Buying compute

GPU-hour vs instance-hour: what are you actually paying for?

GPUIndexes Research··2 min read
THE SHORT ANSWER

A GPU-hour is one GPU used for one hour. An instance-hour is the entire server or virtual machine used for one hour. Dividing an eight-GPU instance price by eight gives a comparison unit, but it does not make a one-GPU rental available at that rate.

Normalize the unit, then check the bill

Consider an illustrative eight-GPU instance priced at $24 per hour. Its normalized rate is $3 per GPU-hour. If the provider rents only the whole instance, one hour still costs $24. A separate one-GPU instance at $4 per hour has a higher normalized GPU price but a smaller total hourly commitment.

These numbers illustrate the calculation; they are not observed provider quotes. A useful table should show GPU count, the normalized rate and the whole-instance price together.

Illustrative offerGPUsInstance-hourGPU-hour
Whole server8$24$3
Single-GPU instance1$4$4

The advertised compute rate is one part of cost

Check how the provider bills partial hours, whether capacity is interruptible, whether a commitment is required, and what CPU, memory and storage are attached. Network transfer and persistent storage may be charged separately. Credits and promotions can also change the payable amount without changing the public rate.

A rental reference compares an explicitly defined set of advertised observations. It is not a complete invoice estimate. Comparing references across pricing lanes requires reconciling the terms rather than treating every hourly number as equivalent.

Why the initial GPUIndexes basket uses one GPU

The initial benchmarks require exactly one GPU per instance. That removes the multi-GPU commitment difference from the basket and makes the instance-hour and GPU-hour base compute prices numerically equal for the included offers.

Other differences remain, including region and non-GPU resources. The methodology names those limits, and the quote export supplies the observed provider and region so you can investigate an offer before booking.

An offer count is not a GPU count

Several configurations or regions can refer to shared underlying capacity. Ten qualifying offers do not establish that ten physical GPUs are simultaneously available. Provider coverage is likewise a count of provider entries with eligible observations, not verified market share.

Use these measures to understand the sample behind a reference price. The provider remains the source for confirmation of bookable capacity and the final bill.

Check the observed rental market

These observations update separately from the guide. Use their capture timestamps when citing a price.

H100 SXM rental reference →

The current H100 SXM provider-floor median is $3.2820 per GPU-hour, based on 7 qualifying providers in the snapshot from Sep 16, 2026 at 21:56 UTC. The basket covers 80 GB and exactly one GPU per instance, with on-demand USD pricing.

RTX A6000 rental reference →

The current RTX A6000 provider-floor median is $0.5300 per GPU-hour, based on 5 qualifying providers in the snapshot from Sep 16, 2026 at 21:56 UTC. The basket covers 48 GB and exactly one GPU per instance, with on-demand USD pricing.

Common questions

Does $3 per GPU-hour mean I can rent one GPU for $3?

Only if a qualifying single-GPU offer is available at that price. A normalized rate for an eight-GPU server may require renting all eight GPUs and paying the whole-instance bill.

Do GPUIndexes references include storage and network charges?

They describe the advertised base rental observations within the defined basket. Other resources, transfer charges, taxes and provider billing terms can affect the final invoice.

Sources & further reading

  1. GPUIndexes basket and pricing-unit methodology
  2. Inspect the underlying source quotes

Published by GPUIndexes, operated by Quanta Cloud LLC. Ownership, editorial policy and corrections.

Continue reading