GPU Cloud Price Index 2026: AWS Capacity Blocks, Azure, GCP, Lambda, CoreWeave & RunPod

What This Index Is (and Isn't)

This index normalizes published cloud GPU pricing to a single comparable unit — USD per GPU-hour — across two hyperscalers with full on-demand pricing (Azure, GCP), three GPU-specialist "neoclouds" (Lambda, CoreWeave, RunPod), and AWS EC2 Capacity Blocks for ML reservation pricing (see below). Multi-GPU instances (for example, an 8×H100 CoreWeave or GCP instance) are normalized by dividing the published instance price by the GPU count in that SKU; providers that already price per single GPU (Lambda, RunPod, GCP's T4 GPU-attachment SKU) are used as-is. Every price is a snapshot as of the "checked" date shown on each row — cloud GPU pricing changes frequently, so treat any single figure as a point-in-time reference and re-check the source link before budgeting against it.

AWS's regular EC2 on-demand and Spot GPU pricing is still not represented in this index: the standard EC2 pricing pages remain JavaScript-rendered, and the unauthenticated bulk pricing offer file is too large to retrieve in a single request (see the methodology page for detail). This edition does add 6 AWS EC2 Capacity Blocks for ML rows (checked 2026-07-20) — a distinct reservation-fee pricing model where you pre-schedule a future block of GPU capacity in advance, not an instant-launch on-demand rate. Capacity Blocks rows are labeled “Capacity Blocks” everywhere they appear in the tables below and are never blended with on-demand figures. Vast.ai is excluded on principle: it is a peer-to-peer marketplace with prices that vary by listing rather than a single vendor-published rate.

Cheapest Normalized $/GPU-hour by Model

The table below shows, for each GPU model covered by this index, the single cheapest normalized price found across all six providers (on-demand rates for every provider except AWS, whose entries reflect Capacity Blocks reservation pricing instead — see the “Price type” column). This is the citable, at-a-glance view — the full per-provider SKU breakdown (with spot and committed-use pricing) follows in the next section.

Full Pricing by Provider

Every priced SKU in this index, grouped by provider, with on-demand, spot/preemptible (where publicly listed), and 1-year/3-year committed-use pricing, all normalized to USD per GPU-hour. Spot and preemptible prices are volatile point-in-time snapshots, not guaranteed rates — GCP, for example, states that Spot prices are dynamic and can change up to once every 30 days.

AWS

AWS rows below are EC2 Capacity Blocks for ML reservation-fee prices — you reserve a future block of GPU capacity in advance for a fixed duration, which is a different purchasing model from the instant-launch on-demand rates shown for every other provider in this table (see the “Price type” column below and the methodology page for the full explanation). Regular AWS on-demand and Spot GPU pricing is still not represented in this index.

Azure

CoreWeave

GCP

Lambda

RunPod

Methodology Summary

Full detail — including region selection per provider, exactly how multi-GPU instances are normalized, what's excluded and why, and the complete source list — lives on the dedicated GPU Price Index methodology page. In short: one representative, named region per provider where applicable (Azure: eastus; GCP: us-central1/Iowa; CoreWeave: North America; Lambda and RunPod: single global/tier-wide price lists; AWS: no single representative region — EC2 Capacity Blocks pricing is quoted per SKU/region on AWS's own pricing page, so each AWS row states its own region individually); prices sourced only from each vendor's own pricing page or public API, never from third-party aggregators; and any SKU that could not be live-verified at build time is dropped from the index rather than estimated.

This index applies the same GPU-specific purchasing-model logic covered in GPU Cost Optimization and the same commitment/spot mechanics covered generally in AWS Spot Instances Guide — this page exists to make the actual current numbers behind those strategies checkable in one place. For how GPU spend fits the broader AI cost picture, see FinOps for AI.