How billing works
You pay for two things, and never for both at the same time: the GPU slice while your instance runs, and the disk while your instance sits stopped. Both are drawn from a prepaid credit balance that belongs to your organization, not to you personally.
The two meters
| Instance state | Metered | What you pay |
|---|---|---|
| creating | no | nothing |
| starting | no | nothing |
| running | yes | the offer's hourly price for the whole slice, charged by the second |
| stopping | no | nothing |
| stopped | yes | disk size multiplied by the storage rate per GB-hour, charged by the second |
| destroying | no | nothing |
| destroyed | no | nothing |
| error | no | nothing |
Two consequences are worth internalizing.
Provisioning is free. The GPU meter turns on when the instance reports running, not when you press deploy. Time spent creating or starting — pulling a multi-gigabyte image, booting the container — costs nothing. The same applies on the way down: stopping and destroying are unmetered.
A stopped instance is not a free instance. The storage meter turns on the moment the instance reaches stopped and runs until the instance is destroyed. You are paying to hold your disk and your slot on that specific machine.
Stopping ends the GPU charge, not the bill. If you are finished with the work, destroy the instance. See Stop vs destroy.
The offer price covers the whole slice
An offer is a slice of a machine — 1x, 2x, 4x or 8x GPUs — and the hourly price you see on the marketplace is the price of that entire slice. A 4x slice at $9.60/hour costs $9.60/hour, not four times that.
Rates are fixed at rent time
The hourly GPU price and the storage rate are copied onto your instance when you deploy. If the marketplace price for that GPU model changes afterward, a live instance keeps the rate it was rented at. The rates in force for a given instance are shown on its Billing tab as "Burn rate (running)" and "Burn rate (stopped)".
The settlement tick
A settlement pass runs about once a minute. For every instance that is running or stopped, it:
- Computes the time accrued since that instance's last settlement, at the rate for its current state.
- Appends one entry to the organization's ledger.
- Subtracts the amount from the balance.
Settlement also happens on every state change, inside the same transaction as the change itself. When you press stop, the GPU time up to that instant is settled at the running rate before the state flips — so the boundary between GPU charges and storage charges is exact, not rounded to the nearest tick.
Each settlement bills whole seconds and carries any leftover fraction of a second into the next one, and amounts are rounded once per settlement. Nothing is billed twice, and nothing is lost between ticks.
Because a tick settles usage that already happened, a balance can pass zero by up to about a minute of usage before enforcement sees it. That is by design, and it is why the second threshold exists. See Auto-stop and auto-destroy.
A worked example
The rates below are round numbers chosen to make the arithmetic legible. Substitute the ones the deploy page quotes for your own offer and disk size.
- Offer: $2.40 per hour for the slice, which is $2.40 ÷ 3600 = $0.00067 per second while running.
- Disk: 200 GB, with storage quoted at $0.0002 per GB-hour.
- Storage rate for this instance: 200 GB × $0.0002 = $0.04 per hour while stopped.
You deploy in the morning, work for an hour and a half, stop the instance overnight, come back for half an hour, and destroy it.
| Time | State | Duration | Charged |
|---|---|---|---|
| 09:00–09:02 | creating | 2 min | nothing |
| 09:02–10:32 | running | 90 min | $2.40 × 1.5 = $3.60 |
| 10:32–10:33 | stopping | 1 min | nothing |
| 10:33–22:33 | stopped | 12 h | $0.04 × 12 = $0.48 |
| 22:33–22:34 | starting | 1 min | nothing |
| 22:34–23:04 | running | 30 min | $2.40 × 0.5 = $1.20 |
| 23:04 | destroying, then destroyed | — | nothing |
Total for the day: $3.60 + $0.48 + $1.20 = $5.28.
In the ledger this appears as roughly 90 GPU-usage rows of about $0.04 each (one per minute of running time), then roughly 720 storage rows of about $0.0007 each (one per minute of stopped time), then roughly 30 more GPU rows. Sub-cent amounts are shown to four decimal places so that a per-minute charge never reads as $0.00.
Now change one thing. Leave that same instance stopped for a week instead of twelve hours:
$0.04/hour × 24 hours × 7 days = $6.72
The parked disk costs more than the 90 minutes of GPU time that produced the data on it. Large disks left stopped are the most common source of a surprising balance.
Where to see what you are spending
| Where | What it shows |
|---|---|
| Header balance chip | Current credit balance, refreshed about once a minute |
| Billing page | Balance, a 30-day spend chart, and the full transaction ledger |
| Instance → Billing tab | Settled spend for that one instance, plus both burn rates |
The Deploy page shows a live cost estimate before you commit, and Cost estimate explains how to read it.