GPU Cloud Providers Compared Independently

Four kinds of GPU cloud

Kind	Examples	Good for	Watch out for
Neocloud (GPU-first cloud)	Lambda, Hyperstack, CoreWeave, Together AI, Jarvislabs	training and inference at lower hourly rates	capacity can sell out; some rent only whole nodes
Developer cloud	RunPod, DigitalOcean, Vultr	quick single-GPU work, templates	billing minimums, storage rules
Serverless	Modal, RunPod Serverless	bursty inference you pay for by the second	CPU and memory billed on top; cold starts
Hyperscaler	AWS, Azure, Google Cloud, Oracle	integration with an existing cloud, compliance	much higher on-demand GPU rates; 8-GPU VMs
Marketplace	Vast.ai	lowest prices from independent hosts	reliability varies by host; prices set by hosts

Provider by provider (26 October 2026)

Provider	H100 on-demand	Billing step	Egress	Notes
Hyperstack	$2.50 PCIe, $3.20 SXM	per minute	free	prepaid balance (source)
RunPod	$2.89 PCIe, $3.99 SXM	per second	not stated	idle volumes $0.20/GB-month (source)
Jarvislabs	$3.49	per minute	free in region	spot from $1.54 (source)
Modal	$3.95	per second	$0.04/GiB after 1 TiB	serverless; CPU and RAM extra (source)
Lambda	$3.99	not stated	free	(source)
Together AI	$3.99	hourly	not stated	spot $1.99 (source)
DigitalOcean	$4.41	per second, 5-minute minimum	large allowance included	billed while powered off (source)
CoreWeave	$6.16 per GPU	not stated	free	8-GPU nodes only (source)
Azure	$12.29 per GPU	per second	paid	ND H100 v5 is an 8-GPU VM (source)

Vultr lists H100 only on long contracts, and Vast.ai prices are set live by hosts, so neither has an H100 on-demand price in our table. Google Cloud, AWS and Oracle are being checked by hand.

How to choose

1. Size the job first with the VRAM calculator.
2. Price the whole job with the cost calculator.
3. Check capacity and region on the provider's page before you start.