Cost guide · 6 min read

# When monthly rental beats per-hour.

Per-hour pricing is the right answer for a job that finishes today. It becomes the wrong answer sooner than most people expect — and the break-even is arithmetic, not opinion.

The short answer

- **Our monthly price:** $1,469 for a NVIDIA H100 PCIe — $2.01 an hour if you divide it out
- **The break-even:** Against a $2.99/hour cloud, monthly wins past **492 hours** — about 20 days
- **The real question:** Not "how many hours will I compute" but "how many hours will the machine exist"
- **Where hourly wins:** A job that finishes today, or a load that is genuinely bursty and can tolerate eviction

## The break-even

One NVIDIA H100 PCIe, at our monthly price of $1,469, against three ways of renting the same card by the hour. The break-even is the point where the meter has cost you what a whole month costs us.

**Hours until a monthly term is cheaper than hourly billing**

| Hourly option | Per hour | A full month | Break-even | You save at 24/7 |
|---|---|---|---|---|
| Hyperscaler, on demand Committed-use discounts need a year and a contract | $6.98 | $5,095730 hours | 211 h 9 days | $3,626 every month |
| Specialist GPU cloud Per-second billing, no commitment | $2.99 | $2,183730 hours | 492 h 21 days | $714 every month |
| Marketplace, community tier Someone else's machine, interruptible | $1.89 | $1,380730 hours | 778 h 32 days | Hourly is cheaper |
| Our monthly term Dedicated, no meter, no eviction | $2.01derived, not billed | $1,469 | — | — |

Hourly figures are public list prices for the same card, observed in August 2026. We do not bill by the hour; the per-hour figure on our own row is the monthly price divided by 730, shown only so the columns can be compared.

## The arithmetic

Three lines, and worth writing out because the second one is where most estimates go wrong.

Break-even, in hours

```
break_even_hours = monthly_price / hourly_price

# For this card, against a $2.99/hour cloud:
$1,469 / $2.99 = 492 hours = 20.5 days

# The mistake: counting only the hours you compute.
# The meter runs while the machine exists, not while it works.
hours_billed = hours_the_instance_is_running   # not hours of useful work
```

A month is 730 hours. If your machine needs to exist for more than about 20 days out of thirty, a monthly term is cheaper — and every hour past that point is pure saving.

## What an hourly price hides

The published rate is the part you can compare. These three are the part you cannot, and together they usually matter more than the rate itself.

### Storage that bills separately

On most hourly platforms the disk is a separate line item, charged per GB per month, and it keeps charging while the instance is stopped. A 2 TB dataset can quietly cost more than the GPU on a lightly used month.

**Here:** 2 × 2 TB NVMe is in the price, and nothing is metered.

### Egress

Getting your model weights, checkpoints or generated output *out* is charged per gigabyte almost everywhere. It is the cost that only appears once the work succeeds.

**Here:** 1 Gbit/s, unmetered, both directions, no overage at any volume.

### Cold start and eviction

Pulling a 140 GB model onto a fresh instance is fifteen to forty minutes you pay for and get nothing from. On interruptible tiers you pay it again every time you are evicted mid-run.

**Here:** the machine is yours for the term. Weights are on its disks and stay there.

## The idle-time trap

This is the one that turns a careful estimate into a bill three times larger. Per-second billing feels safe because you can stop the instance — and almost nobody does, because stopping it means losing the state on it.

**Monthly cost at different levels of actual utilisation, hourly versus monthly**

| The machine runs | Hours | At $2.99/h | Our monthly | Cheaper |
|---|---|---|---|---|
| A few hours a week | 12 h | $36 | $1,469 | Hourly, by $1,433 |
| Working hours, weekdays | 176 h | $526 | $1,469 | Hourly, by $943 |
| Every day, working hours | 240 h | $718 | $1,469 | Hourly, by $751 |
| Continuously | 730 h | $2,183 | $1,469 | Monthly, by $714 |

Read the second row carefully. "Working hours, weekdays" sounds modest and is already the wrong side of the line — and it assumes you diligently stop the instance every evening, which requires that nothing you care about lives on it.

**Ask the honest question.** Not "how many hours will I use the GPU" but "how many hours will this machine need to exist". An endpoint that answers requests, a model you are iterating on, a dataset you do not want to re-download — all of those need the machine to exist continuously, whatever the utilisation graph says.

## Longer terms

If the answer is "for a year", the monthly price is not the best one available either. The same card on our three terms:

**The same configuration on each term**

| Term | Per month | Paid up front | Saved over a year | Trade |
|---|---|---|---|---|
| Monthly monthly billing, cancel any time | $1,469 | $1,469 | — | None. Cancel before the renewal date and it simply stops. |
| 3 months 3-month term, 5% off every month | $1,396 | $4,188 | $876 | Paid up front, not refunded pro rata. That is what the discount buys. |
| 12 months 12-month term, 12% off every month | $1,293 | $15,516 | $2,112 | Paid up front, not refunded pro rata. That is what the discount buys. |

## When hourly really is right

We do not sell hourly, so take this as what it is: the cases where you should go elsewhere.

**A job that finishes today.** One fine-tune, one benchmark, one render. Rent it by the hour, pay for four hours, be done.

**Genuinely bursty load that tolerates eviction.** If your work checkpoints cleanly and can be killed mid-run, interruptible capacity is dramatically cheaper than anything dedicated.

**You need eight cards for six hours, once.** A monthly term on an eight-GPU node to run one experiment is the wrong shape of purchase.

**You do not know yet.** Rent by the hour until the utilisation graph tells you what you actually need, then come back with a number.

Everywhere else — an endpoint that has to answer, a model you are living with for a quarter, a card you want to leave loaded — a month costs less than the meter and behaves better. And unlike a committed-use discount, a monthly term here needs no contract, no company, and no identity check: it is [one payment](https://gpuserver.io/guides/pay-in-crypto) and the machine is yours.

## Every price in the catalogue is a monthly price.

No meter, no egress charge, no storage line item, and no setup fee to add to it.

[Browse the catalogue](https://gpuserver.io/#catalog) [Read the guides](https://gpuserver.io/guides)

## Other guides

- [Cost · 9 min Self-hosting against a per-token API The break-even between paying an API per token and renting a GPU, worked out on our own prices and throughput — and the four things the arithmetic leaves out. Read the guide](https://gpuserver.io/guides/api-vs-self-hosting)
- [Practice · 11 min Serving Llama 3.3 70B on one node From a delivered machine to an OpenAI-compatible endpoint, with the flags that matter and the two that silently halve your throughput. Read the guide](https://gpuserver.io/guides/serve-llama-70b)
- [Practice · 10 min Fine-tuning a 70B model on one card QLoRA on a single 48 GB GPU: what fits, what it costs for a month, and why full fine-tuning is a different order of machine. Read the guide](https://gpuserver.io/guides/lora-finetune)

---

Source: https://gpuserver.io/guides/monthly-vs-hourly/. This file is generated from the same data as the website; if a figure here differs from a page, the page is authoritative and this file is stale — the canonical source is https://gpuserver.io/.