GPU Price Index
Language / English

Methodology & sources

Cheapest among observed offers only. Coverage depends on the last collection; this is not a global-market minimum.

Latest collection: complete; pages: 1, normalized records: 114; finished .

Verified public sources: tier A official APIs, including Microsoft Azure Retail Prices API, may be displayed and exported. B/C pages are display only, fetched at most once per hour per provider. JSON includes only tier A rows with export_allowed. Link-only sources are never fetched.

Minimum, median and maximum use current, exact USD prices for whole GPUs across all providers, separately for on-demand and Spot. Standard and unspecified regions are eligible; edge and restricted regions are excluded. Unknown tax basis is eligible and labelled; tax-included, fractional, GPU-component-only and advertised starting prices are excluded. GPU-hour rates divide instance rates by GPU count; per-GPU rates retain their scope. Calculations use unrounded values. Minima may refer to different regions or SKUs. Region counts include an unspecified-location group when present.

Maximum = the highest price across the same eligible offers. Models tracked = the number of distinct GPU models represented in today's eligible offers.

Offers in edge and restricted regions stay visible with badges but do not enter headline comparisons or history. An unspecified region is eligible and labelled “Region not specified by provider”. Unrecognized region IDs default to standard and are logged for review. Restricted badges include sovereign clouds and the conservative exclusion of West India and legacy Jio regions.

Dollar values use two decimals at or above USD 1 and three decimals below USD 1, using unrounded source values for all calculations. Counts n accompany every offer-price median. No discount percentage is displayed when the Spot minimum is at least the on-demand minimum; discounts are never negative.

The workflow may run every 15 minutes; each page provider is fetched at most once per rolling hour. SQLite meta persists the time across runs. Skips retain the observation and collection status. UTC dates and outcomes indicate freshness; missing history days are not filled.

Every build independently recalculates published prices and statistics from SQLite, checks JSON exports and localized HTML including titles, descriptions, FAQ structured data and charts, and fails on a mismatch. Optional live sampling compares stored VM prices with Azure; a newer effective date is reported as a source update rather than a historical error.

We request Virtual Machines in NC, ND and NV families across all commercial regions with USD currency, Consumption prices and primary meters. Windows, paid Linux license add-ons, Dev/Test, reservations, savings plans, legacy Low Priority, non-hourly, future-effective, zero-price and invalid meters are excluded. Base VM products without a Windows designation are treated as Linux base-OS list rates.

Known exact ARM SKUs map to documented GPU model, GPU count, vCPU and RAM. Unknown SKUs and conflicting same-date rates are logged and skipped. The most recent effective date wins within a collection. L40S and other unverified sizes are not guessed. Curated coverage is not exhaustive.

FX uses dated ECB reference rates: EUR × USD rate, or CNY × USD rate / CNY rate. Failed refreshes retain the last valid date. Original currency and hourly price remain visible. Monthly prices are never divided into hourly prices.

Minimum, median and maximum use current, exact USD prices for whole GPUs across all providers, separately for on-demand and Spot. Standard and unspecified regions are eligible; edge and restricted regions are excluded. Unknown tax basis is eligible and labelled; tax-included, fractional, GPU-component-only and advertised starting prices are excluded. GPU-hour rates divide instance rates by GPU count; per-GPU rates retain their scope. Calculations use unrounded values. Minima may refer to different regions or SKUs. Region counts include an unspecified-location group when present. Verda uses public instance types and separate on-demand and Spot fields. Regions and live stock are unknown. CPU core semantics and RAM byte convention are unverified, so their original descriptions are shown. No authenticated availability endpoint is queried.

Daily history retains the latest observed rate per provider, SKU, region and price type. Minimum, median with sample count and maximum use the same eligible whole-GPU population as headlines, including unspecified regions, with on-demand and Spot kept separate. Missing days are not filled and stale providers are excluded from today. Coverage changes affect the series.

Past days retain the cheapest offer per provider, GPU model, allocation, price type and region from the latest daily offer observations. Historical distributions use those group representatives; today uses the current listing.

Stale · latest fetch failed or is from an earlier day. Cheapest among observed offers only. Coverage depends on the last collection; this is not a global-market minimum.

HTTP uses approved HTTPS endpoints and refuses redirects, without credentials, cookies or browser impersonation. APIs use backoff and Retry-After. Page requests have no automatic retries; every attempt, including failure, consumes the persisted hourly budget. HTTP 403 stops the source.

Provider methodology

Sesterce

D · Link only Link only: automated price access has not been cleared. Check the official catalog for configuration, region and current prices. Source

Seeweb

D · Link only Link only while review of public-price reuse is incomplete. GPU count and the selected configuration must be verified at the source. Source

Gcore

D · Link only Link only while the website reuse review is incomplete. Check the official pricing model and configuration; inconsistent multi-GPU specifications are excluded. Source

Jarvislabs

D · Link only Link only: access rules and the complete terms could not be verified. Check official prices and billing for your country. Source

Modal

D · Link only Link only pending access review. Serverless, per-second billed; GPU, CPU and memory charges are separate. Source

Koyeb

D · Link only Link only pending access review. Serverless, per-second billed; check instance specifications, region and hourly rates at the source. Source

AWS EC2

A · Official API Official EC2 Linux shared-tenancy prices in two regions, refreshed daily. Spot entries are regional lowest prices, not availability-zone quotes; these floors do not set comparison headlines. Source

Google Cloud

A · Official API Official Cloud Billing Catalog GPU-hour components, including A3/A2/G2 GPU SKUs. GPU only; VM vCPU/RAM billed separately. Collection skips when GCP_BILLING_API_KEY is absent; catalog attempts are limited to once per six hours. Capacity is unknown. Source

Hetzner

B · Public page · display only Public dedicated GPU servers, billed hourly with a monthly cap and an additional one-time setup fee. Configuration starting prices remain marked; display only. Source

Alibaba Cloud International

D · Link only Link only: international website terms prohibit systematic retrieval to build a database and crawling. No pricing collection or export. Source

Tencent Cloud International

D · Link only Link only: international terms restrict benchmarking and use competitive with the services. Anonymous pricing was not verified; the quote API needs a signature. No collection or export. Source

Clore.ai

A · Official API Community marketplace with individual hosts; reliability varies. Whole-server daily USD quotes divided by 24, then by the explicit GPU count. Platform fees may apply. Source

AnyGPU 算力云

C · Display only Normal hourly CNY package prices; member prices are notes only. GPU allocation is unknown unless explicitly stated. Source

智星云

C · Display only Promotional hourly CNY floor prices; allocation and availability unconfirmed. Excluded from headline comparisons. Source

Azure

A · Official API Azure uses curated Linux GPU SKUs, with primary hourly meters and effective-date checks. Source

Linode / Akamai

A · Official API Linode joins exact GPU SKUs to the public region/plan availability feed. Regional hourly overrides take priority; VPU devices are excluded. GPU use is rounded up to hours. Source

Scaleway

A · Official API Scaleway uses complete GPU instance rows with hardware metadata. Money units and nanos retain full precision; minutes and seconds are explicitly converted to hours. Legacy aliases, Apple cores and monthly prices are excluded. EUR is converted with dated ECB rates when available; otherwise EUR remains visible without a USD comparison. Source

Verda / DataCrunch

A · Official API Verda uses public instance types and separate on-demand and Spot fields. Regions and live stock are unknown. CPU core semantics and RAM byte convention are unverified, so their original descriptions are shown. No authenticated availability endpoint is queried. Source

Featurize

C · Display only Featurize agreement 7.1/7.3 restricts reuse of site materials: tier C, display only. Record one GPU only if per-card rental is explicit; otherwise show the instance price with unknown count. Preserve CNY and dated ECB conversions. Source

Nebius

B · Public page · display only Nebius prices cover the GPU only; CPU/RAM are billed separately. Starting Spot prices are excluded from headlines. Do not infer GPU counts or memory. Source

Lambda

B · Public page · display only Lambda prices take effect 2026-11-16 at 00:00 UTC and are excluded from current offers before then. Preserve per-GPU prices and explicit counts; exclude cluster commitments. Source

Paperspace

C · Display only Paperspace hourly GPU compute prices are display only. Unknown counts remain unknown. Starting prices, three-year commitments and special promotions retain their conditions and are excluded from headlines. Source

Baidu GPU

C · Display only Baidu GPU post-paid instance prices are display only. Preserve explicit counts and fractions; exclude prepaid monthly prices. Show CNY originals and dated ECB conversions. Tax basis is unknown. Source

Runpod

C · Display only Link only. Plain HTTP requests returned 403; no bypass, collection or price export. Source

DigitalOcean GPU

C · Display only Link only. Plain HTTP requests returned 403; no bypass, collection or price export. Source

Salad

C · Display only Link only. Plain HTTP requests returned 403; no bypass, collection or price export. Source

Thunder Compute

C · Display only Link only. Plain HTTP requests returned 403; no bypass, collection or price export. Source

Massed Compute

C · Display only Public instance-hour prices, attributed to Massed Compute. Display only under EULA reuse restrictions; region and tax unspecified. Source

Cloud Clusters

B · Public page · display only Public monthly plans with the selected commitment; no verified hourly rental price. Source

Novita AI

D · Link only Link only: site terms prohibit automated extraction of website data. Source

GPUonCLOUD

D · Link only Link only: terms exclude collecting and using product listings and prices. Source

Hyperbolic

D · Link only Link only. No prices are fetched or exported. Source

Vultr

A · Official API Official hourly catalog with explicit GPU counts. Missing locations do not imply available capacity. Preemptible rates are labelled separately; unverified virtual GPU allocations are excluded. Source

Oracle Cloud

A · Official API Official GPU-hour rate multiplied by the reviewed shape GPU count. OCPUs remain distinct from vCPUs. Region and inventory are unspecified. Source

OVHcloud

A · Official API Official FR subsidiary catalog in EUR, excluding VAT. Linux hourly instance prices; exact regions and capacity are unspecified. Local-zone and three-zone prices are not inferred. Source

Vast.ai

D · Link only Link only. No prices are fetched or exported. Source

Together

D · Link only Link only. No prices are fetched or exported. Source

Specification sources

Some links may be affiliate links. This does not affect rankings or prices shown. Provider-only cards carry no scraped price and are excluded from every comparison.

Know the full job cost before you rent: GPU Budget Kit ($12)