What a GPU-hour costs, and why one chip has three prices

Ask three providers what a B200 costs per hour and you will get three honest answers. On 27 September 2026: $4.86 on a twelve-month contract, $6.69 by the hour at a neocloud, $14.24 on demand at AWS. A B300 runs $6.25, $7.85 and $17.80 on the same three routes.

Why they differ

Commitment. A contract moves the risk of an empty machine from the provider to you. You are paid for taking it, in a lower rate.

Delivery. A block that arrives set up, networked and tested costs more than a bare virtual machine. Hyperscalers add credit, support and a bundle of services on top, and charge for all of it.

Scarcity. When on-demand capacity is sold out, and in 2026 it mostly is, the hourly price is whatever the last buyer would pay.

Which price you should pay

A rule of thumb. If you know you need the machines for more than a year, sign the contract. If you need them for one to three months, buy a window. If you need them for hours, use a marketplace and accept that the box may disappear.

One warning: term is not always cheaper. On the H100, an older chip, a one-year contract was $2.10 to $2.70 in spring 2026 while marketplaces rented the same card for $1.30 to $1.90 by the hour. As a chip ages, the contract can end up above the spot price. Buy term on the chip you will still want next year.