Model Training Cost Calculator (GPU-Hours)
Turn a training run — number of GPUs, wall-clock hours and price per GPU-hour — into total GPU-hours and dollar cost, the way cloud accelerator bills are actually rated.
How it works
GPU-hours = GPUs × wall-clock hours · Cost = GPU-hours × price per GPU-hour
Cloud accelerators are billed per GPU-hour, so a training run’s cost is simply the number of GPUs multiplied by how long they run, multiplied by the per-GPU price. Adding GPUs usually shortens wall-clock time but keeps GPU-hours (and therefore cost) roughly constant — you pay for the compute either way, with some overhead from imperfect scaling. Spot or reserved capacity can cut the per-GPU rate by 50–70%, which is why the rate field, not the GPU count, is often the biggest lever on the bill.
Worked example
A run uses 8 GPUs for 72 hours at $2.50 per GPU-hour. That is 8 × 72 = 576 GPU-hours, and 576 × $2.50 = $1,440 for the run, finishing in 3 days.
Frequently asked questions
Should I use the instance price or the per-GPU price?
Use the per-GPU price. If your quote is for a whole 8-GPU instance, divide the hourly instance price by 8 before entering it here so the GPU count in this calculator does the multiplication.
Does doubling GPUs halve the cost?
No — it roughly halves wall-clock time but leaves total GPU-hours (and cost) about the same, since you are renting twice as many GPUs for half as long. In practice large clusters lose some efficiency to communication overhead, so cost can even rise slightly.
What is not included here?
Storage, data egress, networking, checkpoint storage and any managed-service premium sit on top of raw GPU time. For a rough training budget, GPU-hours dominate, but add 10–30% for these extras.
Related calculators
Educational estimate — real costs and memory depend on your provider, framework, model architecture and pricing, which change often. Verify against your own bills and profiling before committing budget.