What AI compute actually costs
Every month, from our own price index: what an hour of GPU compute cost, how far spot ran below on-demand, and how much the same machine varied by region.
Latest: September 2026
5.52 million spot price observations on AWS, 41 GPU machine types.
- On spot, an hour on an H100 80 GB cost $2.57, less than an hour on an A100 80 GB at $3.85. Newer and bigger is not always more expensive.
- Since August, T4g 16 GB spot rose 18% and L40S 48 GB spot rose 15%. A100 80 GB held steady at $3.85 per GPU-hour.
- H100 80 GB spot cost 64% less than on-demand: $2.57 against $8.43 per GPU-hour.
- Per GB of GPU memory (which decides how large a model fits), T4g 16 GB was cheapest at $0.012 an hour. Among cards with 48 GB or more, L40S 48 GB was cheapest at $0.029.
- Where a machine runs matters: g5.16xlarge (A10G) cost 10.2 times as much in São Paulo (sa-east-1) as in Stockholm (eu-north-1).
- Stockholm (eu-north-1) was the cheapest region overall: machines there cost about 54% less than their typical price across regions, and every one of them came in under it. Tokyo (ap-northeast-1) was the most expensive, about 29% more.
- H100 80 GB spot barely moved all month: on a typical day it was within about 2% of its average. L40S 48 GB moved around much more, about 14%.
Prices by GPU
Spot price per GPU-hour, latest report. Cheapest first.
Get the next report
GPU prices every month, by GPU and by region. One email a month; we’ll ask you to confirm first.
Free. One-click unsubscribe in every issue.
Let the price come to you.
Otium is in closed alpha. Set a deadline for your batch AI work, and it waits for the lowest-cost compute available before then, so you don’t have to watch these prices yourself. Leave your email and we’ll send an invite when a spot opens.
No credit card, no spam — one email when your invite is ready.
Closed alpha — onboarding is gated while we calibrate. Already invited? Sign in.