A ledger for one GPU

What it covers

On-demand H200 prices per GPU-hour, shown per GPU and per server, from the pricing pages of GPU clouds, marketplaces and hyperscalers. Spot and reserved listings sit on their own page. Comparison pages set the H200 beside other GPUs wherever both have at least three providers with a price.

Three working pages sit beside the ledger: the cost of an 8-GPU server, what 141 GB of memory holds, and what a monthly budget buys.

How the ledger is ordered

Rows are grouped by GPUs per server. The ladder on the front page ranks each provider by its lowest solid on-demand price, whatever the size of that configuration, and the step between neighbouring rungs is the difference in dollars.

Solid means the provider sets the price. Marketplace figures set by hosts are shown with a chip and left out of every ranking and statistic. Prices that start later are shown as announced and are not counted until they take effect.

Spot, reserved and the month

Statistics use on-demand rows only. Spot and reserved prices are labelled and never mixed into the on-demand figures. A month in any calculation is 720 hours, which is 30 days of 24 hours.

What it does not cover

Availability, speed, tax, storage, network egress and support. A listed price is not a promise that GPUs are free when you ask, and this is one snapshot with no price history.