A GPU-hour has at least three prices, and comparison tables that average them are comparing nothing. Of the 69 GPU plans PalCloud tracks today, 9 publish a spot price alongside their on-demand price and 3 publish a committed or reserved price. That is why our tables keep the three in separate columns and sort only by the published on-demand price.
On-demand: the price you can always get
You start an instance, you pay by the hour or the second, you stop it and the billing stops. No commitment, no interruption. It is the highest of the three prices and the only one that is comparable across providers without caveats, which is why it is the sort key on every PalCloud table.
Spot: cheaper, until it is taken away
Spot (sometimes “interruptible” or “preemptible”) sells capacity the provider is not using, and takes it back when someone pays full price for it. The discount is real — across the plans that publish both, the median spot price sits about 45% below the same plan’s on-demand price — but so is the interruption. Spot works when the job checkpoints and can restart: batch inference, hyperparameter sweeps, data preparation, rendering. It does not work for a production endpoint with a latency promise, and a “cheap” run that is killed at 90% and restarted from zero was not cheap.
Committed and reserved: a term contract
You agree to pay for capacity for a period — a month, a year, three years — and get a lower rate. Two things to check before signing: whether the commitment is to a specific GPU in a specific region (usually yes), and what happens if your workload changes. Most committed GPU pricing is negotiated rather than published, so the number on a public page is often a starting point. If you are running for weeks, it is worth asking: tell us the shape and term and we will pass the request to up to three providers that match.
How to compare them honestly
- Compare like with like: on-demand against on-demand, per GPU-hour, in one currency.
- Price the whole run, not the hour — our run-cost calculator does both on-demand and spot.
- For spot, add the cost of the restarts you expect, and only then compare.
- For committed, divide the total contract by the hours you will genuinely use, not the hours in the term.
Every figure behind this post is in the GPU cloud comparison, each row linked to the provider’s own pricing page with the date it was checked.