Vector Databases: The 2024 Shakeout
Pinecone, Weaviate, Qdrant, pgvector - who's still standing?
The vector DB market exploded in 2023, consolidated in 2024. Here's where each player stands and which one I'd pick for what.
Two years ago, "vector database" was a category that didn't exist. By late 2023, there were 20 of them. By late 2024, four matter for production.
The four
Pinecone. Managed-only, fully-hosted SaaS. Premium pricing. Mature operations. The default choice if you want to forget about ops.
Weaviate. Self-host or managed. More flexible than Pinecone. Stronger hybrid (vector + keyword) story.
Qdrant. Open-source, fast Rust core. Self-host friendly. The choice if you want control.
pgvector. Postgres extension. Use this until you can't.
Who I'd pick for what
- Up to 10M vectors, simple architecture: pgvector. No new operational surface.
- 10-50M vectors, want managed: Pinecone.
- 10-50M vectors, want flexibility: Qdrant.
- Hybrid (vector + keyword) is critical: Weaviate.
- >50M vectors: real conversation. Probably Pinecone if you can pay; Qdrant or Weaviate if you can't.
What killed the also-rans
Most of the 2023 vector DB startups have either pivoted or shrunk. Why?
- pgvector improved fast - it's now within shouting distance of specialty DBs for most workloads
- Pinecone matured - the gap between "managed Pinecone" and "self-host alternative" widened
- Demand for "JUST a vector DB" softened - most production systems need hybrid retrieval, structured filtering, and observability
The companies that survived are either:
- Best-in-class on a specific axis (Pinecone on managed ops, Qdrant on self-host performance)
- Built into a broader product (Weaviate's hybrid, pgvector's ubiquity)
What I'd build today
If I were starting today:
- Day 1: pgvector inside the same Postgres
- Day 100 (if you've crossed scale limits): migrate to Pinecone or Qdrant
The migration cost is real but manageable. Don't pay for premium vector infra before you need it.
What's next
I expect more consolidation. The remaining four will all be fine; some of them will be acquired. Bet on the protocol (your application code) being portable, not the vendor.