GPU cloud prices per GPU-hour: H100, H200, B200, A100, L40S, RTX 4090 and more, with AWS, Google Cloud and Azure

Lowest listed price per GPU: GPU clouds and hyperscalers

The cheapest on-demand listing for each GPU among the GPU clouds, marketplaces and serverless platforms above, next to the cheapest on-demand listing from AWS, Google Cloud or Azure. Instance prices are divided by the number of GPUs in the instance. Raw data: prices.json.

The comparison is not like for like. A hyperscaler GPU instance includes vCPUs, system memory, local NVMe storage and fast networking (EFA or InfiniBand on the large ones), and most hyperscaler GPU capacity is bought with reservations, savings plans or committed-use discounts, which are not shown here. Hyperscaler prices are for one region each: AWS EC2 us-east-1, Google Cloud us-central1, Microsoft Azure eastus.

Provider notes

RunPod

Billing: Per second. Community Cloud runs on vetted third-party hosts; Secure Cloud runs in RunPod's partner data centers. Both tiers are recorded. The page said "Updated September 27, 2026".

Lambda

Billing: Per minute. Per-GPU price for 1x, 2x, 4x and 8x instances; larger instances cost less per GPU. 1-Click Cluster prices (16+ GPUs, 2 weeks or longer) are not on-demand and are not recorded.

DigitalOcean GPU Droplets

Billing: Per second, 5-minute minimum; powered-off Droplets are still billed. 12-month reserved and spot prices on the same page are not recorded. The page said "New pricing is effective as of August 1, 2026".

Paperspace (DigitalOcean)

Billing: Per hour. Prices marked with asterisks on the page are 3-year-commitment prices and are not recorded as on-demand. A footnote says "On-demand pricing for H100 is $5.95/hour under our special promo price"; promotional prices are not recorded.

Vast.ai (no prices recorded)

Billing: Per second; prices set by individual hosts. A marketplace: each host sets its own price, so there is no single list price. We record a figure only when Vast.ai's public pages state one in their HTML. Live marketplace prices load in the browser; the page HTML contained no prices, so none are recorded.

CoreWeave

Billing: Per hour. North America region. Listed per instance (8 GPUs; 4 for NVL72 systems); divided by the GPU count here.

Hyperstack

Billing: Per minute. The table lists 'H100', 'H100 NVLink' and 'H100 SXM' (and the A100 equivalents) as separate products without stating the form factor of the first two.

Nebius

Billing: Per hour. Two price columns: 'On-demand, GPU-hour' and 'GPU-hour (Effective October 1, 2026)'. Recorded: the second (now in effect).

Modal

Billing: Per second while a container runs; scales to zero. CPU and memory are billed separately. Modal's plan table lists non-preemptible execution at 3x base prices and region selection at 1.15-1.75x; the prices here are base prices.

Crusoe

Billing: Per hour.

JarvisLabs

Billing: Per minute; USD prices for customers outside India.

Voltage Park

Billing: Per hour. The pricing table says 'Contact for pricing'; the FAQ on the same page says one H100 can be rented 'starting at' this price without a contract.

AWS EC2 (us-east-1)

Billing: Per second for Linux, 60-second minimum ("by the hour or second (minimum of 60 seconds)"). Hyperscaler GPU instances bundle vCPUs, system memory, local NVMe storage and high-speed networking with the GPUs, and the price shown per GPU is the whole-instance price divided by the GPU count. Most hyperscaler GPU capacity is bought with commitments (reservations, savings plans, committed-use discounts or short-term capacity reservations), which can be far below on-demand; those prices are not recorded. US East (N. Virginia), us-east-1. On-demand Linux, shared tenancy. GPU names and memory per GPU from AWS's instance-type pages (aws.amazon.com/ec2/instance-types/p5/, p4/, p6/, g6e/, g6/, g5/); AWS does not state the H100/H200 form factor there, so none is recorded. Savings Plans, Reserved Instances, Spot and EC2 Capacity Blocks for ML are not recorded. Read from AWS's public Price List file for EC2 in us-east-1, version 20261006172454 (published 2026-10-06T17:24:54Z).

Google Cloud (us-central1)

Billing: Per second after a 1-minute minimum. Hyperscaler GPU instances bundle vCPUs, system memory, local NVMe storage and high-speed networking with the GPUs, and the price shown per GPU is the whole-instance price divided by the GPU count. Most hyperscaler GPU capacity is bought with commitments (reservations, savings plans, committed-use discounts or short-term capacity reservations), which can be far below on-demand; those prices are not recorded. Iowa, us-central1. Machine-type prices include the GPUs, vCPUs, memory and any bundled Local SSD. GPU names from Google's machine-type documentation (https://cloud.google.com/compute/docs/accelerator-optimized-machines), which states SXM for the H100 (A3) and H200 (A3 Ultra). Google's Cloud Billing Catalog API needs an API key from a Google Cloud project, so the public pricing page is read instead. DWS Flex-start, Calendar mode, Spot and committed-use prices on the same page are not recorded. The page's tables are read in its default region, Iowa (us-central1); the region name is checked on every read.

Microsoft Azure (eastus)

Billing: Per full minute of running time. Hyperscaler GPU instances bundle vCPUs, system memory, local NVMe storage and high-speed networking with the GPUs, and the price shown per GPU is the whole-instance price divided by the GPU count. Most hyperscaler GPU capacity is bought with commitments (reservations, savings plans, committed-use discounts or short-term capacity reservations), which can be far below on-demand; those prices are not recorded. East US, eastus; a size the API lists no eastus price for is read in eastus2, and its row says so. Linux pay-as-you-go prices from the Azure Retail Prices API (prices.azure.com, no account needed). GPU names from Azure's GPU size-series documentation on learn.microsoft.com, which states H100 NVL for NCads H100 v5 and PCIe for NC A100 v4. Spot, Low Priority, reservation and savings-plan prices are not recorded.