Egress fees: the GPU cost buyers ask about before they sign

The question asked before the card

Egress is the charge for moving data out of a provider, and a GPU provider should state its egress policy on the pricing page, next to the GPU rates, in a sentence. Anyone who has trained a model and then paid to download the checkpoints, or served traffic and paid for every response, has reason to ask about egress before asking about the card. When the page says nothing, the cautious reading is that there is a charge and that it is large.

That reading is unfair to providers that charge nothing, and the page is the only thing that can correct it before the first invoice does.

The fix costs one line of text. The rest of this note is about what that line should say, what the detail behind it should cover, and why the buyer reads it before anything else on the page, including the rate itself.

Why the burned buyer asks first

GPU work moves a lot of data. Training sets go in, checkpoints come out, and a model under test may be copied between machines and regions several times in a week. Someone who once found that moving the results cost a real share of the compute bill learns to price the exit before the entry.

The hourly rate cannot show any of this. A provider can be the cheapest seller of a card per hour and still be the dearer choice for a job that writes large outputs, and anyone who has been caught once knows that. The egress line gets read before the rate, and where there is no egress line the choice is a support chat or another tab.

Serving makes it sharper. A model answering requests sends every response out of the provider, and an image or audio model sends a lot more than a few tokens of text. A buyer planning to serve from a rented GPU is pricing a stream of outbound data that never stops, and will want to know the cost per unit before the cost per hour.

What the calculator cannot add for you

The GPU cost calculator works out a rental bill from the published hourly rate and the hours the buyer enters. It does not add storage, network or egress, because those are not in the rates we hold, and it says so beside the result. A buyer using it gets the compute line of the bill and nothing else.

That makes the provider’s own page the only place the rest of the bill can come from. If the egress policy is on the page, the buyer adds it to the calculator’s figure themselves. Where it is not, the buyer adds a guess, and a cautious guess runs high.

We do not collect egress terms and do not compare them. How each figure is gathered lists what is read and what is not, and egress is in the second list.

How to state an egress policy

Say it in one sentence near the GPU rates, then link to the detail. "Data transfer out is free" is a full policy. So is "Data transfer out is charged per GB after a monthly allowance; see the network pricing page". What the buyer needs from the first glance is whether there is a charge, not the full schedule.

Then give the detail plainly: the unit, the rate per unit, any allowance, and whether traffic between your own regions or to object storage counts. If ingress is free, say that too. Buyers reading an egress section are looking for the sentence that will cost them money later, so give it to them before they go hunting.

If the policy differs by region or by product, say which rate applies to the GPU instances. A blanket network page written for every product a provider sells is hard to apply to one GPU rental.

What an unstated policy costs the seller

A provider that charges nothing for egress and does not say so loses the advantage. The buyer cannot credit a policy they cannot see, so the free egress earns nothing in the comparison where it would have helped most.

A provider that does charge and does not say so loses something harder to win back. The charge turns up on the first invoice, and the buyer who found it there remembers the invoice, not the rate. A clear line on the page converts that complaint into a cost the buyer accepted in advance.

The questions an egress paragraph should already answer

Is traffic to your own object storage counted as egress, or only traffic that leaves your network? Buyers who keep datasets and checkpoints in the same provider want to know that moving them between a GPU instance and a bucket is free, or exactly what it costs.

Is traffic between your regions counted? A buyer who trains in one region and serves in another will move every model they ship across that boundary. If it is charged at a different rate from traffic to the internet, say so, and say which applies to which.

What about inbound data, a public IP address, or a private link to the buyer’s own network? None of these is egress in the strict sense, and all of them are network charges a buyer will be asked to pay. A short list of what the network costs, in one place, beats a sentence on egress that leaves the rest to be discovered.