What will your GPU workload cost on each GPU cloud, and when is it cheaper to buy the card?

Pick a GPU, enter the busy GPU-hours per month and how much of the rented time the GPU is really busy. Get RunPod, Lambda, Vast.ai, CoreWeave, AWS, Google Cloud and Modal priced side by side, the 12-month spend, and the hours a month above which owning the card wins — before you leave an H100 idling on a bill.

Nothing is uploaded. Calculated in this browser from numbers you enter and approximate list prices you can edit. It cannot see your cloud account or your invoice.

1. Your workload

GPU

Number of GPUs

Busy GPU-hours per month (per GPU)

Utilisation of rented time (%) — instances bill while up

Persistent storage (GB)

Growth in busy hours per month (%)

Monthly budget in USD (optional)

Budget applies to

2. Comparison

Provider	USD per GPU-hour	Per month	Per year	12-month spend with growth

Rent or buy?

Owning = card plus host overhead, depreciated after resale value over your horizon, plus electricity at busy hours. It ignores your time, cooling, noise, failures and idle power.

Rate cards (editable)

Provider	USD per GPU-hour	Storage per GB-month

Owning the card

Card price (USD)

Electricity (USD per kWh)

Horizon (months)

Resale value at the end (%)

Host overhead (% of card price)

Press Compare after changing a value. A blank rate means the provider does not offer that GPU. Interruption overhead and per-second billing stay at their defaults here; use the CLI --prices file to edit them.

Prefer a CLI, or want a budget gate in CI?

node cli.mjs compare --gpu h100 --hours 200 --utilization 60 --growth 5
node cli.mjs compare --gpu a100 --gpus 4 --hours 300 --provider aws --budget 2000   # exit 1 over budget
node cli.mjs compare --gpu rtx4090 --hours 120 --prices prices.json --json
node cli.mjs providers

Get cli.mjs and engine.mjs — Node 18+, nothing to install, no network call. More: should I rent or buy a GPU?

Common questions

Does this read my cloud account?

No. It uses only the GPU, GPU-hours, utilisation and storage you type and rate cards you can edit. It cannot see your instances, quotas or invoice. Treat every figure as a way to compare options, never as what you will be charged.

Where do the prices come from?

They are approximate on-demand list prices per GPU-hour as best known in 2025, shown with an as-of line, and every number is editable. GPU rental prices move quickly and vary by region and availability, so check each pricing page before deciding.

What is not included?

Spot or reserved discounts, committed-use contracts, egress, multi-GPU instance shapes that force you to rent more GPUs than you use, regions, support plans and queue or cold-start time. Interruptible capacity is modelled only as a percentage of re-run hours. Use this to compare the hourly lines, not to forecast the whole invoice.

When does buying a GPU beat renting one?

When your busy hours a month pass the breakeven shown: depreciation of the card and host over your horizon, divided by what each busy hour saves against the cheapest rental after electricity. Below it, renting is cheaper; above it, owning is. It ignores your time, noise, cooling, failures and resale uncertainty.

Is my data uploaded anywhere?

No. The page calculates in your browser and sends nothing. The CLI reads local files and makes no network call.