Vector Database Market Splits as Postgres Takes Smaller Workloads

From category to data type

For several years, adding retrieval to an AI application usually meant adding a dedicated vector database to the stack. In 2026 that default no longer holds. Coverage from across the industry now describes vectors as a data type built into general-purpose databases, not a separate product category. VentureBeat's year-ahead outlook argued that 2025 had already made this clear, and pointed to Amazon S3's new ability to store vectors as further evidence that standalone vector stores were becoming optional18. Commentary published in March framed the shift the same way. PostgreSQL (through pgvector and pgvectorscale), MongoDB Atlas Vector Search, and Oracle's database line had all made vector support a native feature, and the major cloud providers had followed11.

The more interesting point is what the specialist vendors did next. They did not fade away. Instead, they spent the year shipping features aimed at problems general-purpose databases still handle poorly: agent workloads, very large scale, edge deployment, and memory efficiency. At the same time, the streaming-data world began building vector search directly into its own processing engines. The result is a market splitting into tiers, and the reporting broadly agrees on where those tiers sit.

Postgres takes the low end

The numbers most often cited for Postgres are striking, but their origin matters. One widely repeated benchmark has pgvectorscale reaching about 471 queries per second at 99% recall on 50 million vectors, against roughly 41 QPS for Qdrant on identical AWS hardware13. The same table shows Qdrant 39% faster at P95 latency and 48% faster at P9913. A later analysis noted that TigerData, a Postgres-focused company, produced the pgvectorscale result using a fork of ANN-benchmarks. It also pointed out that Qdrant, Zilliz, and Redis each publish their own benchmark suites12. Put simply, the headline result shows that Postgres can win on throughput under one vendor's test conditions, while the specialist engine still wins on tail latency.

The practical threshold is more consistent across reports. One analysis estimated that pgvector now handles workloads below about 50 million vectors and cuts total cost of ownership by 40 to 60%12. The same analysis said latency climbs from around 50ms to 800ms once a deployment passes 10 million embeddings, even on a large instance12. Practitioner guides put the comfortable default lower, recommending pgvector for retrieval applications under about 10 million vectors9. One comparison of agent workloads found that under 10 million vectors, all eight databases it tested performed adequately3. Postgres also has a structural weakness. Its multiversion concurrency control (the way it handles simultaneous reads and writes) is inefficient for write-heavy workloads, so tables and indexes bloat under constant writes11. That matters because write-heavy is exactly what agent workloads tend to be.

Specialists ship for agents and scale

Release activity from the purpose-built vendors has been heavy. Qdrant raised a $50 million Series B on March 12, passed 250 million downloads, and shipped version 1.1712. Its May 1.18 "TurboQuant" release added a quantization method derived from Google Research. Qdrant says it matches scalar quantization's recall using about half the memory, and the release also lets users add or remove named vectors without rebuilding a collection20. In April, Qdrant Cloud launched GPU-accelerated indexing with up to 4x faster HNSW index builds19. Qdrant also introduced Qdrant Edge, a lightweight embedded engine for robots, point-of-sale systems, and phones14.

Weaviate's v1.37.0, released April 16, added a built-in Model Context Protocol (MCP) server in preview. This lets AI agents talk to the database directly, and the release also brought diversity search, query profiling, and incremental backups20. Milvus 2.6, focused on lowering cost at scale, became generally available on Zilliz Cloud on January 204. Pinecone, meanwhile, released Nexus and a declarative query language called KnowQL. KnowQL lets agents state intent, latency budgets, and confidence requirements instead of working with raw vectors12. Pinecone's internal benchmark on SEC filings claimed task completion rising from 50–60% to above 90%, along with a 95% cut in LLM token usage. One analyst noted that these claims still need independent validation12.

Pinecone's position is less clear than its product releases suggest. According to reporting summarized in one review, the company that arguably created the vector database as a product had hired bankers to explore a sale and replaced its founder-CEO with an executive from Google20. Analysts were also predicting two to three acquisitions in the category before year-end19. Read alongside Qdrant's funding round, this points to consolidation that rewards open-source engines with flexible deployment more than managed-only offerings. That is an inference, but the reporting supports it.

Release details don't always agree

Coverage of specific releases is uneven, and readers should check version claims against vendor release notes. One comparison describes Qdrant's GPU-accelerated HNSW indexing and Multi-AZ clusters as arriving in "v1.14 (April 2026)"2. That conflicts with reporting that Qdrant had already reached 1.17 by March12. Another guide lists Qdrant at v1.12 and Weaviate at v1.28 as their 2026 releases7. Neither matches the v1.37 and v1.18 releases documented elsewhere20. Benchmark tables differ even more. One independent-style test found Pinecone Serverless the cheapest option at 1 billion vectors8, while another guide rated Pinecone the most expensive tier6. The safest conclusion is that cost depends heavily on workload shape and query volume.

The edge-deployment story deserves the same care. Much of the argument that cloud-native vendors are neglecting offline, air-gapped, and regulated deployments comes from Actian, which sells a product built for that gap1112. The underlying pressure appears real. Data residency rules and EU AI Act penalties of up to €35 million or 7% of global revenue are cited as making cloud-only vector stores difficult for regulated industries13. The AI Act's high-risk system rules were expected to take full effect in August 202619. Qdrant Edge suggests specialists are moving to compete for these deployments too14.

Agents change the workload

The sources agree most strongly that AI agents are changing how retrieval systems are used. Multiple analyses claim agents issue about 10 times more queries than humans1112. They also write frequently, which causes index bloat, traffic spikes, and pressure on multi-tenant isolation that read-optimized indexes were not built for12. Recommended fixes include semantic caching (reusing answers to similar earlier queries), exposing index parameters so engineers can tune speed against recall, and connection pools that resize dynamically with demand11.

This explains why the 2026 feature lists look alike. Production-grade now means hybrid search, metadata filtering, multi-tenant isolation, and the ability to re-embed data in place as embedding models change faster than annual release cycles17. Eight of the ten databases in one roundup now support hybrid search out of the box17. Quantization has moved from research into production, with memory savings of 4x to 32x for small recall losses1716.

Streaming engines take on retrieval

The least-discussed change may matter most over time: stream processors are starting to perform retrieval themselves. Apache Flink 2.2, released in December 2025, added a VECTOR_SEARCH function for streaming similarity search inside Flink SQL. It joins ML_PREDICT, an earlier function for LLM inference23. The Flink project said that before this, the engine could generate embeddings and send them to downstream storage but could not run similarity queries itself23. Instaclustr describes this as enrichment, scoring, and retrieval happening while data is in motion, not in a separate batch step26.

Streaming analyst Kai Waehner lists "streaming powers agentic AI" among the six data-streaming trends for 2026, alongside diskless Kafka, Apache Iceberg storage, and regional deployments driven by sovereignty rules27. Reference architectures from Conduktor and others describe Kafka topics carrying raw events, processors calling embedding models, and change data capture keeping vector stores such as Pinecone, Weaviate, and Qdrant current in real time2821. Engineering write-ups emphasize exactly-once processing via Flink checkpoints and idempotent upserts keyed on vector IDs to keep indexes consistent with the event log2224.

What it adds up to

The vector database is not dying, but it is losing its claim to be the center of the AI stack. Postgres and other general-purpose systems are absorbing the small-to-mid workloads where most retrieval projects begin. Specialists are pushing toward billion-scale deployments, agent protocols, memory-efficient quantization, and on-premises or edge installs. Streaming platforms are building retrieval into the pipeline itself. Growth forecasts remain healthy, with one projection putting the market at $3.73 billion in 2026 rising to $10.6 billion by 203212. That growth increasingly favors vendors that fit into existing data infrastructure, not ones that try to replace it. Teams choosing a stack this year should start with their scale, how write-heavy their workload is, and where the data is allowed to live, and treat vendor benchmarks as marketing until they can reproduce them.