1. Is the price per GPU or per machine?
Some providers quote one GPU, others quote a whole instance. This site shows a per-GPU figure; where a provider lists only an instance price (CoreWeave does), the figure is the instance price divided by its GPU count and the row says so. On Lambda the per-GPU rate also depends on the instance size. At the time of writing its page lists the H100 SXM at $4.59 per GPU-hour in an 8x instance, $4.69 in 4x, $4.79 in 2x and $4.89 as a single GPU, so a smaller machine costs a little more per GPU. Lambda also states that the prices on its page take effect on 16 November 2026 at 00:00 UTC.
2. How finely is time billed?
Billing granularity changes the real cost of short jobs. RunPod states that Pods are billed per second. Hyperstack states that its pricing is billed per minute. A provider that rounds up to a full hour makes a ten-minute test cost a full hour. Check the billing line on the provider’s pricing page and on its checkout screen.
3. Which tier is the price for?
- Secure vs community. RunPod lists two prices for many GPUs. For the H100 SXM it shows $4.39 per hour on Secure Cloud and a Community Cloud rate starting from $2.69. The Community figure is a starting price, and availability can differ, so this site labels it “from”.
- On-demand vs spot. Hyperstack lists a spot price for the H100 PCIe of $2.00 per hour against $2.50 on-demand. Spot capacity can be interrupted, so it suits work that can restart from a checkpoint.
- On-demand vs reserved. Hyperstack’s reservation prices start from $2.72 for the H100 SXM against $3.20 on-demand, and from $1.36 against $1.60 for the A100 SXM, in exchange for a commitment. This site compares on-demand prices only.
- Pods vs serverless. RunPod lists its serverless H100 workers at $5.89 per hour against $4.39 for a Secure Cloud pod. Serverless can still be cheaper overall if your GPU would otherwise sit idle, because you only pay while workers run.
4. What else will appear on the bill?
Storage is the usual surprise. RunPod, for example, lists container disk at $0.10 per GB per month, volume disk at $0.10 per GB per month while running and $0.20 while idle, and standard network storage at $0.07 per GB per month under 1 TB. A 500 GB dataset kept for a month on a $0.10 volume adds $50 before any GPU time. Also check data-transfer charges, public IP charges and any minimum top-up or expiry on prepaid credit.
5. Work out the cost per job, not per hour
job cost = price per GPU-hour × number of GPUs × hours used
+ storage kept × storage price × months
+ idle time you were still billed for
Example: a training run needs 8 GPUs for 6 hours at $3.00 per GPU-hour. That is 8 × 6 × $3.00 = $144. If the machine then sits idle for two hours before you stop it, you pay another 8 × 2 × $3.00 = $48. Automatic shutdown after a job finishes often saves more than switching to a slightly cheaper provider.
6. Does the GPU fit your model?
A cheap GPU that runs out of memory costs nothing to rent and everything to debug. Check VRAM first with the VRAM calculator, then compare providers using the filters on the price table. Prices change; every row on this site shows the date it was last checked and links to the provider’s own page, which is the source of record.