GPU Rental Pricing

Per-GPU hourly rates are shown in the pricing table above. Entry-tier consumer GPUs like RTX 4090 sit at the low end and are available on demand; larger models such as RTX PRO 6000, H100 and H200 are reserved, so their on-demand rate is quoted on request. Rates update from live capacity.

For inference workloads that fit in 24GB of VRAM, RTX 4090 is typically the lowest hourly rate among the GPUs in the table above. For larger models, RTX 5090 and RTX PRO 6000 deliver better cost per token. CloudRift adds new capacity regularly, and older models such as NVIDIA V100 can be a more cost-effective fit for certain workflows. Contact us if your workload may suit a GPU not listed above. See our LLM inference benchmarks for throughput numbers.

Hourly rental is cheaper than buying when GPU utilization stays below roughly 40 to 50 percent of the depreciation window of the hardware. For continuous workloads above that, reserved capacity or owned hardware tends to win on total cost. CloudRift offers both hourly rates and multi-month reservations so you can match the billing model to actual utilization.

CloudRift lists transparent per-GPU per-hour pricing with no minimum spend and no egress charges. Billing is per second of active runtime, similar to RunPod and Lambda. Hyperscaler GPU instances on AWS, GCP, and Azure typically run several times higher per hour at list price and require longer reservation commitments for discounts. The cloud GPU price comparison puts the current AWS, Azure, Google Cloud, RunPod and Lambda rates next to ours, card by card. For a specific comparison against another provider, contact us with the GPU type and duration you are evaluating.

On-demand is pay-as-you-go billed by the second while the instance is running. Reserved locks a specific GPU type for a fixed duration of 1 week, 1 month, 3 months, or 1 year at a lower hourly rate, paid upfront. Reserved capacity is the right choice when utilization is high and predictable.

Yes. On-demand instances have no minimum duration and bill in per-second increments while the instance is running. You can spin up a GPU for a few minutes of testing and pay only for that runtime.

Billing starts when the VM reports ready, not when you click launch, and stops the moment you terminate, even if the node is still cleaning up. If a VM never starts, the hold on your balance is returned and nothing is charged. Paused VMs keep their GPU allocated and keep billing.

The hourly price covers compute, the attached local storage shown in the pricing table, networking, and standard support. There are no separate egress, ingress, or API-call charges on CloudRift.

For GPU rentals that aren't available in the console (such as NVIDIA H100 and H200), we may still be able to help. Reach out on Discord or via our contact form with your requirements and we'll work with you to find a solution.