GPU cost calculator: monthly vs hourly GPU rental

This GPU cost calculator compares the two ways teams usually pay for GPUs: an hourly rate multiplied by the hours you run, or a flat monthly price for a server that is yours all month. Which one costs less depends mostly on utilization, so this page explains the math, the break-even point and the costs an hourly rate leaves out.

How the GPU cost calculator works

On hourly billing, a month of GPU time costs:

hourly cost per month = hourly rate × GPUs × hours per day × days

On monthly billing, the cost is the plan price, whatever the hours. To compare both on the same basis, turn the monthly price into an effective rate per hour of real work:

effective rate per GPU-hour = monthly price ÷ GPUs ÷ hours used per month

A month averages 730 hours (8,760 hours in a year ÷ 12). That is the figure behind the per-GPU-hour equivalents on our pricing page: our single H100 at $2,190 per month comes to about $3.00 per GPU-hour when it runs around the clock.

The break-even point between monthly and hourly

break-even hours per month = monthly price ÷ hourly rate

Below that many hours, hourly billing costs less; above it, the monthly plan does. Divide by 730 to read it as a share of the month. Compare like with like: the same GPU model, the same GPU count and similar CPU, RAM and storage.

Seen from the monthly side, the fewer hours you use, the more each used hour costs. For our single H100 plan at $2,190 per month:

If an hourly offer for the same GPU costs less than the right-hand figure at your real utilization, hourly is the better deal for that workload, before the extra costs below.

Utilization in practice

• Always-on inference. A model served around the clock runs close to 730 hours a month. This is where a flat price is easiest to justify, and a fixed cost also makes budgeting simple.
• Bursty training. Runs last days, then stop for evaluation, data work or code changes. A server busy two weeks out of four sits near 50%.
• Development and experiments. A GPU used 10 hours a day on 22 working days totals about 220 hours, roughly 30% of the month.

Use measured hours rather than planned ones. A few weeks of scheduler logs or GPU metrics show how long jobs really run and how long instances stay up between them; our GPU monitoring guide explains how to collect them.

Costs an hourly rate does not show

The hourly rate is rarely the whole invoice. Depending on the provider, also count:

• Data egress: per-gigabyte charges on outbound traffic, which add up when you serve users, ship checkpoints or move datasets.
• Storage while stopped: volumes and snapshots are often billed even when the GPU instance is off.
• Setup fees: some dedicated offers charge a one-time setup or installation fee.
• Idle time between jobs: an instance left running while you prepare data or debug is billed at the full rate.
• Re-provisioning after preemption: when spot or interruptible capacity is reclaimed, you lose the work since the last checkpoint, plus the time to find capacity again, rebuild the environment and reload data.

On our servers, these items are known in advance: no setup fee, local NVMe storage included in the plan, and bandwidth included with no egress fees up to 20 TB per month on single-GPU servers and 50 TB on 8-GPU servers, under fair use, with inbound traffic not counted. See bandwidth and egress for the details.

When hourly billing is the better choice

A monthly server is not always the right answer. Hourly or per-second billing usually wins when:

• your usage stays below the break-even hours, for example a few short experiments a month;
• you need many GPUs for a few days, then none;
• traffic is spiky enough that scale-to-zero serverless inference beats an always-on server;
• your jobs tolerate interruptions and can run on spot capacity;
• you want to try several GPU types before settling on one.

We rent by the month only: no per-second billing, no serverless endpoints, no spot capacity. A common setup combines the two: a monthly server for the steady baseline and hourly capacity for peaks.

How to use your result

1. Enter measured hours and the number of GPUs you actually keep busy.
2. Add the extra costs from your current invoices to the hourly side.
3. Check that your model fits in memory with the VRAM calculator.
4. If the monthly plan comes out ahead, deploy it: choose a configuration and pay the first month from a prepaid balance that you top up in crypto. The server is delivered in under 10 minutes after payment.

The minimum term is one month, and you can cancel any month with no fee from your dashboard. For a fuller discussion of both billing models, read monthly vs hourly GPU pricing.