GPU Rental Market.

EC2 accelerated instances (P and G families) at public on-demand list price for us-east-1, read from the same pricing feed that backs the AWS calculator. The most expensive way to rent an H100 by the hour, and the most available.

ND and NC series VMs at pay-as-you-go Linux list price for East US, from Azure's public Retail Prices API. Windows, spot and low-priority meters are filtered out so the rate is comparable to everyone else's on-demand price.

Container platform renting single GPUs by the second across Secure Cloud and Community Cloud. Its public GraphQL endpoint returns the lowest live on-demand and interruptible rate per GPU model.

Peer-to-peer market where independent hosts set their own price, so it is almost always the cheapest quote on the board — and the most variable in reliability, location and uptime. Only offers listed as rentable are quoted.

One of the original GPU clouds. Publishes flat per-GPU pricing on 8-way H100 and B200 nodes with no commitment, and is usually the reference rate the rest of the neocloud market prices against.

Energy-first GPU cloud running on stranded and flared power. Competitive on 8-way H100/H200 nodes with InfiniBand.

European AI cloud with large NVLink-connected H100/H200 and B200 fleets, sold both on demand and on reserved contracts.

Non-profit-backed operator of a very large H100 fleet, consistently among the cheapest on-demand 8-way H100 rates published anywhere.

NexGen Cloud's GPU-as-a-service platform, with broad coverage from A4000 through H100 and RTX PRO 6000.

North American AI cloud running H100, A100 and Intel Gaudi capacity on demand.

Bare-metal and virtualized GPU rentals with unusually deep A6000, A5000 and L40S inventory.

Bare-metal cloud with GPU nodes across the Americas, Europe and Asia.

DigitalOcean's GPU arm. Strong on single-GPU workstation-class rentals (A4000 through A100) billed by the second.

GPU Droplets — H100 and MI300X nodes sold on the same simple hourly model as the rest of DigitalOcean.

French cloud provider with EU-resident H100, L40S and L4 instances.

Global edge cloud renting fractional and whole GPUs across a wide region footprint.

How much does it cost to rent an H100 per hour?

It depends entirely on who you rent from, and the spread is enormous — on this board the dearest provider is routinely five to eight times the cheapest for the same GPU. Peer-to-peer hosts on Vast.ai sit at the bottom, dedicated GPU clouds like Voltage Park, Lambda and Crusoe in the middle, and AWS and Azure list prices at the top. The board shows the current low per GPU-hour and every provider quoting it.

Are these GPU rental prices real-time?

Yes. Every provider here is fetched live server-side on a rolling sweep — AWS's public pricing feed, Azure's Retail Prices API, RunPod's GraphQL endpoint, Vast.ai's marketplace listings, and Shadeform's catalog for a dozen-plus dedicated GPU clouds. Nothing on this board is estimated, modeled or hand-entered.

What does $/GPU-hour mean, and why not the node price?

Providers sell very different shapes — a single GPU on RunPod, an 8-way NVLink node on Lambda, a 96-vCPU instance on AWS. Dividing the node's hourly price by the number of GPUs attached gives one number that compares across all of them. The node price is shown too, since that is what you are actually billed and the only honest way to price a cluster.

How do I price an 8-GPU cluster node?

Each GPU's page lists every multi-GPU configuration on offer with its full node rate, so you can see the cheapest 8x H100 or 8x B200 box per hour and which provider has capacity for it. Interconnect matters as much as price for training: an NVLink or InfiniBand-connected node and eight loose PCIe cards are not the same product, so both are labelled.

Is it cheaper to rent or buy a GPU?

Every GPU tracked on both sides of this site shows a breakeven — the number of hours of rental that cost the same as buying the card outright at its current street price. Below that many hours, renting wins outright. Above it, buying starts to pay back, though the breakeven ignores power, cooling, networking and the cost of capital, so it is a floor rather than a full comparison.

Why is Vast.ai so much cheaper than AWS?

They are not the same product. Vast.ai is a peer-to-peer market where independent hosts rent out their own machines and set their own prices, with wide variation in reliability, location, bandwidth and how long a machine stays up. AWS sells guaranteed capacity with an SLA, enterprise networking and a support contract. The board quotes both honestly and labels which is which.