Skip to content

How we compare GPU prices

Cloud providers didn't design their catalogs to line up. Here's how we turn provider listings and observations into numbers you can compare, plus where those comparisons stop being useful.

Where the data comes from

Once an hour, a scanner signs in to each provider's API with an ngrok account and records every GPU instance the provider lists: its specs, its list price, and its availability. A scan that fails records nothing, so a gap in a chart means we got no reading, not that anything changed.

This site loads the last 30 days of those scans. Each instance's specs come from its newest scan. The catalog lists GPU instances only, so host vCPU and RAM describe the machine a GPU comes in, not a separate CPU offering.

Here's what the scanner reads from each provider:

AWS
EC2 instance types and spot price history, plus on-demand Linux prices from the AWS Price List API, in 17 regions. AWS publishes no availability.
Azure
Linux on-demand and spot prices from the Azure Retail Prices API, with GPU VM specs from a size table the scanner keeps, in 42 regions. Azure publishes no availability.
DigitalOcean
Droplet sizes and regions from the DigitalOcean API, with an available flag on each size and region.
Google Cloud
On-demand and Spot VM prices from the Cloud Billing Catalog API, with GPU machine type specs from a table the scanner keeps, in 32 regions. A machine type's price is the sum of its GPU, vCPU, RAM, and bundled Local SSD rates. Google Cloud publishes no availability.
Hyperstack
Flavors, regions, and the pricebook from the Hyperstack API. A flavor's price is the sum of its GPU, vCPU, and RAM rates, and each flavor carries a stock flag.
Lambda
Instance types from the Lambda API, each with the regions that have capacity for it right now.
Linode
Plan types, regions, and per-region plan availability from the Linode API.
Oracle Cloud
Per-GPU-hour prices from the Oracle Cloud price list API, multiplied by each shape's GPU count, with GPU shape specs from a table the scanner keeps. Oracle charges one price in every region, so each shape has a single global row. Only the V100 virtual machine shapes sell preemptible capacity, at half the on-demand price. Oracle publishes no availability.
RunPod
Secure Cloud GPU prices from the RunPod GraphQL API, and each data center's stock level from its GPU catalog. Every RunPod row is one GPU.
Scaleway
Server products and their availability from the Scaleway API, in nine zones, priced in euros.
Verda
Instance types, locations, and the on-demand and spot availability lists for each location from the Verda API.
Vultr
Plans, regions, and each region's list of plans it can deploy now from the Vultr API.

When a provider's scans stop, its last scan stays in the catalog. Its provider page says how old that scan is, and it drops out of the count of providers with current prices until scans resume. The scanner also reads providers that don't have a page here yet. Their data stays off the site until they do.

How fresh the data is

Every page carries a freshness line. It measures the age of the newest scan across all providers, except on a provider page, where it measures that provider's own last scan. As the data ages, the line changes, and so does what the page lets itself say:

Under 2 hours
“Updated hourly. Last scan 12 minutes ago.” Prices are current, and in-stock language is fine.
2 to 48 hours
“Last updated 6 hours ago.” The numbers stay up, but nothing claims to be true right now.
Over 48 hours
“Historical snapshot from 2026-08-19.” The data is history. Pages stop saying available now, latest, or in stock.

The Markdown pages and /llms.txt carry the same state, so an agent that quotes a snapshot quotes its date too.

How we normalize prices

We compare hourly list prices in US dollars. Each observation also keeps the provider's own currency and price. A price quoted in another currency converts at the European Central Bank's daily reference rate, and each observation stores the rate and its date, so you can reproduce any conversion.

An instance's price history counts a change only when the price in the provider's own currency moves. A euro price converted to dollars drifts with the exchange rate every day, and diffing the dollar figure would report repricings that never happened.

Two exceptions to “list price”: a provider that quotes zero, or nothing, gets no price rather than a free instance, and Hyperstack's dollar price is the rate our account pays, which matches its list price unless Hyperstack applies a discount.

Price per GPU-hour
Hourly instance price divided by the number of GPUs in the instance. Every price ranking on the site sorts by this, so an 8-GPU node ranks by what each of its GPUs costs. A slice of one GPU counts as that fraction of a GPU, so a 1/8 L4 at $0.20 an hour costs $1.60 per GPU-hour.
Price per GB of VRAM-hour
Price per GPU-hour divided by the memory on one GPU.
Monthly price
The provider's own monthly quote, converted to dollars, shown only when the provider publishes one. We don't estimate a monthly price from the hourly one.

Per-unit prices make differently sized instances easier to compare. They don't normalize GPU performance, network throughput, storage, support, or any other difference in what you get for the money.

How the headline numbers work

Every number on the site starts from each instance's latest observation. Price rankings use on-demand prices only, so a spot price can't win a contest with different terms.

Providers with current prices
Providers with at least one GPU instance whose latest observation has a price, isn't marked unavailable, and falls within 24 hours of the newest scan. The provider count on the homepage and on /providers uses this rule. A GPU page counts only the providers that sell that GPU, and the GPU Index counts the providers in its basket. The instance and region counts beside it use the same rule and count every purchase model. Each chart's own rule is defined below.
Cheapest listed
On a GPU page, the lowest on-demand price per GPU-hour among instances listed now, at any availability. It's the number a price list shows you.
Cheapest in stock
The lowest on-demand price per GPU-hour among instances listed now whose latest observation reads available. The prices on the homepage use the same rule, and so does the last-scan column of the provider table on a GPU page.
Price to budget for
On a GPU page, the lowest on-demand price per GPU-hour among instances listed now that read available in at least 90% of their availability readings over the last 30 days. When no instance clears that bar, it's the cheapest in stock instead. The provider table beside it reads the same 30 days, so a provider that drops sold-out instances from its catalog still has a row, marked sold out.
Cheapest in stock, by hour
The homepage chart, for each of the four GPU models that readers search for most, and the chart at the top of each GPU page. For each hour, the lowest on-demand price per GPU-hour among instances of the model that their provider reported available in its newest scan before the hour ended. A provider whose newest scan is more than 24 hours old doesn't count. An hour with none in stock is a gap in the line. The change beside the homepage price compares it with the median of the in-stock hours in the chosen range, and the dashed line marks that median. The strip under the chart marks the hours when the cheapest in-stock price equaled the range low, to the cent.
Available only
The catalog filter that keeps instances listed now whose latest observation reads available. The homepage's “Browse GPUs in stock” button opens the catalog with this filter.

An instance is listed now when its provider's newest scan includes it and that scan is less than 24 hours older than the newest scan in the catalog. A provider that drops sold-out instances, or whose scanner stopped, has instances that aren't listed now. The prices on GPU pages and the Available only filter use only instances listed now. The catalog, on the homepage and on each GPU page, keeps every instance with its last price and availability, and every catalog row shows when it was last seen.

A GPU gets its own page once at least two providers sell it, with five or more priced instances between them. One provider isn't a comparison.

What availability means

Availability is what the provider reported at observation time. It isn't a reservation or a capacity guarantee.

Available
The provider's API reported the instance as available.
Unavailable
The provider reported the instance as unavailable, or dropped it from its list when it sold out.
Not reported
The provider gave no availability signal for the instance. That doesn't mean it's sold out.

Providers report availability in one of three ways, so the same percentage means different things from different providers. Each provider page says which kind it is, judged from its last 30 days of observations:

Reports both states
The provider marks instances available and unavailable. Its percentages mean what they say.
Reports nothing
Every instance reads not reported. The data can't tell you whether you can buy at that price.
Only lists what has capacity
The provider returns an instance while it has capacity and drops it from the list when it sells out. The scanner records each drop, and every scan that misses the instance counts as a sold-out reading. When the scanner can't tell a drop from a sellout, it marks the drop unknown. Its stock always reads 100% available, which is a missing signal. GPU pages leave a provider like that out of every in-stock share.

Availability percentages leave not-reported observations out of both sides of the calculation, and they cover the 30 days up to the newest scan. Trend lines stay separate by provider, because adding a provider with a different catalog mix would otherwise look like a market-wide change.

Purchase models and derived fields

On-demand and spot prices stay separate in every price ranking. An instance counts as spot when the provider labels it spot, or when its name says spot: DigitalOcean lists some spot droplets as on-demand.

Most specifications come straight from the provider. The site derives a few fields itself:

  • Interconnect (SXM, PCIe, or PCIe NVLink) from the GPU model and instance name, because no provider reports it as structured data, and “Unspecified” when the name doesn't say.
  • GPU model groups from name patterns, because providers spell the same GPU differently.
  • VRAM per GPU from the GPU's published spec when the provider leaves it out or reports more than the spec allows.
  • Region cities from a hand-kept table or from names the provider publishes, and only the country when neither exists.

The data has no CPU model or silicon-generation field, so the site can't compare instances on either.

What these comparisons don't tell you

A listed price is a starting point, not total cost of ownership. The comparisons don't include:

  • Negotiated discounts, credits, taxes, or support plans.
  • Network egress, attached storage, or other usage-based charges.
  • Benchmarked performance, noisy-neighbor effects, or CPU generation.
  • A guarantee that capacity will still be available when you open the provider's console.

Use the normalized numbers to narrow the field, then verify the provider's current price, capacity, and full billing terms before you buy.