H100 rental price per hour

$1.99 per GPU-hour at Voltage Park, which publishes it as a "starting at" figure for one H100 for one hour, read 2026-11-12. That is the lowest of 44 current on-demand rows from 19 providers. The median provider's lowest price is $3.90, and the highest is $11.06 at Google Cloud.

The lowest 8-GPU configuration is Vultr at $2.30 per GPU, which is $18.40 per hour for the server. 10 providers list 8 GPUs, from $2.30 to $12.29 per GPU ($18.40 to $98.32 per server-hour).

Amazon Web Services, DigitalOcean (Paperspace), Runcrate, RunPod, and Thunder Compute state per-second billing on at least one H100 row. 8 of 19 providers state no billing unit on any of their H100 rows, so the data records it as not stated rather than assuming one.

variant not stated: $1.99 to $12.29 across 15 rows from 9 providers; HGX: $2.30 to $6.16 across 7 rows from 6 providers; PCIe: $2.50 to $2.89 across 2 rows from 2 providers; NVL: $2.92 to $6.98 across 10 rows from 3 providers; SXM: $2.95 to $5.09 across 10 rows from 6 providers. The variant is whatever the provider's page names; a row with no named variant is grouped as not stated. Rows from marketplaces and announced prices are left out of these ranges.

8 of 52 current rows are marketplace or "from" figures set by hosts rather than by the provider: RunPod, Vast.ai, and TensorDock. They carry a marketplace chip and are left out of every statistic and the ranking.

Yes. 6 rows from Lambda and Crusoe carry a future start date, with the earliest on 2026-11-16. They are listed under the rate card and are not counted until they take effect.

The lowest solid spot row is Microsoft Azure at $1.42, 29% below the lowest on-demand price. The lowest solid reserved row is Hyperstack at $1.75, 12% below the lowest on-demand price. Spot capacity can be interrupted and reserved capacity is paid for over a term, so the two are not like for like with on-demand.