GPU Cloud Rental Cost Calculator

GPU cloud rental cost calculator

✓ Last verified: 2026-07-15· Source: official provider pricing page· Auto-monitored — report change →

Training or self-hosting a model? Cloud GPUs bill by the hour, on or idle. Pick a card — H100, A100, L40S, L4, RTX 4090 and more — set your hours and count to see the monthly bill, the spot vs on-demand saving, and what always-on (730h) would cost. Compare every GPU side by side.

A rented GPU bills every hour it's switched on

The thing that surprises teams moving from APIs to their own GPUs is that the meter never pauses. A per-token API costs nothing when traffic is quiet; a rented H100 costs the same whether it's saturated or sitting at 2% utilization, because you're paying for the hour, not the work. That's why the single biggest lever on your bill isn't the GPU you pick — it's how many hours you keep it running. An H100 you only need eight hours a day costs a third of one left on around the clock. This calculator makes that explicit: set the real hours, not the theoretical maximum, and watch the monthly number move. The quick-set dropdown covers the common patterns — always-on inference, a working-day training box, weekend batch runs.

Tier is the second lever. On-demand capacity is guaranteed for as long as you hold it and priced at a premium; spot, community and interruptible instances are spare capacity sold 20–50% cheaper, with the catch that the provider can take them back. For checkpointed training or fault-tolerant batch jobs, spot is close to free money; for an inference endpoint that can't blink, the on-demand premium buys you stability. The table shows both rates for every card so the trade-off is in front of you. Once you have the rental number, the real question is whether owning GPU-hours actually beats paying per token — work that out with the self-hosted LLM vs API calculator, then price the bandwidth your model serves with the cloud egress calculator and the inference endpoints themselves with the serverless function cost calculator.

How to use it

1. Pick your GPU — its typical on-demand and spot rates load automatically.
2. Choose on-demand or spot, and set how many GPUs you're running.
3. Set the hours per month, or use the quick-set dropdown for a common duty cycle.
4. Read the monthly cost and spot saving, and compare every GPU in the table.

Common mistakes

Assuming always-on. 730 hours is the maximum, not your usage — most workloads need far fewer. Forgetting idle time bills. A GPU left running between jobs still costs full rate; shut it down or use autoscaling. Picking the biggest card. An L4 or 4090 often serves a small model at a fraction of an H100's cost — match the GPU to the model. Ignoring storage and egress. Volumes, snapshots and data transfer are separate line items; add storage above and price egress separately.

FAQ

Which GPU do I need for my model?

Roughly: 7–13B models fit on an L4, RTX 4090 or A100 40GB; 70B-class models need an A100/H100 80GB or several cards. Bigger isn't cheaper if a smaller card handles the load — match VRAM to the model and batch size.

How many hours is "always-on"?

About 730 — that's 24 × 365 ÷ 12. Anything less than continuous use should set the real hours, since cost scales linearly with them.

Why do prices differ so much between providers?

Specialist clouds (RunPod, Vast.ai, Lambda) and community/spot markets undercut the hyperscalers significantly for the same silicon. The rates here are a mid-market blend — your provider's live price is the one that counts.

Does this include the cost of the model API itself?

No — this is raw GPU rental. To compare renting-and-self-hosting against a per-token API, use the self-hosted vs API calculator.

Estimate only. GPU rates move constantly and vary by provider and region — verify live pricing before relying on them.

Frequently asked questions

How much does it cost to rent an H100 in the cloud?

As of mid-2026 a single NVIDIA H100 80GB rents for roughly $2 to $3.50 an hour on-demand, depending on the provider, and noticeably less — often $1.80 to $2.50 — on spot or community tiers. Running one flat-out around the clock is about 730 hours a month, so on-demand works out near $1,500 to $2,500 monthly per GPU. This calculator lets you set the exact GPU, hours and count to see your own number, and shows the spot savings beside the on-demand price.

What's the difference between on-demand and spot GPU pricing?

On-demand (or reserved) GPUs are guaranteed for as long as you keep them running — you pay a premium for that stability. Spot, community or interruptible instances are spare capacity sold at a discount, typically 20–50% cheaper, but the provider can reclaim them with little warning. Spot is ideal for fault-tolerant batch jobs and checkpointed training; on-demand is safer for inference endpoints and anything that can't be interrupted. The calculator shows both so you can weigh the saving against the risk.

Is renting a GPU cheaper than using an inference API?

It depends almost entirely on utilization. A rented GPU is billed every hour it's switched on, busy or not, so it only beats a per-token API once you keep it densely loaded — usually well above 30–50% utilization. For spiky or low traffic, a per-token API is cheaper because you pay nothing when idle. For steady high throughput, owning the GPU-hours wins. Use this calculator for the rental side, then compare against per-token pricing with the self-hosted vs API calculator.