HubSpot Scales Vector Database Infrastructure to 20B+ Vectors with In-House Qdrant
Key Takeaways
- ▸HubSpot's VaaS platform has scaled from a 2023 POC to storing 20+ billion vectors across 140+ clusters in 5 regions, serving 38+ teams and 200+ indexes
- ▸The platform handles extreme traffic spikes: 5,000+ RPS for writes (up to 100,000 RPS during backfills) and 1,000+ RPS for reads, with cluster isolation by product team to minimize blast radius
- ▸In-house Qdrant deployment enables cost optimization through corporate AWS rates, full data control, and tight integration with HubSpot's database infrastructure expertise and security practices
Summary
HubSpot has built VaaS (Vector-as-a-Service), a centralized platform for vector storage and semantic search that has grown from a proof-of-concept to critical infrastructure handling 20+ billion vectors across multiple regions. The platform, built on top of the open-source Qdrant vector database, powers AI agents, retrieval-augmented generation (RAG), contact deduplication, and semantic search across 38+ teams and 200+ indexes. VaaS operates as an API layer in front of Qdrant, handling embeddings generation, access control, and request routing across a fleet of 140+ clusters deployed in 5 regions and 2 environments.
The platform currently processes 5,000+ requests per second for writes (spiking to 100,000 RPS during large backfills) and 1,000+ RPS for reads. The largest single index stores 9.5 billion vectors, with an average index size of 95 million vectors. By running Qdrant in-house rather than using a managed service, HubSpot gained control over costs through corporate AWS rates, security compliance, infrastructure tuning, and seamless integration with existing internal tooling for tracing, monitoring, rate limiting, and cost tracking—reflecting a strategic decision to own AI retrieval infrastructure as it becomes essential to enterprise AI operations.
- Vector search has become foundational to HubSpot's AI strategy, powering agents, RAG systems, and semantic use cases—demonstrating the criticality of reliable vector infrastructure at enterprise scale
Editorial Opinion
HubSpot's investment in in-house vector database infrastructure signals a broader enterprise trend: as semantic search and RAG become core to AI products, companies are building purpose-built infrastructure rather than relying solely on managed services. This decision trades operational complexity for control—something HubSpot is well-positioned to handle given its existing database expertise. The case study underscores a maturing market reality: vector databases are transitioning from experimental tools to essential infrastructure, and enterprises with scale are increasingly choosing to own and operate them in-house for cost, security, and performance reasons.
