Reading the cost estimate
The Cost card sits beside the deploy form and updates as you configure. It answers one question — what does this machine cost per hour right now — and it is often read as answering more than that.
What the card shows
| Row | Where the number comes from |
|---|---|
| GPU (model ×N) | The offer's hourly price for the whole slice |
| Storage (size) | Your disk selection × the machine's per-GB-per-hour storage rate. Moves when you move the slider |
| Total | The two rows added together |
| ≈ per day | Total × 24 |
| Your balance | Your organization's current credit balance, not a projection |
Only one meter runs at a time
The total adds two rates that never bill at the same moment:
| Instance state | What is metered |
|---|---|
running | The GPU rate, per second |
stopped | The disk, per GB per hour |
creating, starting, stopping, destroying | Nothing |
destroyed | Nothing |
So the Total row is a ceiling, not a forecast. An hour of running costs the GPU line; an hour stopped costs the storage line. You never pay both for the same hour.
The same two numbers reappear on the instance page as Burn rate (running) and Burn rate (stopped), which is the clearer framing once the instance exists.
What the estimate does not include
- The storage bill after you stop. The card is a rate for one instance as configured, and the per-day row is that same ceiling multiplied by 24. Stopping swaps the GPU meter for the storage meter rather than ending the bill, and nothing on this page projects how long you will leave it stopped. That is the charge people do not see coming.
- Anything else your organization is running. Credits are shared. The balance row is today's number, not what is left after your teammates' instances settle.
- Provisioning time, which is free, and equally a deploy that fails, which costs nothing.
- Later price changes. The rates are copied onto the instance when it is created, so an instance keeps the price it was deployed at even if the catalog moves. The estimate reflects the catalog price at this moment.
Sanity-check a monthly figure
Work in whole hours and use a 30-day month — 720 hours — which is the convention the console already uses when it quotes storage per month.
Suppose the card reads:
GPU (H100 SXM ×2) $1.80/hr
Storage (200 GB) $0.04/hr
Total $1.84/hr
≈ per day $44.16
Three figures fall out of that, and they are very different:
| Scenario | Arithmetic | Month |
|---|---|---|
| The card's ceiling | $1.84 × 720 | $1,324.80 |
| Running continuously | $1.80 × 720 | $1,296.00 |
| Running 8 hours a day, stopped the rest | ($1.80 × 240) + ($0.04 × 480) | $451.20 |
Two habits keep the estimate honest:
- Multiply before you deploy, not after. If the monthly figure is larger than the balance you are willing to hold, choose a cheaper slice or a smaller disk now.
- Check the first hour against reality. An hour after the instance reaches
running, Spend so far on the instance page should sit close to the GPU rate. Charges settle about once a minute, so the number climbs in small steps rather than continuously. If it is far off what you predicted, find out why on the billing page before it runs for a week.
At a zero balance your running instances are stopped automatically — and a stopped instance still bills its disk, against a balance that is already empty. At −$5 the instances are destroyed and the disks go with them. See Auto-stop and auto-destroy.