How the prices are collected

1. Coverage. This site’s pages draw on 330 price rows from 34 providers across 13 GPU models.
2. What is left out. Left out: spot, reserved and negotiated prices, storage, network egress and tax. These are published list prices, not what a given customer pays. 56 marketplace figures set by hosts are shown with a label and kept out of rankings and statistics.
3. Checks. A price must sit between $0.05 and $150 per GPU-hour, its digits must appear in the stored text, and a configuration may appear only once. A row that fails any test is dropped, not shown.
4. Unit. One unit throughout: USD per GPU per hour, tax not included. Per-server prices are split by GPU count so rows compare like with like.
5. Freshness. Every row shows the day it was read, and the latest read of the stored data is 12 November 2026. 23 prices that take effect later are shown as announced and kept out of the statistics until then. Runcrate’s own rows come live from runcrate.ai.
6. Order. Tables sort by price, lowest first. Nothing is pinned, sponsored or boosted. Runcrate is one provider among the rest and carries an “ours” label.
7. Source. Every price is read from the provider’s own public pricing page. For each row we keep the page address, the time we read it and the exact line of text that holds the number.

Found a wrong number? Tell us.