How to read a GPU price page before you rent

Five passes before the number

Before you trust the hourly figure on a GPU price page, find five things: the unit it is quoted in, the commitment it assumes, the exact card including form factor and memory, the region it applies to, and what else the price includes. A rate that fails any of these cannot be compared with another provider's rate, and a rate that cannot be compared is how a first invoice surprises people.

Do the passes in that order. The unit and commitment are the ones that change the figure the most, and they are also the ones a page is most likely to leave to a footnote.

Keep the provider's page open beside a blank row of your own sheet and fill a cell for each pass. A pass you cannot fill from the page is a question to send, and the answer belongs in the same row as the rate.

Pass one: what the price is per

Look for "per GPU", "per instance" or "per node" beside the figure. An instance with several cards at one price is not cheap per hour; divide the price by the number of cards to get the figure you can compare. A node price that looks high next to a per-card price may be the cheaper of the two once you divide.

Then look for the time unit and the billing increment. Per hour is common, but some pages quote per second or per minute, and some bill a minimum. Where the page is silent on partial hours, assume they are rounded up until you hear otherwise.

This matters most for short work. A ten-minute test repeated through a day costs a few hours of card time under per-second billing and a great deal more under hourly rounding, for exactly the same work.

Pass two and three: the commitment and the card

A single table can mix on-demand, reserved and interruptible rates under one heading. Find the label on each row. A reserved rate is only available if you sign the term, and an interruptible one can be taken back while it runs, so neither belongs beside another provider's on-demand rate.

Then read the card name to the end. "A100" might be the 40 GB or the 80 GB part, SXM or PCIe. On this site an A100 rate listed with no form factor sits on its own page (cheapest on record: $1.09), apart from the A100 SXM 80GB (cheapest on record: $0.41), because the two cannot be assumed to be the same purchase. Where the provider's page does not say which A100 it is, ask before you rent.

The same applies to the H100. The SXM part and the PCIe part share a name, sit in different machines and are priced differently. A page that says only "H100" has told you less than it seems.

Pass four and five: the region and what is bundled

Some providers charge different rates by region, and a headline figure may be for the cheapest one. If your data has to stay in one jurisdiction, or your users are in one place, find the rate for that region and use it.

Then find what the rate includes. CPU cores, system memory, local disk and a network allowance are bundled by some providers and charged separately by others. Storage that persists after the instance stops and data moved out of the provider are two extras that can appear on a bill without appearing on the price page. A rate that bundles generous local disk can be cheaper in practice than a lower rate that does not.

When the region you need is missing from the page, ask before you plan around the rate. A headline figure for a region you cannot use is not a price you can pay, and a provider that sells in your region at a higher rate may still be the cheaper option once the cost of moving data across borders is counted.

When the page leaves something out

Most price pages answer some of the five passes and not all. A missing answer is not a hidden cost by default, but it is a question you have to ask before the invoice answers it for you. Write the question down as you read, so you ask every provider the same thing.

Common gaps: no billing increment, no statement of what CPU and memory come with the card, no word on egress, a card name without its memory size, and no date on the rates. The last one matters more than it looks. A page with no date may have been updated yesterday or a year ago, and you cannot tell which.

When a provider will not answer in writing, use the more expensive possibility in your estimate. The estimate you make before renting should be the one you can afford if the unknowns go the wrong way.

Writing down what you found

Make one row per provider with the same columns: card exactly, per GPU-hour on-demand rate, billing increment, region, what is included, and the date you read the page. The date matters because pages change; a rate you copied a month ago is a rate you should check again.

The table of every GPU rate is built the same way: per GPU-hour, on-demand, the exact card, the source page and the date it was read, with any rate older than 21 days hidden rather than shown stale. How those figures are gathered is written down. Use it for the first pass, then read each provider's own page for the four things a rate table cannot carry.