Qdrant Cloud or Self-Hosted: What It Really Costs

November 6, 2026

Qdrant Cloud’s free tier is a 1 GB cluster with no time limit on one node, and paid clusters start around $25 per month. Self-hosting Qdrant is free under Apache 2.0 with no vector cap, but you take on backups, monitoring, and upgrade risk. The break-even point is usually earlier than people expect.

What does 1 GB actually hold?

Roughly 200,000 vectors at 1024 dimensions with float32 and minimal payload, or considerably fewer once you add metadata and the HNSW index. In practice, a RAG collection with real chunk text in the payload fills the free tier somewhere between 30,000 and 80,000 chunks.

That is enough for a single-product knowledge base. It is not enough for a multi-tenant SaaS, where each customer adds thousands of chunks.

When is Cloud the obvious choice?

Under about a million vectors, with no one on the team who wants to be an operator. The paid tiers include snapshots, backups, and monitoring dashboards, and an outage becomes a support ticket rather than your Saturday.

There is a second case: regulated environments where you need a vendor SOC 2 report more than you need control. Self-hosting your own vectors does not remove the compliance obligations around them.

A useful rule: if your time costs more than $25/hour, and self-hosting costs half a day a month of maintenance, Cloud is already cheaper before you count the server.

What does self-hosting actually require?

A machine with RAM, a persistent volume, and someone who checks it.

  • Backups — Qdrant snapshots exist, but nothing schedules them for you. A cron that writes a snapshot off-box is the minimum.
  • Monitoring — a scrape of the metrics endpoint into whatever you already run, with an alert on memory and on collection size growth.
  • Upgrades — version bumps occasionally change collection formats. Read release notes and pin the image tag.
  • Security — port 6333 open to the internet is a data leak, not an installation. Bind to localhost and put it behind your app.

That is the whole job for a small deployment. It is not hard, but it is recurring.

How does memory scale, and can quantisation help?

Plan 4 bytes per dimension per vector before index overhead. A million 1536-dimension vectors is about 6 GB, and the HNSW graph adds more.

Scalar quantisation to int8 cuts memory roughly 4x with a rescoring pass to recover accuracy. Binary quantisation claims up to 32x, with more accuracy loss. Enabling quantisation is often the difference between a $25/month node and a $200/month node — try it before buying capacity.

Under roughly 10 million vectors, the default HNSW parameters are fine. Tuning early tends to cost accuracy for a speed gain you cannot measure.

What are the real risks of each path?

Cloud: vendor lock at the infrastructure layer, price changes you do not control, and data residency questions if your customers are in a specific jurisdiction. Mitigation is keeping the collection definition in code so you can export and rebuild elsewhere.

Self-hosted: silent failure. A disk fills, the service restarts, and your app returns empty results instead of errors. Plan a health check that queries a known vector rather than pinging a port.

Which should you pick if you are unsure?

Start on the free Cloud tier. Migrate later once you know the actual vector count, query load, and growth rate. Migrating a collection is straightforward — snapshot and restore into a new cluster — so the decision is reversible.

Do not build self-hosted infrastructure for a product you have not validated. That is the expensive mistake, not the $25.

Quick Summary

  • Free Cloud tier: 1 GB cluster, no time limit, roughly 30,000-80,000 RAG chunks with payloads.
  • Paid Cloud clusters start near $25/month and include snapshots and monitoring.
  • Self-hosting is free with no vector cap but needs scheduled snapshots, monitoring and pinned versions.
  • Int8 quantisation cuts memory about 4x — try it before scaling up.
  • Start on Cloud, measure real volume, then decide; migration via snapshot is cheap.

Distances, quantisation modes, and cloud pricing tiers are listed on the Qdrant tool page.

Tags: