← GPU Training Pricing Calculator
Put your GPU, busy hours a month and utilisation into the calculator to price RunPod, Lambda, Vast.ai, CoreWeave, AWS, Google Cloud and Modal side by side and see the buy-versus-rent breakeven. These are the questions to settle before you trust the answer.
When your busy hours a month stay above a breakeven that is usually well under half of a month for consumer cards. Divide the card and host cost, less resale value, by the months you will keep it, then divide by what each busy hour saves against the cheapest rental after electricity. A 24 GB RTX 4090 against a roughly $0.69 an hour rental breaks even at around 70 busy hours a month with the calculator's defaults; an H100 needs far more because rental prices are lower relative to the card. Below breakeven rent; above it, own.
Instances bill while they are up, not while the GPU is busy. If you launch a pod, run for an hour, and leave it up for two, you paid for two. Enter your real utilisation: at 50% every hourly rate effectively doubles, and a per-second serverless provider with a higher rate can win. Interruptible marketplace capacity adds re-run hours when work is reclaimed.
Whether the GPU you need is offered where you can get capacity, how fast an instance starts and whether your data and checkpoints sit on persistent storage that is billed even when the GPU is off, and whether egress or minimum commitments apply. Run the calculator with your own hours, utilisation and storage, then confirm against each provider's pricing page.