---
title: "What is a GPU-hour? GPU time vs per-token LLM pricing"
description: "A GPU-hour is one GPU used for one hour, the unit of cloud GPU rental and of compute budgets. How it is billed, and how it differs from an hourly LLM API key."
url: https://broadrouter.com/glossary/gpu-hour/
lang: en
updated: 2026-11-16
publisher: BroadRouter (a BroadNet site); API by BroadNet Inference, https://inference.broadnet.ai
---

# GPU-hour

Updated 2026-11-16

A GPU-hour is one graphics processor used for one hour. Cloud providers bill rented GPU instances by the hour or second, and researchers use GPU-hours to state how much compute a job took. You pay for the machine's time, whether or not it is busy.

Also called: GPU hours, GPU-hours, GPU instance hour, GPU rental hour

A GPU-hour is the unit you pay in when you rent hardware. A per-token API hides it; a self-hosted model makes it the whole bill.

## How it works

Cloud providers price GPU instances per hour, then meter more finely. AWS’s on-demand page says pricing is “per instance-hour consumed”, with partial hours on Linux billed per second and a 60-second minimum. Google Cloud charges “all vCPUs, GPUs, and GB of memory” a 1-minute minimum, then 1-second increments. An instance with 8 GPUs running for 3 hours is 24 GPU-hours.

The meter runs while the instance is up, not while it is useful. An idle GPU at 3 a.m. costs the same as a busy one.

Researchers use the same unit for compute. The GPTQ paper, for example, quantizes a 175B-parameter model “in approximately four GPU hours”.

## Why it matters

To compare a GPU-hour with per-token pricing, convert through [throughput](https://broadrouter.com/glossary/throughput/):

```text
cost per 1M output tokens = hourly GPU cost ÷ (tokens per second × 3,600 ÷ 1,000,000)
```

The catch is the denominator. Throughput depends on the model, precision, batch size and how busy you keep the GPU. A rented GPU at 20% load costs five times as much per token as the same GPU at full load. That is the main reason teams with uneven traffic use APIs; see [self-hosting vs API cost](https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/).

## On BroadNet Inference

BroadNet does not rent GPUs by the hour. It sells API access to its models; its Terms say they run on BroadNet-owned or leased hardware in the United States, or with US inference partners. Two of its options are priced per hour, and neither is a GPU-hour:

|  | GPU-hour (cloud rental) | BroadNet hourly key | Meteor plan |
| --- | --- | --- | --- |
| What you get | A whole GPU and the machine around it | One stream on a shared model | One session on a shared model |
| Model | Anything that fits | BroadNet Gemma 4 VL 26B only | DeepSeek V4.1 Flash only |
| Limits (per BroadNet’s docs) | The hardware | 120 requests/min, 300,000 tokens/min, 1 stream | One request at a time, no token meter |
| Price | Set by the cloud | $0.50 per hour | $1 per hour per session |

You do not manage drivers, scaling or idle time with either BroadNet option, but you also cannot choose the hardware or load a different model. Running the hourly key all month would be $0.50 × 24 × 30 = $360.00. For details, see [hourly LLM API keys explained](https://broadrouter.com/blog/hourly-llm-api-key-explained/) and the [flat-rate plan](https://broadrouter.com/glossary/flat-rate-llm-plan/) entry.

## Questions

### Is a BroadNet hourly key a GPU-hour?

No. An hourly key buys one hour of one request stream on a shared model, BroadNet Gemma 4 VL 26B, with fixed limits. You do not get a GPU, a choice of model or any control over the server.

### Are GPU-hours billed for the full hour?

Often not. AWS bills many Linux instances per second after a 60-second minimum, and Google Cloud charges GPUs a 1-minute minimum, then per second. Check the instance type and operating system.

### How do I turn a GPU-hour price into a per-token price?

Divide the hourly cost by the millions of tokens the GPU produces in an hour, which is tokens per second × 3,600 ÷ 1,000,000. Measure throughput at your real load, because idle time is billed too.

## See also

-   [Pay-as-you-go](https://broadrouter.com/glossary/pay-as-you-go/) (term)
-   [Flat-rate LLM plan](https://broadrouter.com/glossary/flat-rate-llm-plan/) (term)
-   [Inference provider](https://broadrouter.com/glossary/inference-provider/) (term)
-   [Serverless inference](https://broadrouter.com/glossary/serverless-inference/) (term)
-   [Throughput](https://broadrouter.com/glossary/throughput/) (term)
-   [How the unlimited LLM API plans work for agents](https://broadrouter.com/guides/unlimited-llm-api-sessions/) (guide)
-   [Self-hosting an LLM vs an API: the real cost and break-even](https://broadrouter.com/blog/self-hosting-llm-vs-api-cost/) (article)
-   [Hourly LLM API key: a flat price per hour, explained](https://broadrouter.com/blog/hourly-llm-api-key-explained/) (article)

Try it on your real workload. New accounts start with $5 of credit for Gemma and transcription. DeepSeek, GLM and Qwen3.8 need a top-up.

[Get started](https://inference.broadnet.ai/account/?utm_source=broadrouter&utm_medium=referral&utm_campaign=glossary-gpu-hour&utm_content=page-cta) ·

**Sources**

-   [Amazon EC2 On-Demand pricing (per-hour and per-second billing)](https://aws.amazon.com/ec2/pricing/on-demand/)
-   [Google Cloud Compute Engine pricing (GPU billing increments)](https://cloud.google.com/products/compute/pricing)
-   [Frantar et al., GPTQ (quantizing a 175B model in about four GPU hours), arXiv 2210.17323](https://arxiv.org/abs/2210.17323)
-   [BroadNet Inference API reference, hourly keys and limits](https://inference.broadnet.ai/docs?utm_source=broadrouter&utm_medium=referral&utm_campaign=glossary-gpu-hour&utm_content=sources#limits)
