Skip to main content
PalCloud

23 Sept 2026 · 6 min read

How to choose a GPU cloud provider

The questions that actually decide the bill: which GPU, on-demand or committed, where the data sits, and what is not in the hourly price.

Most GPU cloud comparisons start with the hourly price. That is the right number to start with and the wrong number to finish with. PalCloud currently tracks 63 providers and 291 published price rows, and the spread between the cheapest and dearest hour for the same GPU is routinely larger than any discount you will negotiate. The job is to work out which of those hours you can actually use.

1. Fix the GPU before you compare anything

The GPU decides the shortlist, because not every provider offers every card. Work out the memory you need first — our GPU memory calculator shows the arithmetic — and then look only at providers who publish a price for that card. Across the 14 GPU familys with a published price on PalCloud today, the lowest per-GPU-hour figure of any family is $0.44/GPU-hr for the L4, but a cheaper card that forces you to shard a model across four GPUs is not cheaper.

2. Decide how you will buy the hour

The same card is sold three ways: on-demand, spot (interruptible, cheaper) and committed or reserved (a term contract). On PalCloud, on-demand and spot are shown in separate columns and never blended, because they are not the same product. Committed prices are usually negotiated rather than published, which is why you will see fewer of them. The rule of thumb: bursty experiments belong on spot, production inference on on-demand, and anything running for weeks is worth a conversation about a term price. More on the three pricing models.

3. Check what the hourly price excludes

  • Storage — block and object, billed per GB-month, and often the reason a “cheap” provider is not.
  • Egress — per TB out. Some providers charge nothing, some charge more than the compute.
  • Public IPv4 addresses, load balancers, snapshots and images.
  • Minimum billing increments and whether a stopped instance still bills for its disk.
  • Taxes: an Indian provider’s price may be quoted before GST, a US provider’s before sales tax.

4. Work out where the data has to sit

Data residency is a shortlist filter, not a preference. If your contracts or your regulator require the data to stay in one country, only providers with a region there qualify. Our provider directory records each provider’s data-centre countries with a source, and for India there is a separate question of MeitY empanelment and IndiaAI status.

5. Ask the boring operational questions

  • Can you actually get the capacity, in the region you need, this week? Published price and available capacity are different things.
  • Is there an API and a Terraform provider, or is it a web console only?
  • What happens to your data and your running jobs if you stop paying?
  • Which certifications does the provider hold, and can you see the certificate rather than a logo?

6. Then compare prices — from the provider’s own page

Every price on PalCloud links to the provider’s pricing page and carries the date we last checked it. Use the GPU cloud comparison to shortlist, the run-cost calculator to turn an hourly rate into the cost of your actual job, and the provider’s own page to confirm before you commit. If a figure here and a figure there disagree, the provider’s page wins — tell us and we will fix ours.

Want the shortlist done for you?

Tell us the workload, region and budget. We pass it to up to 3 matching providers.

Get Best Quotes

More guides

← All guides

Compare published cloud prices

Get Best Quotes