Hourly × cards × hours, computed locally

GPU cloud cost calculator

Hourly rate × GPU count × hours = cloud bill. Compare marketplace and hyperscaler pricing for the same card before you rent compute for your next model.

 
Total cost $—
AD
Live comparison

Same job, every provider

Your GPU count × hours, across the board · verified October 2026

ProviderGPU$/hourThis job
RunPod RTX 4090 $0.69 $—
RunPod A100 80GB $1.89 $—
RunPod H100 $3.49 $—
Vast.ai RTX 4090 $0.44 $—
Vast.ai A100 80GB $1.65 $—
Vast.ai H100 $2.75 $—
Lambda A100 40GB $1.29 $—
Lambda H100 $2.49 $—
Lambda H200 $3.29 $—
AWS A100 40GB (p4d) $4.10 $—
AWS H100 (p5) $8.00 $—
GCP A100 80GB $5.60 $—
GCP H100 $7.40 $—
Azure H100 $6.90 $—
Official list prices per pricing pages; region, contract, and spot pricing vary. Verify before large commits.
How the math works

Cost is rate × cards × hours.

Cloud GPU billing has no hidden formula: providers quote dollars per GPU-hour, you multiply by how many cards and how long. The value of a calculator is the comparison — the same H100 ranges from about $2.50 on marketplaces to over $8 on hyperscalers, so an identical training job can cost three times more depending only on where you launch it.

Chain this with the other AI calculators for full planning: the VRAM calculator tells you how much memory the model needs, multi-GPU tells you the card count, training-time converts that into hours — and this page turns hours into dollars. One workload, priced end to end.

Good to know

Frequently asked questions

Why do the same GPUs cost 2-4× more on AWS, GCP, and Azure?

Hyperscalers bundle enterprise SLAs, VPC networking, compliance certifications, and support into the price. Marketplaces like RunPod, Vast.ai, and Lambda sell raw compute closer to spot cost. For research and personal projects the discount is real; for regulated production work the hyperscaler premium buys things you may actually need.

How accurate are these hourly rates?

Reference prices from provider pricing pages as of October 2026, re-verified quarterly. Actual quotes vary by region, availability zone, contract length, and demand — burst pricing on marketplaces can undercut list by 30-50% at quiet hours.

How do I pick between renting and buying?

A rough rule: a $2,000 consumer card costs about 3-5 months of 24/7 cloud rental on marketplaces. If your workload runs most of the day every day, buy; if it bursts for hours at a time, rent and only pay while training or serving.

More AI calculators

Keep computing