On 25 September 2026 we read every rentable on-demand offer for nine GPU models on a public marketplace. The medians formed a clear ladder: RTX 4090 at $0.49 an hour, RTX 5090 at $0.56, A100 SXM at $0.67, H100 SXM at $2.27, H200 at $4.71 and B200 at $7.94.
$2.27
median H100 SXM hour
$4.71
median H200 hour
$7.94
median B200 hour
Cheapest is not typical
The cheapest H100 offer that day was $1.79 an hour, about a fifth below the median. Cheap offers are real, but they are few, and they tend to sit in one region. That is why the index shows both the minimum and the median, and where the cheapest offer is located.
What the ladder says
Each step up the ladder buys more memory and faster interconnect, not only more speed. A model that fits on a smaller card is often far cheaper to serve there. Newer generations cost several times more per hour, which only pays off when the job needs what they add.
- Prices are a snapshot of one open market, not a quote.
- COREVANE does not rent or resell capacity.
Figures come from public sources on the date shown: Robinhood Chain, Chainlink and a public GPU marketplace. This article is general information, not investment advice.