turbopuffer

turbopuffer is a vector and full-text search database built object-storage-first, aimed at engineering teams running retrieval for AI products at very large scale. It keeps S3 or Google Cloud Storage as the single source of truth and uses memory and NVMe SSD only as a cache tier, which removes the cost of holding an entire index in provisioned RAM and makes enormous multi-tenant corpora economically viable.

turbopuffer is a search and retrieval database that treats object storage — S3 or Google Cloud Storage — as the single source of truth, with memory and NVMe SSD acting only as a cache tier. That inversion of the usual in-memory vector database design is the entire product thesis: because cold data lives in object storage rather than provisioned RAM, turbopuffer meters what customers actually consume instead of charging per node or per gigabyte of provisioned memory, and the company advertises roughly an order of magnitude lower cost than RAM-resident alternatives. A single engine serves vector search at over 90% recall, BM25 full-text ranking, regex via trigram indexes, and metadata filtering, so hybrid retrieval does not require running two systems side by side. The documented production envelope is 1 trillion-plus documents, 10 million-plus writes per second, 25,000-plus queries per second and up to 256TB per namespace, with copy-on-write branching and customer-managed encryption keys available per namespace. Founded in 2023 by former Shopify infrastructure engineers Simon Hørup Eskildsen and Justine Li and based in Ottawa, Canada, turbopuffer took a small angel round from Lachy Groom in early 2024 and an undisclosed seed round with Thrive Capital in December 2025. Its namespace-per-tenant model made it the default retrieval layer for AI-native products with very large multi-tenant corpora: Cursor indexes more than a trillion chunks of code across over 80 million namespaces, and Anthropic, Notion, Linear, Atlassian, Ramp, Grammarly, Harvey, TELUS and Superhuman are named production customers.

Category

Data & Analytics

Pricing

Usage-based, Subscription

Target Market

CTOs, Enterprise Developers, Platform Engineers, Data Scientists, ML Engineers

Deployment

Cloud-first, API-based, Hybrid

Founded

2023

Headquarters

Ottawa, Canada

Team Size

11-50

• ✓
  Object-storage-native architecture

  S3 or GCS is the sole source of truth with memory and NVMe SSD as a cache layer, so storage cost tracks object storage rates instead of provisioned RAM

• ✓
  Hybrid vector, BM25 and regex search

  Vector search at over 90% recall, full-text BM25 ranking and trigram-indexed regex run in one engine, removing the need to operate a second search system

• ✓
  Namespace-per-tenant multi-tenancy

  Up to 256TB per namespace and tens of millions of namespaces per account, which is how Cursor isolates more than 80 million customer codebases

• ✓
  Copy-on-write branching

  Namespaces can be branched cheaply for testing index changes or re-embedding runs without duplicating the underlying stored data

• ✓
  Consumption-metered pricing with no platform fee

  Charges are storage per GB-month, writes per GB and queries per petabyte scanned, so idle data costs almost nothing rather than holding a paid node

• ✓
  Customer-managed encryption and private networking

  CMEK per namespace, IP allowlisting, audit log streams and private networking are available for regulated enterprise deployments

• ✓
  Five official SDKs generated from one API spec

  Python, TypeScript, Java, Go and Ruby clients are Stainless-generated from a single spec, so language clients stay consistent as the API evolves

• •
  AI coding assistant code retrieval

  Cursor indexes over a trillion chunks of code across more than 80 million namespaces so each developer's repository stays isolated and searchable

• •
  In-product AI search over customer workspaces

  Notion migrated from Pinecone to power Notion AI, and Linear runs issue search on it, both requiring per-workspace isolation at scale

• •
  Enterprise RAG over large document corpora

  First-stage retrieval narrows millions of documents down to dozens or hundreds before a reranker or LLM sees them, which is the pattern the engine is tuned for

• •
  Email and communication search

  Superhuman uses it for email search, where each mailbox is a separate namespace and most historical mail is queried rarely

• •
  Cost-driven migration off provisioned vector databases

  Teams whose Pinecone or Weaviate bill scales with index size in RAM move to object-storage economics without giving up hybrid search

There is no free tier. Pricing is pure consumption metered against a monthly tier minimum — customers pay the greater of usage or the floor, which is $16 on Launch (cut from $64 in June 2026), $256 on Scale and $4,096 on Enterprise, where a 35% usage premium also applies. Published rates are up to $0.33 per GB-month of storage, up to $2.00 per GB written (copy_from_namespace operations get a 50% write discount), and $1 per petabyte scanned on queries with volume discounts of 80% above 32GB scanned and 96% above 128GB, subject to a 1.28GB minimum billable scan per query. The query rate was cut from $5/PB to $1/PB in February 2026. HIPAA BAA and SSO require Scale; BYOC, single-tenancy, private networking and CMEK are gated behind Enterprise at $4,096/month or more.

This page was written from 7 sources, 5 on domains other than turbopuffer.com.

1. 1.turbopuffer.com — pricingvendor
2. 2.turbopuffer.com — docsvendor
3. 3.sacra.com — turbopuffer
4. 4.usagepricing.com — turbopuffer
5. 5.modern-datatools.com — turbopuffer
6. 6.jxnl.co — turbopuffer object storage first vector database architectur
7. 7.stainless.com — turbopuffer