Multiple winners in vector databases
Jeff Tang, CEO of Athens Research, on Pinecone and the AI stack
The key implication is that vector databases look more like databases or observability tools than like social networks, where one company takes nearly everything. Different buyers want different things, some want the easiest hosted tool for a prototype, some want cloud native options bundled into AWS or Google Cloud, and some just add vector search to Postgres with pgvector. That leaves room for several large winners, even if no single one owns the category.
-
Pinecone sits in a horizontal infrastructure layer. Teams store embeddings, run similarity search, and feed the top matches back into an LLM for RAG. But the actual applications differ a lot, from receipts to Jira tickets to recommendations, so one generic app layer rarely fits every workflow. That variety supports a broad market with many product niches.
-
The market is also structurally fragmented by distribution. Pinecone sells a standalone managed database, AWS offers vector search inside OpenSearch, Google offers Vertex AI Vector Search, and developers can use pgvector inside ordinary Postgres. When buyers can get acceptable vector search from several existing stacks, share spreads across multiple vendors instead of collapsing to one.
-
What decides winners is less a hard moat and more execution at a specific layer. Pinecone argues databases can still become sticky through reliability, performance, and switching costs, while builders like Jeff Tang often choose tools based on developer mindshare, ease of setup, and speed to prototype. In practice, brand, distribution, and product quality matter more than exclusivity.
Going forward, the category should keep splitting into a few durable layers. Cloud platforms will bundle vector search, open source will cover the default path, and specialists like Pinecone will compete by being faster, easier, and more production ready. That is how a market gets very big without becoming winner take all.