Fine-tuning, without paying for certainty you do not need.
Training is the workload spot capacity was made for: it runs for hours, it can checkpoint, and it does not care if it pauses for four minutes. The trade is 30–53% off the hourly rate.
- LoRA on 8B
- $0.06/h1× A40 on spot
- Spot discount
- 30–53%Against our own on-demand price
- Notice
- 2minTime to flush a final checkpoint
- Node size
- up to 8×NVLink between the GPUs of a node
What a training run actually needs.
The rule we apply below, stated so you can check it: LoRA ≈ weights × 1.25 (adapters, gradients, activations). Full fine-tune ≈ weights × 4 (parameters, gradients and two optimiser states), before activations. Both at FP16 / BF16.
| Model | QLoRA, 4-bitnode · spot price | LoRA, FP16node · spot price | Full fine-tunenode · spot price |
|---|---|---|---|
| Llama 3.1 8B8B parameters | 1× RTX A4000 $0.03/h | 1× A40 $0.06/h | 1× A100 PCIe $0.09/h |
| Gemma 2 27B27B parameters | 1× A30 $0.05/h | 2× A40 $0.12/h | 4× A100 PCIe $0.36/h |
| Qwen2.5 32B32B parameters | 1× A40 $0.06/h | 2× A100 PCIe $0.18/h | 4× A100 PCIe $0.36/h |
| Llama 3.3 70B70B parameters | 1× A100 PCIe $0.09/h | 4× A100 PCIe $0.36/h | 8× RTX PRO 6000 $1.52/h |
They depend on the batch size, the sequence length and whether you use gradient checkpointing. Take the node above as the floor, and give yourself one step of headroom for a first run — the hourly difference is usually smaller than the cost of a job that dies at 90%.
A run that resumes is a run you should not pay full price for.
Three habits turn an interruption into a four-minute pause: checkpoint to the volume, watch the notice, resume from the newest checkpoint. Every serious training stack already does the first and the third.
- Checkpoints go to
/mnt/volThe persistent volume survives the reclaim. Local scratch does not — it is the single most expensive mistake on this platform. - Poll the notice
/v1/noticeflips 2 minutes before the stop; flush a final checkpoint and exit cleanly. - Turn on auto-relaunchThe platform queues the same disk and image for the next free GPU of that model. You pay nothing while it waits.
What the tier is worth
H100 · 1 GPU- A 12-hour run on spot$6.60
- The same run on-demand$10.68
- Saved$4.08
- Checkpoint storage, 250 GB for a month$20.00
- Spot run + storage$26.60
At a <5% reclaim rate, a 12-hour run on this model is interrupted rarely — and when it is, you lose the minutes since the last checkpoint, not the run.
From an empty account to a training loop.
-
Create a volume for the run
250 GB costs $20.00 a month and holds a 7B fine-tune with several checkpoints. It is what makes the job resumable.
/mnt/vol/run-42 -
Launch the node on spot
Pick the model from the table above, the PyTorch template, the region closest to your data, and turn on auto-relaunch.
-
Point the trainer at the volume
Hugging Face
Trainer, Lightning and most training scripts resume from the newest checkpoint in their output directory.--output_dir /mnt/vol/run-42 --save_steps 200 -
Watch the run, not the machine
The console shows the reclaim notice and the relaunch. Your job shows its loss curve. Nothing else needs your attention.
Training questions
How much memory does fine-tuning need?
Far more than inference. A LoRA or QLoRA run needs roughly the weights plus a quarter for adapters, gradients and activations. A full fine-tune with an Adam optimiser needs about four times the weights — the parameters, the gradients and two optimiser states — before activations. The tables above apply those rules to the catalogue.
Is spot capacity safe for a training run?
Yes, if the run checkpoints. You get a 2-minute notice, the instance is stopped rather than deleted, and the persistent disk keeps every checkpoint. With auto-relaunch the job resumes on the next free GPU. Without checkpoints, a reclaim costs you everything since the start — which is true of any interruption, including a crash.
How often should I checkpoint?
Often enough that losing the interval does not hurt: every 5 to 15 minutes for most runs. Writing a checkpoint to the persistent volume takes seconds and costs nothing extra. See the handling guide for the pattern.
Which template should I use?
The PyTorch template for most stacks, or the CUDA base if you install your own. If your team already ships a training image, bring it — any OCI image runs, and the requirements are short.
Do you support multi-node training?
A single node holds up to 8 GPUs, which covers most fine-tuning. Anything wider is arranged as reserved capacity so the machines are guaranteed to be there together — tell us the shape and we will say what is possible.
Start the run on spot and keep the difference.
Pay as you go — no contracts, no minimum commitment. Add credit, launch, stop whenever you want.
Billed per minute from the moment the instance is reachable. Minimum credit $40, no subscription.