Vector Search Becomes a Native Database Feature, Yet Cost Math Persists
Dedicated vector databases have largely been absorbed into mainstream engines that now offer vector columns and approximate nearest-neighbor indexes. This consolidation removes an operational layer but not the underlying memory, compression, and RAM costs that determine spending at scale. The article argues that the economics depend on data representation rather than the database vendor.
Dedicated vector stores such as Pinecone, Milvus, Weaviate, and Qdrant appeared around 2021, when mainstream engines lacked usable ANN indexes. That gap closed as Elasticsearch, OpenSearch, Postgres via pgvector, MongoDB, Redis, and ClickHouse added HNSW-based search, often because HNSW reaches high recall without rebuilding.
Cost remains tied to RAM. HNSW traversal follows pointers, so disk spills shift latency to page faults. OpenSearch documents a per-vector estimate of 1.1 times the sum of four bytes per dimension and eight bytes per graph connection; one million 1,536-dimension float32 vectors consume roughly 6.1 GB as raw payload before graph overhead.
Teams building AI search may benefit from fewer systems to operate, but memory costs could still shape access. Smaller organizations may find consolidated features easier to adopt, while large-scale semantic search may remain expensive because of RAM and dimensionality. Database vendors may compete less on novelty and more on efficiency. Users could see faster, more integrated search experiences, though cost pressures may influence which services are offered and at what price.