Per card or per node: pick one unit for your GPU prices

The unit buyers compare in

Publish GPU prices per GPU per hour. Buyers compare cards across providers, and the only unit that lines up across every page is the cost of one card for one hour. A price for a whole node or a bundled instance can be correct and still useless in a comparison, because the buyer has to divide it by the card count before it means anything, and the next tab may need no arithmetic at all.

You can show the node price too. Put the per-GPU figure first and the node figure beside it, and the buyer never has to do the arithmetic you could have done for them.

The division a buyer will not do

A buyer with a spreadsheet of H200 rates has a column of per-card figures; the lowest listed hourly rate in it is $2.09, from Beam. Your page shows a single figure for a node. To place it in the column, the buyer has to find how many cards the node holds, divide, and trust that the node price contains nothing else. Three steps, each a chance to give up.

Worse, the undivided figure is sometimes pasted straight into the sheet. A node price sitting in a column of card prices looks absurdly dear, the row is sorted to the bottom, and nobody looks at it again.

The reverse mistake happens too. A buyer reading a per-card rate as a node rate thinks they have found a machine for the price of one card, and then feels misled when the invoice arrives.

Working the arithmetic in public

The arithmetic is simple, which is why it belongs on the page. Take the lowest listed rate per GPU-hour, which is $2.09: one H200 run at that rate for a month of 730 hours costs $1,525.70. A node of several cards costs that figure times the card count, before anything else is added. Writing both numbers down lets a buyer check one against the other.

If a node price includes CPU, system memory and local disk that a per-card price elsewhere does not, the division alone flatters the competitor. In that case say what the node includes, so the buyer can see the per-GPU figure carries more than the card. The GPU cost calculator multiplies hourly rates by hours and card counts, and it uses per-GPU rates because that is the unit every row can be put in.

Where the comparison breaks

We compare listed on-demand rates per GPU per hour for an exact card. The H200 SXM 141GB page holds rates in that unit, and the count of sellers there is 26. A rate published only per node has to be divided before it can stand in that list, and the division depends on knowing the card count from the same page.

Instances that bundle a fraction of a card, or a card with a fixed slice of a host, raise the same problem from the other side. If you sell part of a GPU, say what part, in memory and in share of the card, and give the rate for that slice clearly marked as a slice. It is a different product and should not sit in a table of whole cards.

A layout that answers the question once

One row per card type. The first figure is the price per GPU per hour, labelled that way. Beside it, the node configuration in words, such as the card count and the memory per card, then the node price per hour. Last, a short line saying what the node includes besides the GPUs.

Billing should follow the same unit, or say clearly where it does not. A provider that prices per card but only rents whole nodes should state the minimum, because a buyer who needs one card will otherwise discover the minimum at checkout.

The billing increment belongs on the same row. Per second, per minute or per hour changes what a short job costs, and a buyer running many short tests will multiply the rounding by the number of runs. A rate per GPU per hour with a per-second increment is a different offer from the same rate billed in whole hours, and the buyer can only tell them apart if the page says which.

What a node price has to disclose

A node price covers more than the cards, and a buyer dividing it needs to know what else is in it. The number of CPU cores, the system memory, the local disk and its type, and the network between nodes if the buyer can rent more than one. These are the things another provider may bill separately, so they are the things that make two per-GPU figures unequal.

Say how the cards are connected inside the node, in words. SXM modules sit on a board with a fast link between them; PCIe cards sit in ordinary slots. A node of SXM cards is a different product from a node of PCIe cards with the same name, and the per-GPU figure cannot show that on its own.

Then say whether the node can be rented in part. If a buyer can take half of it at half the price, the per-GPU rate is real. If they cannot, the per-GPU rate is a way of describing a price they can only pay in full, and the page should make that plain on the same row.