What an 8-GPU node really costs

The newest cards are often sold only as whole machines of eight. The price page divides the machine price by eight and prints a per-GPU figure. That figure is useful for comparing machines. It is not something you can buy.

So the first question about any B300 or B200 price is how many GPUs come with it. The tables on this site print that beside each price as the smallest rental.

Multiply the per-GPU price by eight, then by the hours you will hold the machine. On 12 Nov 2026, the lowest listed hourly total for an 8-GPU B300 machine was $52.80. For the B200 it was $28.00, and for the H200 it was $27.92.

A day costs twenty-four times the hourly total, whether all eight cards are busy or one is. A node left running over a weekend is billed for the weekend.

A node is right when the job needs eight GPUs in one machine, linked closely enough to train one model across them. Then per-GPU prices compare fairly, because every option is a node.

It is wrong when the job needs one or two GPUs. You would pay for six cards to sit idle. Look instead for a provider that rents the same GPU singly, even at a higher per-GPU figure. The hourly total is what you pay.

Ask five things. Is the node on demand, or does it need a term? What unit is it billed in? Are the CPU cores, memory and local disk in the price? Can a new account start one today, or is there a quota or a sales step? And when was the price read?

Runcrate lists each GPU with its own price per GPU-hour and bills by the second on most models. Its smallest rental for each GPU is in the price tables on this site, read from its live feed.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.