NovaPodA friend asked me this week whether an H200 was still worth renting after the price jump. I did not want to send a screenshot from January. I opened the public trackers dated October 30, 2026, then I opened NovaPod's own catalog and wrote down what each one actually said.
I am not going to pretend I locked a year of capacity. I rented the way I usually do: pick the memory, read the ask, run the job, stop the machine. That is the feeling these notes are about.
What changed in the H200 rental price
Through mid-October, published rental indexes put the H200 about 160% above where it started the year. The B200 was up about 120% over the same stretch, and the H100 about 35%. The chips did not get slower. Demand got ahead of the machines people can actually book.
The tracker I used on October 30 put the H200 near $4.65 an hour on demand, about $3.09 on spot, and about $3.62 reserved, averaged across providers. Hyperscaler H200 rates in the same reporting sat roughly from $7.91 to more than $10.60. Those are market averages. They are not a coupon.
What NovaPod listed next to that
NovaPod's catalog, the one the price page reads, showed an H200 at $2.87 an hour, an H100 SXM at $1.82, an H100 PCIe at $1.26, an L40S at $0.77, and an RTX 4090 at $0.33. Billing on this site is still that ask, counted per second, until you stop the machine.
I like that the H200 line is only a bit under the public spot average, not a fantasy number from last winter. I do not like treating $2.87 as a promise. If the marketplace offer in front of you is different, pay the offer. This paragraph will be stale before the card is.
| Card | Public figure I checked | NovaPod catalog ask | What I used it for |
|---|---|---|---|
| H200 | Spot about $3.09/hr; on-demand about $4.65/hr | $2.87/hr | 70B weights plus a long context |
| H100 SXM | Up about 35% since January, still the default 80 GB card | $1.82/hr | FP8 serving when 80 GB is enough |
| RTX 4090 | Loose supply, noisy floor | $0.33/hr | A short LoRA or a ComfyUI batch |
What felt good
I did not talk to a salesperson. I put USDT on the account, picked the H200 because the job had already died on 80 GB, and the clock started when the machine did. When the run finished I stopped it. The seconds after that did not keep the hourly ask. That part matches how I want to work.
The memory is the reason I paid up from the H100. Same Hopper family, 141 GB of HBM3e instead of 80 GB. The job that failed was a context problem, not a FLOPS problem. Renting the cheaper card again would have repeated the failure.
What I would warn a friend about
The deposit floor is $100. A smaller transfer does not become time. Send the token on the chain the page shows. I have watched people send USDT on the wrong network and then wait for a credit that cannot arrive.
There is no card checkout yet. If you need a corporate Visa receipt this afternoon, this is the wrong desk. Credits do not turn back into a refund once you have agreed to the terms.
A stopped machine ends the GPU clock. It does not rewind the seconds you already used. And an article, including this one, is not the bill. Open the GPU list when you are about to launch.