Live pricing
Re-collected every week from each provider's own pricing API, and kept so the series grows. The interesting number is not today's floor but which way it is moving.
$6.88
cheapest H100 per GPU-hour
94 regions · 82 GPU SKUs
$6.98
cheapest H100 per GPU-hour
59 regions · 59 GPU SKUs
$4.20
cheapest H100 per GPU-hour
46 regions · 912 GPU SKUs
Every cloud region selling GPU compute. Dot size = GPU SKU breadth.
GPU-equipped regions per continent.
Matched region + SKU pairs.
Each dot = one region × SKU.
Median on-demand USD per GPU-hour. One point per weekly snapshot.
How these prices are collected, and what they are not. Azure comes from the public Retail Prices API (pay-as-you-go and Spot, Linux); AWS from the Vantage mirror of the official Price List (on-demand and spot, Linux); Google Cloud from the Cloud Billing Catalog API (on-demand and preemptible). The three are not like for like. AWS and Azure per-GPU figures are derived from whole GPU instances, so they carry bundled vCPU and RAM; Google bills GPUs as standalone per-GPU-hour SKUs that exclude the host VM. Read roughly the host-VM cost as the wedge between them — which is most of why Azure looks cheapest here. List prices only: negotiated and committed-use discounts are not reflected anywhere on this page. The map shows primary cloud regions; AWS Local Zones, Wavelength and operator edge zones are in the dataset but off the map, and same-metro dots are nudged apart so all three providers stay visible.