Estimator
The full fine-tune cost calculator
Seven inputs, five steps, one table of providers. Change any assumption and the arithmetic updates. Prices are published on-demand list prices read from provider pages.
Estimate, step by step
- 1. Training FLOPs = 4 × 7 billion parameters × 1000 million tokens = 2.80 × 10^19
- 2. Useful FLOP/s per GPU = 362.05 TFLOPS × 35% MFU = 1.27 × 10^14
- 3. GPU-hours = step 1 ÷ step 2 ÷ 3,600 = 61.4
- 4. Wall-clock on 8 GPUs = step 3 ÷ 8 = 7.7 h, assuming the work spreads evenly
- 5. Cost = step 3 × each provider's lowest listed price per GPU-hour
| Provider | Billing | $/GPU-hr | Estimated cost |
|---|---|---|---|
| Runcrateours, live | per second | $0.55 | $34 |
| CloudRift | per second | $0.69 | $42 |
| Massed Compute | billing not stated | $0.97 | $60 |
| RunPod | per second | $1.09 | $67 |
| Koyeb | per second | $1.20 | $74 |
| Civo | per hour | $1.29 | $79 |
| Crusoe | per hour | $1.50 | $92 |
| DigitalOcean (Paperspace) | per second | $1.57 | $96 |
| Vultr | per hour | $1.67 | $103 |
| OVHcloud | per hour | $1.80 | $110 |
| Amazon Web Services | per second | $1.86 | $114 |
| Verda | billing not stated | $2.00 | $123 |
| CoreWeave | billing not stated | $2.25 | $138 |
| Oracle Cloud | billing not stated | $3.50 | $215 |
Range $34 to $215, median $94 across 14 providers. An estimate, not a quote: real runs differ in utilisation, memory fit, communication and restarts, and providers that bill per hour round up.
The inputs
What each field means
- Model parameters
- Billions of parameters being trained or run through the model. For LoRA this is still the base model’s size, because every token passes through it.
- Training tokens
- Millions of tokens processed during training. For several epochs, multiply the tokens by the number of epochs.
- Method and factor
- Full fine-tune uses 6 FLOPs per parameter per token, LoRA or QLoRA 4. The factor is editable: raise it to model activation checkpointing, which recomputes work.
- GPU
- Only GPUs with a dense BF16 figure printed by the maker. The figure is peak, so it is multiplied by MFU.
- Number of GPUs
- Divides the GPU-hours into wall-clock time on the assumption that work spreads evenly. It does not change the cost.
- MFU
- Model FLOPs utilisation: the share of peak throughput that training achieves. The default is 35%. It is an assumption, not a measurement of any provider.
Questions
The calculator: common questions
Why does the GPU count not change the cost?
GPU-hours are the same whether you use one GPU for a long time or many for a short time, if the work spreads evenly. Real clusters lose some efficiency to communication, which this estimate does not model.
What MFU should I use?
We default to 35% and do not claim it is right for your job. Compare results at a few values; the scenario pages show a small table of GPU-hours at other MFUs.
Does it include storage or egress?
No. It prices GPU time only, at each provider’s lowest listed on-demand rate for the GPU.
Disclosure
Fine-Tune Cost Lab is published by the Runcrate team, who also run Runcrate, an AI cloud. Runcrate is listed in the table, marked "ours, live". It is sorted by price like every other row, with no pinned position. Other providers' prices are read from their own public pricing pages. We have no paid placement.