turbopuffer

turbopuffer is a serverless vector and full-text search engine on object storage, offering very cheap cold storage (~$0.02/GB) and pure usage-based pricing, ideal for large multi-tenant search and RAG.

turbopuffer rethinks vector search economics by putting data on object storage (like S3) instead of keeping everything in RAM and SSD. Because most namespaces are cold most of the time, turbopuffer charges near-zero for inactive data and only bills meaningfully when data is queried. This yields an effective storage cost of roughly $0.02/GB — dramatically lower than legacy RAM-plus-SSD architectures — while a caching layer keeps hot queries fast. The engine supports both vector similarity search and full-text search, making it suitable for RAG, semantic search, and hybrid retrieval at large scale. Its architecture particularly favors multi-tenant applications with many namespaces where only a fraction are active at any moment, such as per-customer search indexes. Pricing in 2026 is pure usage-based: storage per GB-month, writes, and queries (with the queried-data rate reduced to $1/PB in February 2026), plus tier-scaled monthly minimums of roughly $16 / $256 / $4,096. Volume discounts sharply reduce marginal query cost at higher data volumes. For large deployments, turbopuffer is often cited as running several times cheaper than in-memory serverless alternatives.