Pricing models
On-demand, spot and reserved compute prices
One GPU carries three prices, on demand, spot and reserved, and only the on-demand price compares across providers without reading the terms.
GPU compute prices
Answer in brief
Pricing models
One GPU carries three prices, on demand, spot and reserved, and only the on-demand price compares across providers without reading the terms.
Billing
The billing unit decides what a short job costs: the same hourly price bills a ten-minute run as ten minutes, or as a full hour.
Rental size
A per-GPU price inside an 8-GPU node is one eighth of what you pay: the node is the smallest thing you can rent, so the hourly bill is eight times the listed figure.
GPU prices
The B300 is the newest and most expensive NVIDIA GPU in these tables, and the one whose listed price is least likely to be a price you can start.
GPU prices
The B200 is the practical Blackwell GPU for most buyers in 2026: priced below a B300, with more memory than an H200.
GPU prices
The H200 is the last Hopper GPU and, in 2026, the easiest large datacenter card to rent by the hour: many providers list it and several rent a single one.
GPU prices
The RTX Pro 6000 puts 96 GB on a single Blackwell card for less per hour than the datacenter GPUs in these tables, and most providers rent it one GPU at a time.
Availability
A GPU on a price page can fail to start for three reasons: the provider is sold out, the card is gated behind sales or approval, or your account has no quota for it.
Billing
The GPU-hour is the largest line on a compute bill but rarely the only one: storage, data leaving the provider and hours a machine sat idle are billed too.
Practical
You can test any compute price in five minutes: find the unit, find the smallest rental, try to start it, and write down the date.
Checklists are original editorial guidance. Published prices do not establish availability.
Compute Price Watch. “Topics.” Content reviewed 12 Nov 2026. https://computepricewatch.com/topics