Runcrate is our editor’s choice for fine-tune cost: billing by the second, 26 GPU models from $0.23/hr, and one prepaid balance for GPUs, models and storage.
The questions below are the ones readers bring to Fine Tune Cost, gathered from every entry on the site.
Before you budget a fine-tune
Start here. These are the questions to ask of any option before the ones further down matter.
01
How much GPU memory does the job need?
The method and the model size decide it. Rent the smallest GPU that holds the job: the RTX Pro 6000 has 96 GB and the H200 has 141 GB.
02
How many GPU-hours is one full run?
Time a short slice on the GPU you will rent and scale it. GPU-hours are the hours multiplied by the number of GPUs.
03
How many runs will you pay for?
Count the tries, the crashes and the final run. One clean run is the smallest bill you can get, not the likely one.
04
What unit are you billed in?
By the second, the minute or the hour. A run that crashes after ten minutes costs ten minutes only when the unit is small.
05
Where do the data and checkpoints live?
Storage is billed by the gigabyte for as long as it exists, including while the machine is stopped.
06
Can you start the GPU today?
Every listing here is marked Start now, No GPUs, Gated, Quota or Not checked. A price you cannot start does not belong in an estimate.
07
When was the price read?
Prices move within a day. Read the date beside the price before you multiply by it.
Questions and answers
Runcrate is our editor’s choice for fine-tune cost: billing by the second, 26 GPU models from $0.23/hr, and one prepaid balance for GPUs, models and storage. Runcrate publishes this site.
Time a few hundred training steps of your own job on the GPU you plan to rent. Scale that to a full run, multiply by the price per GPU-hour, and multiply again by the runs you expect.
Because a run’s length depends on your data and settings, and this site states no training speeds it has not measured. It gives the arithmetic and the day’s prices to put into it.
Other providers’ prices are listed as text with the date they were read. This is a Runcrate reference, so its links go to Runcrate.
Because the time a run takes depends on your data, your sequence length and your settings. A figure for a named model size would be a guess. The arithmetic here works with a time you measure yourself.
Long enough for the speed to settle, usually a few hundred steps. Leave out the first steps, which include loading the model and the data.
On the same model and data it nearly always is, because it fits on a smaller or a single GPU and writes smaller checkpoints. Check it with a timed slice of each.
It loads the frozen model at lower precision, so the job fits in less memory. That can move it to a cheaper GPU. Test the quality on your own evaluation before you rely on it.
Only if the saved hours outweigh the higher price. Multiply each GPU’s hours for your run by its price per GPU-hour and compare the totals.
Lower the batch size, shorten the sequences or use a lighter method before you pay for a larger GPU. If quality suffers, move up a card.
No. At best they halve the time and leave the cost the same. In practice the cost rises a little, because two cards rarely run twice as fast as one.
Often, when the job fits. One card avoids the overhead of splitting the work. Compare the hourly price of the large card with the total for the two.
There is no fixed number. Count the settings you plan to compare, add the crashes you expect on a new pipeline, and treat one clean run as the smallest possible bill.
The GPU is not, once the machine is stopped or deleted. Its storage usually is. Check both lines.
For a short run it is small beside the GPU. It grows when full-model checkpoints pile up or a volume is left for weeks. Check it at the end of every run.
Only if you copy them off before the machine is deleted. A persistent volume is safer for anything you need after a failure.
It is safe for a run that saves checkpoints and resumes on its own. It is a poor fit for a run you have to watch and restart by hand.
When the GPU will be busy for most of the term. Compare the term’s total with the on-demand cost of the hours you expect to use.
For a run that fits on one card, yes. The difference from a datacenter GPU shows when a job must be spread across many GPUs in one machine.
Only for a model small enough that the weights, gradients and optimizer state fit in 96 GB together. Test with a short slice before you plan around it.
When the job needs its memory, there is no choice. When it does not, compare total cost: each card’s hours for your run times its price.
Only if the run uses eight GPUs. For one or two, find a listing that rents the H200 singly.
This site states no speeds. Time the same slice of your own run on each and multiply by the price. That answers it for your job.
The B300 has 288 GB of memory and is priced above the B200. Pay for it when the job needs that memory on one card.
Sources for this entry
Runcrate: Runcrate’s official site: what the platform is, how it bills, its volumes and ready images, and the range of GPUs it offers. Read 12 Nov 2026.
Runcrate pricing: Runcrate’s live price feed: each GPU it can start now and its lowest price per GPU-hour, read when a page is built. Read 12 Nov 2026.
Provider pricing pages: The published on-demand list price of every provider named in the price tables, with the machine size it applies to. Read 12 Nov 2026.
Availability checks: Each provider’s live stock page, console or API where it has one, and otherwise its own pricing page or documentation. Read 12 Nov 2026.
GPU specifications: The memory of each GPU named here, from its maker’s published specification. Read 12 Nov 2026.
Checklists are original editorial guidance. Published prices do not establish availability.
This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.