PetaPrice

Corrections

Prices are collected automatically, so mistakes happen: a changed price list, a mislabeled GPU, a node price shown per GPU. If you spot one, please tell us.

How to report a price

Email us at

Please include:

  • The GPU and the provider
  • The price PetaPrice shows and the one you see
  • A link to the provider's official pricing page

What happens next

We check the price against the provider's official pricing page. If we were wrong, we fix the source or the way we read it, not just the one number, so the error doesn't come back with the next refresh. Fixes are listed below.

Correction log

What was wrong, what it says now, and why. Newest first.

  1. 8 corrections

    • Google Cloud

      Google Cloud prices include the machine

      Before
      Some Google Cloud rows priced the GPU alone, e.g. the H100 at $9.80 per GPU-hour.
      Now
      Every Google Cloud price includes the machine type the GPU needs (H100: $10.98).
      Why
      Google Cloud doesn't rent these GPUs without a machine, so the GPU-only price can't be booked.

    • RunPod

      Fractional GPUs removed

      Before
      A RunPod RTX PRO 6000 MIG slice was listed as a full GPU at $0.59 per hour.
      Now
      MIG slices, vGPUs and time-shared GPUs are rejected from every source.
      Why
      A slice has a fraction of the compute and memory, so its price isn't comparable per GPU.

    • Shadeform

      GPU variants read from the configuration

      Before
      Some A100 40 GB listings were labeled 80 GB, and some PCIe H100s were labeled SXM.
      Now
      Memory and interconnect come from each instance's configuration.
      Why
      Variants differ in price and performance; a PCIe price must not count as SXM.

    • Massed Compute, Latitude.sh

      Read from the providers' own pricing pages

      Before
      Prices came from an aggregator whose rates differed by up to 73%.
      Now
      Both are read directly from their pricing pages.
      Why
      The provider's own price is the one you pay.

    • Vultr

      Bare-metal rates corrected

      Before
      8× H100 at $2.99 per GPU-hour and MI300X at $3.99.
      Now
      Vultr's published rates: $2.30 and $1.85.
      Why
      The old figures came from an outdated catalog.

    • Azure

      Edge zones and non-public SKUs excluded

      Before
      A T4 price from an operator edge zone was 35% below Azure's public price.
      Now
      Only public regions and SKUs are used.
      Why
      Edge zones and private SKUs can't be booked by most customers.

    • Crusoe, Hyperstack, Together AI

      Node prices divided per GPU

      Before
      Some SXM prices were for whole 8-GPU nodes but treated as single GPUs.
      Now
      They are divided by the GPU count like every other node price.
      Why
      All prices on PetaPrice are per GPU-hour.

    • All sources

      Stale catalogs dropped

      Before
      Third-party catalogs were used however old they were.
      Now
      Catalogs not updated for 45 days or more are rejected.
      Why
      An old catalog shows prices that may no longer exist.
Methodology →