On 2026-09-25 eleven clouds we read publish an on-demand rate for the NVIDIA B300. The cheapest is $4.95 per GPU-hour at Sesterce, listed as a from rate; the dearest is $17.80 at Amazon Web Services in us-east-1. The median rate card — each provider counted once, at its own cheapest on-demand rate — is $8.20 per GPU-hour, and the middle half of those rate cards sits between $7.63 and $9.86.
Amazon’s B300 rate is one day old
Our first read of it was at 04:38 UTC on 2026-09-25; before that read, the dearest B300 on our page was Oracle’s $15.00. AWS prices the card as a whole machine: the p6-b300.48xlarge is $142.416 an hour with eight B300s attached, which is the $17.80 per GPU our table prints beside it.
Against the B200 in the same price list the step is exactly 1.25x: the eight-B200 p6-b200.48xlarge is $113.9328 an hour, $14.24 per GPU, against $17.80 for the B300. We read one region for AWS, us-east-1, and the figure is the Linux on-demand rate with no commitment attached.
The B300 is also thinner on the page than the card before it. Eleven clouds list an on-demand B300; nineteen list a B200, at a median rate card of $6.75 per GPU-hour over a range of $4.00 to $16.11.
What a 3.6x spread costs over a month
Between the two ends the spread is 3.6x. Eight cards running a full month at the cheapest rate — $4.95 x 8 x 730 hours — comes to about $28,908, and the same eight at $17.80 x 8 x 730 hours comes to about $103,964.
The two ends are not the same product. Sesterce’s $4.95 is an advertised from rate — a marketing floor on its offer table rather than the price of a configured instance; Amazon’s $17.80 buys an instance with its CPUs, memory and networking attached, and our tables tag which is which. Between them sit the specialists: Hyperstack at $7.40, RunPod’s Secure Cloud at $7.89 and the median rate card at $8.20.
Two hyperscalers, nine specialists
Two of the eleven clouds listing an on-demand B300 on 2026-09-25 are hyperscalers, Amazon Web Services and Oracle. Their median is $16.40 per GPU-hour against $7.89 at the specialists, 108% higher. That gap is a reading, not an explanation: what the dearer rate includes is stated on each provider’s own page, and availability is not something we measure at all.
Other markets go lower than any rate card. Nebius lists spot capacity for the B300 from $0.99 per GPU-hour, a rate for a machine that can be reclaimed mid-run, and RunPod’s community capacity is $6.94, which our risk tag marks the same way, because community capacity can be reclaimed mid-run as well.
How to check any number here
Every figure above is listed under this piece with the page it was read from and the time it was read, and the live B300 page carries the same rates with a capture time on every row. The medians, the spread and the hyperscaler gap are arithmetic on those same readings, computed the way /methodology states. Nothing here is forecast or modelled, and nothing here is advice: the piece states what eleven price pages said on 2026-09-25.