Amazon charges $98.84 an hour to reserve a Capacity Block of eight B200s, about $12.36 per GPU-hour, and says it moves that price on supply and demand. In our July 13 Vast.ai snapshot the B200 asking range rose to $7.25 to $9.92 an hour, up from about $5 a week earlier, while the upper end for last-generation H100 and H200 spot offers eased. Two GPU generations, moving apart.
It is tempting to read that as the shortage narrowing onto Blackwell. Be careful. These are asking-price ranges on one marketplace, and the snapshot merges board variants like H100 SXM and PCIe into a single range, so a move can reflect what is on offer rather than what is scarce. And hourly rent is not what a workload costs. A B200 rents for more than an H100 and can still produce a cheaper token, because throughput, batching, and memory fit decide cost per token, not the sticker on the hour.
So the honest read is narrow. Reserved rates are high across generations; one marketplace's Blackwell asks rose this week while its Hopper asks fell. For what to build on, the number that decides it is cost per token under comparable terms, not the hourly rate.
In the July 13 Vast.ai snapshot, B200 asks rose to $7.25–$9.92/hr from about $5 a week earlier while H100 and H200 upper ranges eased; these are marketplace asking prices, not cost per token.
