How much does it cost to rent an H100? $1.41 to $10 per GPU-hour, median $3.07
Renting one H100 for one hour costs between $1.41 and $10.00 at the 49 specialist GPU providers publishing hourly rates tonight - median $3.07. That is the direct answer, computed from 78 H100 listings on the panel at publication time, each linked to the provider page it was read from. If you include the hyperscalers, the median jumps to $9.07 per GPU-hour across 222 listings. Same silicon, three times the price.
The distribution
The cheapest published H100 hour tonight is Lium at $1.41 for a single H100 80GB HBM3. Then Sesterce at $1.66, Latitude.sh from $1.68, CloudRift at $1.70, Beam from $1.83 for the PCIe variant, and AIC Cloud at $1.88 out of Mumbai. At the top of the specialist range sit the premium interconnect configs at $10.00 per GPU-hour. In between is where most of the market actually trades: the $2.50-$4.00 band holds the bulk of listings.
| Provider | Config | $/GPU/hr | Location |
|---|---|---|---|
| Lium | H100 80GB HBM3 x 1 | $1.41 | Not published |
| Sesterce | H100 x 1 | $1.66 | Not published |
| Latitude.sh | H100 x 1 (floor, marked from) | $1.68 | Multiple regions |
| CloudRift | H100 x 1 | $1.70 | Not published |
| Beam | H100 PCIe x 1 (floor, marked from) | $1.83 | Not published |
| AIC Cloud | H100 x 1 | $1.88 | Mumbai |
SXM versus PCIe
The interface matters more than most buyers expect. H100 SXM - the NVLink variant with full 3.35 TB/s memory bandwidth and the faster interconnect - medians at $3.35 per GPU-hour across 20 providers, ranging from $1.99 to $6.16. PCIe variants run cheaper but give up the NVLink fabric that multi-GPU training runs need. If your workload is single-GPU inference or fine-tuning, the PCIe discount is free money. If you are sharding a training run across eight cards, the SXM premium is the cost of the interconnect, not the chip.
Why hyperscalers cost three times more
Azure's ND96isr H100 v5 works out to $23.33 per GPU-hour on the published on-demand rate. That number buys things a specialist provider does not sell: the surrounding cloud, the compliance attestations, the enterprise agreement, the integration with everything else you already run there. For a training job that needs bare GPUs, it is a 7x markup over the specialist median. For a regulated workload that needs the audit trail, it may be the cheapest compliant option. The panel lists both, labeled, and lets you decide which market you are in.
Frequently asked questions
What is the cheapest way to rent an H100?
As of September 18, 2026, the cheapest published on-demand H100 rate is $1.41 per GPU-hour at Lium. Marketplace-style providers (Vast.ai, RunPod community) can dip lower on spot capacity but with variable availability; the panel tracks published provider rates with receipts.
How much does an H100 SXM cost per hour?
The median across 20 providers publishing H100 SXM rates is $3.35 per GPU-hour, ranging from $1.99 to $6.16. SXM carries a premium over PCIe because it includes the NVLink interconnect required for efficient multi-GPU training.
Is it cheaper to rent H100s from a hyperscaler or a specialist GPU cloud?
Specialist GPU clouds median $3.07 per H100-hour against $9.07 when hyperscalers are included - roughly a 3x gap on the same silicon. Hyperscalers bundle the surrounding cloud services and compliance posture; specialists sell the compute. Committed-use contracts narrow the gap on both sides.
Method note
All figures are published on-demand hourly rates read from provider pages on September 18, 2026, converted to USD at the day's rates, with commitment, spot and promotional pricing excluded. Floors labeled "from" are marked as floors. Every number links its receipt on the panel.