Egress and storage: the extras on a GPU bill

The two extras most likely to surprise a GPU renter are storage that bills while the instance is off and fees for moving data out. Ask about both first.

INFRATOAPP · PUBLISHED 26 OCTOBER 2026 · 5 MIN READ · 3 SOURCES · PRICES LAST CHECKED 18H AGO

What the hourly rate leaves off

The hourly GPU rate pays for the card while the instance runs. It usually does not pay for persistent storage, which keeps billing per gigabyte after you stop the instance, or for egress, the charge some providers make for data you move out of their network. Those two, plus extra public addresses and snapshots, are the usual reasons a GPU bill is larger than hours times rate.

The rental cost calculator leaves all of them out: it multiplies count, hours and the published hourly rate, and nothing else. This site holds no storage or egress prices, so these are questions for the provider.

Asking early also changes the answer you get. Before you have data on a provider, moving away is cheap, and its terms on storage and egress are part of the choice. After a few terabytes are sitting on its volumes, the same terms are a cost of leaving, and you are no longer comparing.

Storage that outlives the instance

A GPU instance normally comes with some local disk that disappears when the instance is deleted. Anything you want to keep, such as a dataset, model weights or checkpoints, goes on a volume or in object storage, and that is billed by size and by month.

The trap is the stopped instance. Stopping the machine stops the GPU charge but not the volume charge, so a box you stopped in March and forgot about can still be on April's bill. A large dataset copied to three regions to chase cheaper cards is three storage bills.

Ask the provider four things: the price per gigabyte per month for a volume, whether a stopped instance keeps billing for its disk, whether snapshots are charged separately, and whether local disk is large enough to avoid a volume altogether for short jobs.

Egress, and why burned buyers ask first

Egress is data leaving the provider: downloading results, serving responses to users, copying checkpoints to another cloud. Some providers charge per gigabyte for it, some include an allowance, and some say nothing on the price page. Ingress, data coming in, is commonly free, which makes the first upload feel cheap and the move away feel expensive.

The size that matters is what you will move out, not what you bring in. A fine-tune that produces a few gigabytes of weights barely registers. An inference endpoint that sends large images or audio to users all month, or a training pipeline that ships checkpoints to another cloud every hour, can make egress a real part of the bill.

A price page that does not mention egress has not told you it is free. Ask in writing, and keep the answer with the invoice.

Smaller charges that add up

Public IP addresses, load balancers, snapshots, container registry storage and support plans can each carry a fee. Each may be small. Across many experiments run by several people, they add up unless someone keeps a list.

Data transfer between regions of the same provider may be charged even when transfer to the internet is not, and the reverse also happens. When the data lives in one region and the cheap card is in another, check whether the saving survives the copy.

Container images are a quiet one. A large image pulled onto every new instance takes time on a billed card, and storing many versions of it in a registry can carry its own charge. Keep images lean and pinned to a version so a launch does not pull a fresh copy each time.

Keeping the extras small

Store one copy of anything large and pull from it, rather than copying datasets onto every new instance and volume. Delete checkpoints you will never load again. Keep the final weights and the few checkpoints you might return to, and let the rest go when the run ends.

Move less data out. Compress results before downloading them, evaluate on the provider's side and download the scores rather than the outputs, and serve users from the same provider as the model where you can. If you must move checkpoints between clouds, move them less often.

Put storage on the same list as instances. Whoever keeps track of what is running should also keep track of what is stored, because a forgotten volume bills as reliably as a forgotten machine and is harder to notice.

A short script for the sales call or the docs

Before the first launch, get answers to these: what persistent storage costs per gigabyte per month; whether stopped instances bill for disk; what egress costs per gigabyte and whether there is a free allowance; whether transfer between regions is charged; and what the billing increment is for the GPU itself.

Then estimate your own sizes: the dataset, the checkpoints you keep, and the data you expect to send out in a month. Multiply each by its rate and add it to the GPU figure from the calculator. The table of every GPU rate and how its figures are gathered cover the card; the rest is between you and the provider's own pages.

Ask the same questions of every provider you are considering, in the same words, and keep the replies together. Answers given in different shapes are hard to compare, and a reply that avoids one of the questions tells you something too. A provider that states its storage and egress terms plainly has made your estimate easier.

What we could not answer

⚑We hold no storage or egress prices for any provider, so this post cannot say what either costs anywhere.
⚑The calculator covers the GPU rate only and cannot be extended to these charges with the data we hold.

What this rests on

infratoapp, “Egress and storage: the extras on a GPU bill”, read 27 October 2026. https://www.infratoapp.com/blog/egress-and-storage-extras-on-a-gpu-bill

Questions people ask about this

What are the hidden costs of renting a GPU in the cloud?

The common ones are persistent storage, which keeps billing while an instance is stopped, and egress fees for data moved out of the provider. Snapshots, public addresses, transfer between regions and billing increments that round up to the hour can also add to the bill.

Do I pay for a stopped GPU instance?

Usually not for the GPU, but often for its storage. Stopping an instance ends the hourly card charge on most providers, while any attached volume continues to bill by size each month until it is deleted. Check the provider's terms for both.

Is GPU cloud egress always charged?

No. Some providers charge per gigabyte, some include an allowance and some do not charge. Policies vary enough that you should ask before renting and keep the answer in writing, especially if you will serve users or move checkpoints to another cloud.

These render as FAQPage structured data · answers are filled from the same live figures