Best GPU Cloud Providers

Every GPU we can price, from every provider whose prices we can actually read, with the day we read each one. Cheapest first.

How to choose a GPU cloud provider

Six things separate providers, and only the first is on the table above. The rest are on each provider's own page here, where we know them, and blank where we do not.

• Price per GPU hour, which is the table above and the only one of the six we read from an API every day.
• Billing granularity. Per second, per minute or per hour. On a job that runs for eleven minutes, hourly billing costs five times what per-second billing costs for the same work.
• Egress. Moving a trained model or a dataset out can cost more than the compute did. Several specialist clouds charge nothing for it and the hyperscalers do not.
• Interconnect. One GPU is one GPU, but eight of them training together are only as fast as the link between them. NVLink and InfiniBand are the words to look for; without them a multi-GPU node can be slower than the count suggests.
• Spot or on demand. A spot instance is cheaper and can be taken away mid-run. That is a bargain for a checkpointed training job and a trap for an inference endpoint.
• Where the machine is. Latency for inference, and the law for data.

What an H100 actually costs, on 2026-10-21

The cheapest H100 SXM we can verify is $2.20/hr at Vast.ai, against a median of $3.09/hr across the listings we read. Left running, that is the difference between $1586 and $2225 a month for the same card.

The spread inside a single model is the part most comparisons miss. A100 80GB ranges from $0.40/hr to a median of $1.49/hr on the same day, across 8 listings. Quoting one number for that card would be quoting one listing and calling it the market.

Specialist clouds and hyperscalers

The two are not competing on the same thing. A specialist GPU cloud or a marketplace sells the card by the hour and little else. A hyperscaler sells the card attached to identity, networking, managed storage, compliance paperwork and a support contract, and charges for all of it.

If the job is "train this and give me the weights", the specialist is usually several times cheaper. If the job has to sit inside an existing estate with its audit trail, the price of the GPU is not the number that decides it.

We do not have verified prices from the hyperscalers yet. They are in the directory with no price rather than with a number we did not read.

What this page is, and what it is not

It is every GPU we can price, from every provider whose prices we can read, with the day we read each one. Today that is 221 live prices from 3 of the 114 providers we list.

It is not a ranking, and there is no "best provider" line anywhere on this site. A ranking of 3 priced providers would be a ranking of what we happen to have read, dressed up as a verdict. How we measure.

Questions

Where do these prices come from?

Each one is read from the provider's own API and carries the day it was read. Nothing here is copied from a marketing page or estimated. How we measure.

Cheapest or median, which should I use?

They answer different questions. The cheapest is one listing, and on a marketplace it can be a single host briefly undercutting everyone. The median is what the card generally costs across the listings we can see.

Why so few providers?

Because we only publish prices we have read ourselves, and today that is 3 of 114. The rest are listed with no price rather than left out.

What is the cheapest GPU right now?

GTX 1050 Ti at $0.0231/hr from Vast.ai, read 2026-10-21.

How do I work out what a month costs?

Multiply the hourly price by 24 and by 30. The table shows that figure beside every price, and it assumes the machine never stops. A job that runs eight hours a day costs a third of it, and only on a provider that bills by the second or the minute.

Is a spot instance worth it?

For a training run that checkpoints, usually yes: it is the same card for less and losing it costs you the time since the last checkpoint. For an inference endpoint somebody is waiting on, no.