Renting an H100 by the hour

The H100 comes in forms that rent at different rates: SXM modules in eight-GPU nodes with a fast interconnect, PCIe cards in ordinary servers, and NVL cards with more memory. Hourly prices split again by tier, from community and marketplace hosts to secure data-center and committed capacity. Compare per-GPU rates on the same form and memory size, and check stock before planning a job around a price. rent.dev starts H100 pods on the agent's own key at the GPU provider rent.dev connects to, and its quotes API lists hourly GPU rates without a key.

rent.dev: one API for the sandboxes and GPU pods an agent needs.

Every H100 row on the board, per GPU hour

Rates as each vendor publishes them, normalised per GPU, from rent.dev's price board, generated 2026-11-14. See the whole board. Browse the full catalog.

GPUVendorTierGPUs per rowPer GPU hourStockRate read
H100 machineBeam1$1.832026-11-11
NVIDIA H100 PCIeRunPodcommunity1$1.99None2026-11-11
Nvidia H100Daytonapreemptible1$2.272026-11-11
NVIDIA HGX H100CoreWeavespot8$2.463752026-11-11
H100 SXM5 80GBVerdaspot4$2.54252026-11-11
H100 SXM5 80GBVerdaspot2$2.54352026-11-11
H100 SXM5 80GBVerdaspot8$2.543752026-11-11
H100 SXM5 80GBVerdaspot1$2.5442026-11-11
H100 SXM5 80GBVerdaspot1$2.5442026-11-11
NVIDIA H100 NVLRunPodcommunity1$2.59None2026-11-11
NVIDIA H100 80GB HBM3RunPodcommunity1$2.69Medium2026-11-11
h100 80Gi SXM5Akash1$2.732026-11-11
NVIDIA H100Northflank1$2.742026-11-11
NVIDIA H100 PCIeRunPodsecure1$2.89None2026-11-11
NVIDIA H100 NVLRunPodsecure1$3.19None2026-11-11

Questions

What is the cheapest H100 GPU rental by the hour?
Usually a PCIe H100 on a community or marketplace tier, or an interruptible SXM slot that can be reclaimed mid-job. Those rates trade reliability, interconnect or support for price. For multi-GPU training the SXM form is worth its higher rate; for single-GPU inference a PCIe card is often enough. The lowest current rows are on rent.dev's price board at rent.dev/prices. The lowest H100 row on rent.dev's price board is the H100 machine at $1.83 per GPU hour (Beam), read 2026-11-11.
How does H100 cloud pricing compare across providers?
Normalise before comparing. Some providers quote a whole eight-GPU node per hour and others a single GPU; some include CPU, RAM and local disk, others charge storage and data transfer apart. Divide node prices by the GPU count, match SXM to SXM and PCIe to PCIe, and note whether a rate is on-demand, interruptible or tied to a commitment. On rent.dev's price board, generated 2026-11-14, 40 H100 rows run from $1.83 per GPU hour (Beam) to $12.29 per GPU hour (Azure); 14 of them price nodes of several GPUs, shown here per GPU.
How do I find on-demand H100 availability by API?
Ask before you create. GET https://api.rent.dev/v1/quotes answers without a key and lists GPU rows by type with an hourly rate and, where the provider reports one, a coarse stock level. The rent_quote tool over MCP is a price list: it prices sandbox tasks and lists the same hourly GPU rates, but it carries no stock level, so read stock from the API. Stock moves by the hour, so read a stock level as a hint. A create for a GPU type that is out of stock fails, and a pod that starts is the only confirmation. On the board generated 2026-11-14, the best H100 stock reading is Medium: the NVIDIA H100 80GB HBM3 at $2.69 per GPU hour (RunPod, community).
Is an H100 worth it over an A100?
For transformer training and large-model inference an H100 usually finishes the same work sooner, which can make it cheaper per job despite a higher hourly rate. For small models, or jobs held back by data loading rather than compute, an A100 or a smaller card often costs less in total. On rent.dev's price board, generated 2026-11-14, the lowest H100 row is $1.83 per GPU hour and the lowest A100 row $0.837 per GPU hour (A100 SXM4 40GB, Verda, spot).