The vector database wars are ending where they started: in Postgres

In 2023, every AI architecture diagram included a dedicated vector database. In 2026, most of the ones we review include Postgres with pgvector, and the dedicated store appears only when the workload genuinely demands it.

The consolidation happened for unromantic reasons. Embeddings want to live next to the rows they describe. Joining vector similarity results against permissions, tenants, and metadata is a SQL problem, and doing it inside one database eliminates a synchronization pipeline that otherwise has to be built, monitored, and debugged.

Performance stopped being the differentiator. HNSW indexing in pgvector handles the single-digit millions of vectors that describe almost every mid-market corpus, with latencies indistinguishable from specialized stores at that scale.

The specialized engines still earn their place above roughly a hundred million vectors, or when filtered search at high recall is the core product. That describes a small fraction of the deployments we see.

The takeaway is a familiar one: infrastructure sprawl is a tax, and the boring consolidation move usually wins. Before adding a database to your stack for an AI feature, check whether the database you already run does the job. Increasingly, it does.