Rex says
Money math without the spreadsheet headache. Plug in your numbers and I'll show you exactly where the dollars land.
Try a scenario
Click to load — tweak from there.Inputs
Result
Total training cost
$1,440.00
Total GPU-hours consumed
576.0
Cost per epoch
$144.00
Effective hourly rate per GPU
$2.500
Psst — share this and help Rex grow
One click, a permanent link with your numbers baked in.
How to use this
- 1Enter number of gpus.
- 2Enter hourly rate per gpu ($).
- 3Enter total training wall-clock hours.
- 4Enter spot/preemptible discount (%).
- 5Enter training epochs (optional).
- 6Read your total training cost on the right — it updates as you type.
- 7Hit Share to keep the scenario or send it to someone.
About this calculator
GPU training cost is driven by three multiplied factors: how many GPU-hours the job needs (GPU count × wall-clock hours), the hourly rate for that GPU type, and whether you're paying on-demand or spot/preemptible pricing. This calculator computes total training cost from those inputs and also reports cost per training epoch if you provide epoch count, which is the number teams actually use to decide whether to keep iterating on hyperparameters or cut losses on a run that isn't converging. Spot/preemptible GPU instances typically cost 50-90% less than on-demand but can be reclaimed by the cloud provider mid-job, so they only make sense with checkpointing that lets a job resume rather than restart — factor that operational requirement in before assuming the spot discount is free money, since a job that gets preempted repeatedly without good checkpointing can end up costing more in wasted compute than it saved on rate.
Worked example
Using the values the calculator loads with:
Inputs
- Number of GPUs: 8
- Hourly rate per GPU: 2.5 $
- Total training wall-clock hours: 72
- Spot/preemptible discount: 0 %
- Training epochs (optional): 10
Results
- Total training cost: $1,440.00
- Total GPU-hours consumed: 576
- Cost per epoch: $144.00
- Effective hourly rate per GPU: $2.50
What each field means
Inputs
- Number of GPUs
- The number of gpus used in the calculation. Starts at 8 so you have a working example on load.
- Hourly rate per GPU ($)
- The hourly rate per gpu used in the calculation, measured in $. Starts at 2.5 $ so you have a working example on load.
- Total training wall-clock hours
- The total training wall-clock hours used in the calculation. Starts at 72 so you have a working example on load.
- Spot/preemptible discount (%)
- The spot/preemptible discount used in the calculation, measured in %. Starts at 0 % so you have a working example on load. Accepted range: 0–90 %.
- Training epochs (optional)
- The training epochs (optional) used in the calculation. Starts at 10 so you have a working example on load.
Results
- Total training cost
- Returned as a money amount in US dollars and shown as the headline result. It recalculates instantly whenever you change an input, so you can compare scenarios without reloading.
- Total GPU-hours consumed
- Returned as a whole number. It recalculates instantly whenever you change an input, so you can compare scenarios without reloading.
- Cost per epoch
- Returned as a money amount in US dollars. It recalculates instantly whenever you change an input, so you can compare scenarios without reloading.
- Effective hourly rate per GPU
- Returned as a money amount in US dollars. It recalculates instantly whenever you change an input, so you can compare scenarios without reloading.
FAQ
Is renting GPUs from a specialty cloud cheaper than AWS/GCP/Azure?
Often yes, sometimes substantially — specialty GPU cloud providers frequently price high-end accelerators 30-60% below the big three hyperscalers because they don't bundle the same breadth of managed services and enterprise support. For a pure training job with no other cloud service dependencies, it's worth comparing rates directly since the discount can be the difference between a run being affordable or not.
How risky is using spot instances for a multi-day training run?
Reclamation rates vary by GPU type, region, and demand, but multi-day jobs on popular GPU types can see preemption multiple times over their run. This is only safe with frequent checkpointing (every 15-30 minutes or every N steps) so a preempted job resumes near where it left off rather than restarting from scratch — without that, spot savings can be wiped out by wasted repeated computation.
Why does GPU count matter separately from GPU-hours?
Multi-GPU training introduces communication overhead (gradient synchronization across GPUs/nodes) that means going from 1 to 8 GPUs rarely gives a full 8x speedup — realistic scaling efficiency is often 70-90% depending on interconnect (NVLink vs. standard networking) and model architecture. Model your actual measured scaling efficiency, not a naive linear assumption, or you'll underestimate total training hours needed.
Accuracy and limitations
- Results are estimates before tax, fees, and inflation unless an input explicitly covers them.
- Rates are treated as fixed for the whole period — variable-rate products will drift from this projection.
- This is educational maths, not financial advice. Check anything contractual with the lender or your accountant.
Related tools
LLM API Token Cost Calculator
Estimate monthly spend on GPT/Claude/Gemini-style token-based API pricing.
Vector Database Storage Cost Calculator
Estimate storage and index memory cost for embedding-based search.
Cloud VM Monthly Cost Calculator
Estimate the true monthly bill for a running virtual machine.
Cite this calculator
Writing about this topic? Grab a citation — every link helps keep these tools free.
RevenueLab. (2026). GPU Training Cost Calculator. Retrieved from https://www.revenuelab.fyi/toolbox/gpu-training-cost
<p>Source: <a href="https://www.revenuelab.fyi/toolbox/gpu-training-cost" target="_blank" rel="noopener">GPU Training Cost Calculator — RevenueLab</a> (2026).</p>
Source: [GPU Training Cost Calculator — RevenueLab](https://www.revenuelab.fyi/toolbox/gpu-training-cost) (2026).