Estimating
How to estimate what a fine-tune costs
A fine-tune costs the GPU-hours of one run, times the price per GPU-hour, times the number of runs you will really make, plus storage.
Fine-tuning on rented GPUs
Topics
Answer in brief
10 short entries on the questions people ask around fine-tuning on rented GPUs.
Each opens with its answer, then gives the detail and the terms behind it.
All topics
Method
LoRA or full fine-tune: what each costs
A LoRA fine-tune trains a small set of added weights and fits on a smaller GPU, while a full fine-tune updates every weight and needs several times more memory.
GPU choice
Which GPU to rent for a fine-tune
Rent the smallest current GPU whose memory holds your job, because memory you do not use is billed every hour all the same.
GPU count
One GPU or several: how GPU count changes the cost
Adding GPUs shortens a run but rarely lowers its cost, because you pay for every GPU for every hour and the speed-up is less than the count.
Failed runs
What failed runs add to a fine-tune bill
Failed and repeated runs are often the largest part of a fine-tune bill, and no price page lists them.
Storage
Storage, checkpoints and data transfer
Storage is the part of a fine-tune bill that keeps running after the GPU stops: datasets and checkpoints are billed by the gigabyte for as long as they exist.
Pricing models
On-demand, spot or a prepaid term for a fine-tune
Rent on demand while you are still finding settings, and price a term only when the GPU will train most hours of a month.
GPU prices
Fine-tune cost on an RTX Pro 6000
The RTX Pro 6000 is where most LoRA fine-tunes should be priced first: 96 GB on one card, at a lower hourly price than the datacenter GPUs in these tables.
GPU prices
Fine-tune cost on an H200
The H200 is the card to price when a fine-tune needs more than a workstation GPU holds: 141 GB on one datacenter GPU, widely listed.
GPU prices
Fine-tune cost on a B200
The B200 is the card for a full fine-tune of a large model, and the one where the smallest rental changes the bill most.
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.
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.
Continue this reading path
Cite this entry
Fine Tune Cost. “Topics.” Content reviewed 12 Nov 2026. https://finetunecost.com/topics