How to compare GPU cloud prices per hour

For H100 80 GB, the 12 public offers in our latest data range from $1.99/hour (Runpod H100 PCIe 80 GB (Community Cloud)) to $4.29/hour (Lambda NVIDIA H100 SXM 80 GB (on-demand) 1 GPU instance) per GPU hour, across 4 providers in total and 4 rental modes. Latest check: Oct 18, 2026. Methodology · All GPU price records.

Published by Caliper Press · Page updated Oct 17, 2026

Why there is no single GPU price

A "GPU hour" is not one product. The price depends on the GPU model, the memory size of that model, the interconnect variant (PCIe, SXM, NVLink), and above all the rental mode. The same card can be offered on demand, as interruptible spot or preemptible capacity, or on a community marketplace, at very different prices. Our tracked offers use these modes: on-demand: 23; Secure Cloud: 12; Community Cloud: 4; spot: 3.

For this reason the Caliper Press GPU Public Price Watch does not pool everything into one number. GPU families, memory sizes and rental modes have separate series; a pooled scalar is not defined. Separate medians by GPU family, memory size and rental mode; no pooled scalar.

Compare within one family and one mode

A useful comparison holds three things fixed: the GPU model, its memory, and the rental mode. The tables below show the GPU families in our data that are offered by at least two providers, with every offer's mode shown in its own column. Rows are sorted by the measured hourly price. Comparing an on-demand row with a spot row in the same table is possible, but they are different services: spot capacity can be reclaimed by the provider.

H100 80 GB: 12 offers from 4 providers

A100 80 GB: 8 offers from 2 providers

A6000 48 GB: 2 offers from 2 providers

A100 cloud pricing

In our latest data, 10 public NVIDIA A100 offers from 3 providers range from $1.08/hour (Hyperstack NVIDIA A100 80 GB (spot)) to $1.99/hour (Lambda NVIDIA A100 SXM 40 GB (on-demand) 1 GPU instance) per GPU hour (latest check: Oct 18, 2026). Memory size (40 or 80 GB), interconnect (PCIe, SXM or NVLink) and rental mode explain most of the spread, so compare offers that match on all three.

What the hourly price covers

Each value is the public USD price per GPU hour from the provider's page or public offer search: Hyperstack, Lambda, Runpod, Together. It is a published price, not a completed rental, a benchmark or an invoice. Providers differ in what the hourly rate bundles: CPU cores, system RAM, local storage and network egress are included by some and billed separately by others, and they are not normalised here. For marketplace offers the rate is the lowest matching offer at the time of the read and includes the offer's storage component; the offer can disappear minutes later, and a change of the cheapest offer can explain a change of price.

Multi-GPU prices are not divided into single-GPU prices. Where a provider publishes a per-GPU rate for an 8-GPU configuration, the record says so. Reservation and commitment discounts are excluded; they require a contract and are not public list prices.

What a price does not say

A public rate does not guarantee available capacity. The cheapest on-demand offer may be sold out in the region you need, while a more expensive one is available. We do not test availability, provisioning time or performance. The deployment location is recorded only when the provider states it.

Reading the history

Each offer has a history of measured prices with dates. Changes are compared only under identical hardware and rental conditions; an offer whose conditions change is a different object. 42 offers are tracked, 42 of them with a successful latest measurement, as shown in the United States. The history spans 3 observation date(s), from Oct 16, 2026 to Oct 18, 2026.

Every table on this page is built from the same records that are available as JSON and CSV; see how to use this data.

All GPU cloud hourly prices records · Methodology · How to use this data