How We Compare GPU Cloud Prices
Where our GPU prices come from, how we date and recheck them every 14 days, how the calculators work, and why no provider can buy a better position.
Where prices come from
Every price is read on the provider's own pricing page, or an official price API (Azure). We record the URL and the date. We don't copy prices from aggregators or provider blogs. If a page doesn't show a price as text, or shows two conflicting prices, we leave it out until we can confirm it.
Per-GPU prices
Some clouds rent only 8-GPU machines. We divide the machine price by eight and mark the row with a star, and the job cost calculator bills you for the whole machine.
How often we check
Every 14 days, and sooner when a provider announces a change. Rows older than 14 days show their age in red. Changes are listed in the updates page.
How we rank
By estimated total cost for the job you describe: compute hours after billing rounding and minimums, storage, idle storage and egress. When a provider doesn't publish a fee, we say "not stated" rather than guess.
No provider pays us, and nothing about money enters the formula. If we ever add a sponsor, it will be labelled and kept outside the rankings.
The VRAM estimates
Weights from parameter count and precision, KV cache from layers, context, batch and KV width, about 16 bytes per parameter for full fine-tuning with Adam, and 10% overhead. Real use depends on the framework, so we show a ±15% range.
What we don't do
- No provider logos and no implied partnerships.
- No paid placement in tables.
- No "live" claims: our prices are dated snapshots.
Spotted an error or a figure that has changed? Let us know