Computeby GlyphSources and method
Every price is read from the provider’s own price list, API or pricing page and normalised to US dollars per GPU-hour. History since Sep 24, 2026.
Where the prices come from
| Source | What | Refresh | Prices |
|---|---|---|---|
| AWS EC2 price list and spot feedKeyless JSON behind the EC2 pricing pages | On-demand per instance in every region; current spot prices | On-demand weekly, spot daily | 777 |
| Azure Retail Prices APIKeyless REST API | Pay-as-you-go, spot, 1- and 3-year reservations per VM size and region | Weekly | 1,331 |
| Google Cloud pricing pagesHTML tables | A2, A3, A4 and G2/G4 machines (Iowa): on-demand, spot, 1- and 3-year CUDs; Cloud TPU per chip-hour in every region | Weekly | 106 |
| Oracle Cloud price listKeyless price list API | GPU list price per GPU-hour | Weekly | 12 |
| CoreWeave pricingPricing page | On-demand and spot per instance | Weekly | 15 |
| Lambda pricingPricing page | On-demand per GPU; 1-Click Cluster contract prices | Weekly | 19 |
| Nebius pricesPricing page | On-demand and preemptible per GPU-hour | Weekly | 14 |
| Crusoe pricingPricing page | On-demand per GPU-hour | Weekly | 11 |
| Runpod pricingPricing page | Secure Cloud and Community Cloud per GPU-hour | Weekly | 30 |
| Together AI pricingPricing page | Preemptible, on-demand and reserved GPU clusters per GPU-hour | Weekly | 11 |
| DigitalOcean GPU DropletsPricing page | On-demand, spot and 12-month reserved per GPU-hour | Weekly | 11 |
| Akamai (Linode) types APIKeyless public API | GPU plans (RTX 4000 Ada and more), hourly list price | Weekly | 9 |
| Vultr plans APIKeyless public API | Cloud GPU and bare-metal GPU plans (A16, A100, H100, L40S, B200, MI300X, MI325X, MI355X, GH200), hourly list price | Weekly | 9 |
| Scaleway products APIKeyless public API | GPU instances in Paris, Amsterdam and Warsaw, in euros converted at the ECB reference rate | Weekly | 19 |
| Verda (DataCrunch) instance types APIKeyless public API | On-demand and spot per instance (GB300, B300, B200, H200, H100, A100 and more) | Weekly | 114 |
| Vast.ai marketplaceKeyless search API | Cheapest verified on-demand offer and lowest interruptible bid per chip | Daily | 12 |
How the list grows
The aim is every cloud that publishes accelerator prices. 16 price readers so far (+1 in the last 30 days). Each day the run also syncs the companies.fru.dev registry and checks a few companies for a public GPU price page: 887 of 2,448 checked so far, 7 price pages found and waiting for a reader. 5 accelerators that sources list are not mapped to a chip yet.
Method
- Per-GPU price = instance price divided by the accelerators in the instance. TPU and Trainium are per chip-hour. Neocloud per-GPU prices are shown as published.
- On-demand is the list price with no commitment (Linux, shared tenancy on AWS and Azure). Spot is the current price from AWS’s and Azure’s feeds, Google Cloud’s page, or a neocloud’s “from” price. 1-year and 3-year are reservations or committed-use discounts as an hourly rate; Contract is a published cluster price for a fixed term.
- Azure spot rows priced at or above pay-as-you-go are dropped: Azure lists them where spot is not really offered. Vast.ai is the cheapest verified offer (reliability 97%+) at read time.
- History: a price that moves is updated and logged with the day it was first seen and the last day the old price was seen. Delisted offers are kept, marked removed. A source that suddenly returns far fewer rows removes nothing.
- Not tracked yet: AWS Savings Plans and Capacity Blocks, Google Cloud prices outside Iowa for GPUs, and prices behind a sales contact.