Why the Lakehouse Vendors and the Clouds Bet on Postgres for AI Applications
Databricks, Snowflake, Google, Amazon, and Microsoft have all built AI-facing products on PostgreSQL, and this piece explains why the open license, extension model, and developer familiarity made it the shared choice. It also warns that "Postgres" now covers four different levels of compatibility, so buyers need to check which one they're getting.
- Databricks paid around $1 billion for Neon and Snowflake paid around $165 million for Crunchy Data, showing the premium the market puts on real Postgres compatibility plus modern architecture.
- Extensions like pgvector, pg_lake, and pg_duckdb turned Postgres into a single engine for state, vector retrieval, and analytics, which suits most AI workloads below very large scale.
- "Postgres compatible" ranges from the real engine to a wire protocol over a different database, and the extensions you rely on determine which layer you need and what leaving will cost.
Five major players have converged on Postgres as the database under their AI applications: Databricks (which paid around $1 billion for serverless Postgres company Neon), Snowflake (around $165 million for Crunchy Data), and Google, Amazon, and Microsoft through AlloyDB, Aurora, and HorizonDB. The reasons are practical. The PostgreSQL license is permissive enough that a vendor can fork it and sell a proprietary service without owing anything back, and no single company controls the project, so the license can't be pulled out from under users the way Elasticsearch, MongoDB, and Redis licenses were changed. Nearly every developer already knows Postgres, which removes adoption friction. And the extension model let the community bolt on the capabilities AI workloads need: pgvector for vector search (good for most workloads under about a hundred million vectors), pg_lake from Snowflake for reading and writing Apache Iceberg tables, pg_duckdb and Aurora's embedded DuckDB for columnar analytics next to transactions, and ParadeDB for full-text search. The vendors aren't abandoning their other engines. Google still invests in Spanner, Microsoft in SQL Server and Fabric, Amazon in DynamoDB. Postgres is simply the choice for applications that want a relational store the whole ecosystem already speaks.
The part worth paying attention to is that "Postgres" now covers four very different things, and the differences matter for lock-in. At the top, Supabase, Cloud SQL, Snowflake Postgres, and Lakebase run the real PostgreSQL engine, so every extension works and leaving means a pg_dump. Next, YugabyteDB keeps the Postgres query layer on distributed storage, which gives you consistency across nodes but behaves differently underneath. AlloyDB and Aurora run modified engines with proprietary storage, so most extensions work but getting your data out means moving it, not just exporting files. At the bottom, CockroachDB and Spanner's Postgres interface only speak the wire protocol and dialect subset, so drivers connect but many extensions don't apply. The extensions you rely on today should determine which layer you need, because the layer determines what switching will cost you later.
Postgres has real ceilings. A single primary can only scale writes so far, and beyond that you're choosing between a Postgres-compatible distributed engine, a distributed engine with its own dialect like OceanBase, SingleStore, or TiDB, or a different data model entirely. Very large analytical scans and billions of embeddings are better served by a dedicated warehouse or vector database. The useful question before committing is not "do we want Postgres?" but "which layer of Postgres, which extensions do we depend on, and what does our workload's write rate and data volume look like?" For most AI applications, one engine with the right extensions is enough, and the operational simplicity counts for a lot.