How a price gets onto this site — and what does not.

One unit

Providers quote what suits them: per instance, per minute, per month, in their own currency. Those figures cannot be compared with each other, so each is converted to US dollars per GPU per hour and the original is kept beside it on the chip's page.

The conversion is arithmetic on numbers the provider published — the instance price and the GPU count — never an estimate. Where a provider does not state how many GPUs are in an instance, the offer is dropped rather than divided by a guess.

Currency

One source quotes in euros. It converts at the European Central Bank's daily reference rate for 2026-08-21 — 1 EUR = 1.1699 USD — and both the euro figure and the rate used are printed next to every converted price. ECB reference rates ↗

Reference rate published each working day at 16:00 CET. Not a dealing rate.

Same chip, or not

An exact variant is vendor, family, form factor and memory. Two providers are only compared when all four match. An H100 SXM and an H100 PCIe are both sold as "H100" and are different silicon at different prices; a 40 GB A100 and an 80 GB A100 likewise. Merging them would produce a tidier table and a false one.

Marketing prefixes are dropped for grouping only — "Tesla V100" and "V100 SXM2" are one card written two ways. Every offer keeps the provider's exact wording, shown on the chip's page, so this normalisation can be checked rather than trusted.

On demand and spot

Headline rates are on-demand: capacity you can rely on. Spot, bid and low-priority prices are a different product — capacity that can be taken back — so they never set a headline figure and are shown separately, in the provider's own term for them.

Freshness

Each source carries its own rule: 12 hours for a live marketplace, 48 for a published catalogue. Past that, the reading is shown as stale everywhere it appears. It is never silently restamped as current.

Specifications

Vendor-published figures. FP16 throughput is dense tensor throughput with no sparsity assumed; vendors often headline the sparse figure, which is twice as large. A chip with no cited vendor figure shows no throughput number and is priced all the same.