bmax

What Does Turbopuffer Cost in Production?

· 1 min read
By

Turbopuffer bills four meters, all usage-based: writes at $2.00 per logical GB, query scans at $0.001 per TB, query results at $0.05 per GB returned, and storage at roughly $0.33 per logical GB-month (discounted rates). There are no instance sizes or cluster tiers. Running 133M+ vectors in production, our invoices land between $1,034 and $1,394 a month — the same workload cost ~$10,000/month on MongoDB Atlas.

Our actual invoices (133M+ vectors, moderate QPS)

Line itemApril 2026May 2026June 2026
Writes ($2.00/GB logical)$869.36$218.84$456.00
Queries — scanned ($0.001/TB)$103.38$1,309.29$715.59
Queries — returned ($0.05/GB)$8.96$80.43$26.35
Storage (~$0.33/GB-mo logical)$52.46$148.99$196.43
Total (gross)$1,034.16$1,757.55$1,394.37

Three things the bills teach:

  1. Storage is nearly free. ~600 GB of logical data costs under $200/month. We have never discussed deleting data to save money.
  2. Scans are the variable line, and they're cheap at absurd scale. In May we scanned over 1.3 exabytes (1,309,290 TB) running corpus-wide batch work — for $1,309. On a provisioned database that month would have been a capacity-planning project.
  3. Writes are a one-time tax per byte. Batch your ingest; the $2/GB meter is on logical GB written.

Turbopuffer also sells pre-purchase credits at a discount (our May net was $1,245.55 after credits) and applies 50% batch discounts on writes over 3MB.

When the math works

The pricing model is object-storage-first: you pay for what moves, not for capacity that sits provisioned at peak. It fits large, growing corpora with bursty or moderate query load. Under a few million vectors, pgvector on your existing Postgres is fine and this math doesn't matter yet.

Full migration story: 133 Million Chunks.

More writing