Cloud GPU pricing: what an H100, A100, L40S or RTX 4090 costs per hour

GPU prices per hour, as of October 2026

These are on-demand prices per GPU-hour in US dollars across 8 providers, as of October 2026, checked on each vendor's own page on 30 October 2026. Marketplace and "from" prices move daily, so confirm the rate on the order page before you launch.

Why the same GPU costs twice as much elsewhere

An H100 goes from $1.73 on one site to $5.95 on another. A lot of that comes down to whether you rent from a marketplace, a community cloud or a secure cloud. Runpod splits its inventory down the middle: on Community cloud an RTX 4090 is $0.34 and an H100 SXM is $2.69, while on Secure cloud the same 4090 is $0.89 and the H100 SXM is $3.99. Newer cards follow the same split: an H200 is $3.59 versus $5.29, and a B200 is $5.98 versus $7.99.

Marketplaces like Vast.ai post live rates that move constantly, which is how Vast.ai lists an A100 SXM4 from $0.40 with a median of $0.80 on 30 October 2026. Interruptible instances there are listed as 50% or more cheaper than on-demand.

Serverless platforms price differently again. Modal charges $3.95 for an H100, $2.50 for an A100 80GB and $1.95 for an L40S, billed per second, and as Modal puts it, "You never pay for idle resources." Lightning AI also bills by the second, at $5.45 for an H100 and $1.85 for an A100 80GB.

Instance size changes the per-card rate too. At Lambda, an H100 SXM is $3.99 per GPU on an 8x instance, $4.09 on a 4x, $4.19 on a 2x and $4.29 on a single GPU, while a single H100 PCIe is $3.29. Lambda's prices exclude sales tax and VAT. Form factor matters as well: at Runpod an H100 PCIe is $1.99 on Community and $2.89 on Secure, against $2.69 and $3.99 for the SXM.

The cheapest options by GPU

For the lowest H100 rate, Vast.ai starts at $1.73 for an H100 SXM with a median of $2.40, and Runpod Community has the H100 PCIe at $1.99. For SXM at a fixed price, TensorDock starts at $2.25 and Runpod Community charges $2.69, followed by Thunder Compute at $3.20, Modal at $3.95, Runpod Secure and Lambda's 8x instances at $3.99, Lightning AI at $5.45 and Paperspace at $5.95.

For an A100 80GB, Vast.ai's live rates for an A100 SXM4 start at $0.40. At fixed prices, Thunder Compute is $1.09, Runpod Community is $1.19 for PCIe and $1.39 for SXM, Lightning AI is $1.85, Modal $2.50, Lambda $2.79 on an 8x instance and Paperspace $3.18.

For the RTX 4090, you're mostly looking at three places: Vast.ai from $0.26 with a median of $0.48, Runpod at $0.34 on Community and $0.89 on Secure, and TensorDock from $0.35. For the L40S, Vast.ai runs from $0.42, Runpod is $0.79 on Community and $1.09 on Secure, Modal $1.95 and Lightning AI $3.54.

Costs that are not in the hourly price

The price per GPU-hour isn't the whole bill. On Runpod, storage gets more expensive when you stop a pod: the volume disk costs $0.10 per GB a month while the pod runs and $0.20 while it's stopped. You need at least one hour's worth of credit to deploy a pod, and when your balance hits $0, pods are stopped, and pods without a network volume are terminated with their data.

Stopping means different things elsewhere. Lambda instances can only be launched, restarted or terminated, never stopped, so an idle instance bills at the full per-minute rate until you terminate it. A powered-off Paperspace machine stops compute billing, but its storage and public IP keep billing by the hour until you destroy them. Thunder Compute bills instances and their disk per minute only while they run, and snapshots separately.

Which provider for which job

For a weekend fine-tune on an RTX 4090, Vast.ai, Runpod Community or TensorDock keep the bill down, and Vast.ai and Runpod bill by the second.

For a multi-day H100 run, weigh marketplace rates like Vast.ai against fixed rates like Thunder Compute at $3.20, Runpod at $2.69 or $3.99 for SXM, or an 8x H100 instance on Lambda at $3.99 per GPU billed per minute. On Lambda, be ready to terminate the instance as soon as the weights are saved, because it can't be stopped.

For bursty jobs where you don't want to manage a machine, Modal bills per second and its Starter plan includes $30 of compute a month.

The biggest cost is the hours you don't use

A run that finishes at 3 a.m. and sits until you wake up bills the same per hour as training. An 8x H100 instance on Lambda costs $31.92 an hour, so an idle night of 8 hours is $255.36. A job that hung hours ago costs the same while the GPU stays allocated.

Keep the stall window at 30 minutes or less. Pair the alert with your provider's stop command at the end of the script, so a run that finishes cleanly stops billing on its own.