electric Cloud get a migration deadline. Builders who already self-host the OSS pieces keep their path.
The right buy for the agent era
This is the right acquisition for the agent era, even with undisclosed terms. Agents already create ~4x more databases than people on Lakebase. Some of those databases live less than 10 seconds. Buying PGlite and the sync engine is Databricks admitting the unit of database work is no longer the long-lived app DB. It is the ephemeral branch. Reuniting Kelvich’s WASM experiment with Neon’s branching model is not nostalgia. It is product strategy catching up to how agents already behave.
Lakebase path after Electric Cloud dies
Watch how fast Electric Cloud customers get a Lakebase migration path, and whether PGlite + Electric sync shows up as a first-class Lakebase/Neon feature for agent sandboxes rather than a side project. Also watch whether the 13M weekly download curve holds once the hosted product is gone.
