H100 SXM Rental Price: Cheapest Cloud per Hour (2026)

This page measures cloud rental prices for H100 SXM using 8 tracked sources, updated 2026-08-23.

Updated 2026-08-23 · refreshed daily

Choose spot at $1.63 for the cheapest H100 SXM rate; choose on-demand at $1.74 instead when interruptions are unacceptable, especially as tracked rates span 3.42 times the cheapest rate.

H100 SXM is the accelerator in the 1x H100 SXM configuration offered through both on-demand and spot capacity.

Before committing, run a one-hour test at $1.74 and confirm that the workload fits, starts correctly, and meets the required throughput.

Can you configure your H100 SXM workload to satisfy FP8, BF16, or FP32 workloads within 80GB VRAM? If the workload meets FP8, BF16, or FP32 workloads within 80GB VRAM, proceed to price H100 SXM on-demand from $1.74; otherwise, exclude this GPU regardless of price.

Do faster time-to-result or a longer uninterrupted run matter more than maximizing price savings? Choose spot at $1.63 when price savings matter more; choose on-demand at $1.74 when faster completion or an uninterrupted run matters more.

Can your H100 SXM workload tolerate interruption and resume from checkpoints without costly recovery? If the workload can restart from a checkpoint, choose spot at $1.63; if it cannot tolerate interruption, choose on-demand at $1.74; if unsure, run a short on-demand test at $1.74 before committing to spot.

Today's prices

USD/hr · lowest on-demand price per provider

1 GPU

Provider	H100 SXM	Availability
Vast.ai	$1.74	In stock
Hyperstack	$3.20	Not reported
DataCrunch	$3.25	In stock
RunPod	$3.29	In stock
Crusoe	$3.90	Not reported
Together AI	$3.99	Not reported
Paperspace	$5.95	Not reported

8 GPUs

Provider	H100 SXM	Availability
CoreWeave	$49.24	Not reported

3.42x

A 3.42 times on-demand price spread makes provider choice material, so compare offers with the $1.74 low end before committing.

6.6%

Potential hourly savings from choosing spot instead of on-demand, before accounting for interruption costs.

$1270/mo

Budget $1270/mo per month at the lowest on-demand rate, and compare providers because the tracked price gap reaches $36879.60 per year.

What your workload needs

The cheapest figure depends on how long you hold the GPU; match your expected run length to the closest row before reading its price.

Workload	VRAM needed	H100 SXM	Estimated cost
70B Q4 inference	42 GB	1 GPU	$1.74/hr
70B FP16 inference	168 GB	3 GPUs	$5.22/hr
70B QLoRA	46 GB	1 GPU	$1.74/hr
7B full fine-tune	112 GB	2 GPUs	$3.48/hr
70B FP8 training	154 GB	2 GPUs	$3.48/hr

What the numbers say

H100 SXM on-demand and H100 SXM spot use the same accelerator configuration, but differ in interruption risk and price. The cheapest spot rate is $1.63, compared with the cheapest on-demand rate of $1.74; across providers, on-demand rates span 3.42 times from low to high, while continuous use at the cheapest on-demand rate costs $1270/mo per month. Choose spot for checkpointable work and on-demand when the run cannot tolerate interruption.

H100 SXM has the same published peak-performance baseline across the compared offers in the NVIDIA Tensor Core GPU datasheet, not measured application throughput, so no provider gains a hardware-performance advantage from the accelerator itself. The effective-cost difference therefore comes from rental terms and operational constraints, and readers should benchmark their own workload before treating the published specification as achieved performance.

Exceptions apply when your workload cannot meet FP8, BF16, or FP32 workloads within 80GB VRAM, which rules out H100 SXM regardless of price, or cannot satisfy checkpointable work that tolerates interruption, which makes spot unsuitable. If either fit is uncertain, test H100 SXM briefly with on-demand at $1.74; the rational default is to keep on-demand for work that cannot restart and move to spot only after checkpoint recovery is proven.

How Marlin helps

To turn these requirements into a selection, Marlin matches the workload to the lowest-priced suitable GPU option across its supported providers.

Full provider reach: Marlin searches every supported provider, including options beyond those priced on this page.
One-pass comparison: Marlin compares supported providers in one matching process, reducing manual price checks across separate sites.
Qualified lowest price: Marlin returns the cheapest option that satisfies your workload requirements, rather than the cheapest row regardless of fit.

Instead of repeatedly checking provider pricing pages and reconciling rates by hand, let Marlin compare supported options when matching your workload to the lowest-priced suitable choice, then validate it with a one-hour test at $1.74.

Before you rent

Data transfer: Egress charges may sit outside the hourly GPU rate; verify them in the provider's network-pricing page and estimate the data your workload will move.

Restart overhead: Capacity interruptions can force a checkpointable job to repeat work, increasing paid GPU hours; measure lost compute between checkpoints during a short trial.

Capacity availability: A listed rate does not ensure that a host will be available for the full run, so verify live inventory and identify a fallback provider before scheduling the workload.

Billing granularity: Per-second, per-minute, or per-hour rounding changes the effective cost of short runs by billing unused portions of a time increment; confirm the applicable rule on the provider's own billing page before committing.

Monthly and annual costs are modeled from hourly rates, not observed invoices or completed workloads. Rental prices were collected from provider pricing pages, while fixed GPU specifications came from the NVIDIA datasheet. Re-check the inputs and their source URLs in the published dataset. https://vast.ai/pricing · https://www.hyperstack.cloud/gpu-pricing · https://datacrunch.io/pricing · https://www.runpod.io/pricing · https://crusoe.ai/cloud/pricing · https://www.together.ai/pricing · https://lambda.ai/service/gpu-cloud · https://www.coreweave.com/pricing · https://resources.nvidia.com/en-us-tensor-core/nvidia-tensor-core-gpu-datasheet

Stop comparing. Start running.

Stop tracking prices by hand and use Marlin to match your requirements to the lowest-priced suitable option across supported providers.

Get started with Marlin