Snapshot 2 · Friday, 14 August 2026
What an hour of GPU time actually costs, normalised across the rental marketplaces.
Every provider names the same silicon differently and buries the price behind a configurator. This is the daily flat table: one row per GPU, the on-demand and spot price from each marketplace we can read, and the cheapest place to rent it right now. Sibling pages: pricing, hot, bench, papers.
Cheapest H100 today
$1.47 /hr
Cheapest today
$0.201 /hr
Cheapest today
$0.134 /hr
18 GPUs · $/hr per single card · cheapest highlighted
| GPU | RunPod on-demand | RunPod community | Vast.ai floor | Cheapest | |
|---|---|---|---|---|---|
| 01 | B200180 GB | $6.79 | $5.98 | $5.31 | $5.31 |
| 02 | H200 SXM141 GB | $3.79 | $0.500 | — | $0.500 |
| 03 | H100 SXM80 GB | $3.29 | $2.69 | $1.47 | $1.47 |
| 04 | H100 PCIe80 GB | $2.89 | $1.99 | $2.13 | $1.99 |
| 05 | MI300X192 GB | $2.39 | $0.500 | — | $0.500 |
| 06 | A100 80GB80 GB | $1.39 | $1.19 | $0.201 | $0.201 |
| 07 | A100 40GB40 GB | — | $1.00 | — | $1.00 |
| 08 | L40S48 GB | $0.990 | $0.790 | — | $0.790 |
| 09 | L4048 GB | $0.820 | $0.690 | — | $0.690 |
| 10 | A4048 GB | $0.440 | $0.350 | — | $0.350 |
| 11 | RTX 6000 Ada48 GB | $0.840 | $0.740 | — | $0.740 |
| 12 | RTX 509032 GB | $0.990 | $0.690 | $0.213 | $0.213 |
| 13 | RTX 409024 GB | $0.740 | $0.340 | $0.134 | $0.134 |
| 14 | RTX 309024 GB | $0.460 | $0.220 | $0.014 | $0.014 |
| 15 | RTX A600048 GB | $0.530 | $0.330 | — | $0.330 |
| 16 | RTX A500024 GB | $0.270 | $0.160 | — | $0.160 |
| 17 | Tesla V10016 GB | — | $0.190 | $0.020 | $0.020 |
| 18 | L424 GB | $0.490 | $0.440 | — | $0.440 |
Each morning the fleet queries the RunPod public GraphQL API and the Vast.ai marketplace bundles endpoint, maps every provider's GPU naming onto a canonical part (an H100 SXM is an H100 SXM whether the listing calls it NVIDIA H100 80GB HBM3 or H100 SXM), and records the lowest price per card. RunPod on-demand is the always-available "secure cloud" rate; RunPod community and the Vast.ai floor are spot-style marketplace prices that can be interrupted and vary by host. GPUs we cannot map with confidence are dropped rather than guessed. Raw JSON: today's snapshot.
Coverage today: RunPod GraphQL (secure + community on-demand); Vast.ai marketplace (cheapest per-GPU offer). Lambda Labs and CoreWeave publish prices as HTML and rate cards rather than clean APIs - they are on the list to add.