NovaPodI went looking for a B200 the same week the October market notes came out. The short version: the card is real, the queue is real, and a lot of people are paying a Blackwell price for a Hopper-shaped job.
This is the note I would send a friend before they refresh a rental page for an hour. It is dated October 30, 2026. Tomorrow's ask can move.
What the October news actually said
In the week to October 10, the providers that publish stock had a B200 rentable in about 76% of checks and an H200 in about 76%. Older cards were much easier to find. The market split in two: a tight band for Blackwell and H200, and a noisy cheap floor for everything else.
On October 30, one cross-provider tracker put the B200 near $7.99 an hour on demand and about $4.38 on spot. Specialist clouds were posting roughly $5.99 to $7.99. The cheapest posted B200 in that early-October sample sat at $3.69 and did not move for seven days, while the typical provider's price rose. Hyperscaler Blackwell quotes were in another league, often described around $14 to $16 an hour.
How I decided between a B200 and the H200 I could start
B200 is the Blackwell datacenter GPU: 192 GB of HBM3e, about 8 TB/s, NVLink 5, and FP4 on the Transformer Engine. H200 stays on Hopper compute and gives you 141 GB at 4.8 TB/s. NovaPod's H200 catalog ask was $2.87 an hour when I looked. I do not have a NovaPod B200 hourly figure to freeze in this note. If a B200 is listed, the marketplace row is the price.
I stayed on the H200. The weights and the context fit. Paying toward a $4 to $8 B200 average so I could say I used Blackwell would have been a story, not a faster job. FP4 matters when the stack uses it and the model does not fit. It does not matter as a sticker.
| Question I asked | H200 | B200 |
|---|---|---|
| Memory | 141 GB HBM3e | 192 GB HBM3e |
| Why people want it | The H100 run ran out of memory | FP4, or the job does not fit in 141 GB |
| Public price I saw on Oct 30 | Spot about $3.09/hr | Spot about $4.38/hr; on-demand about $7.99/hr |
| NovaPod number I could cite | Catalog ask $2.87/hr | Only the live offer, if one is listed |
| How hard it was to book | Tight, about 3 in 4 stock checks | Same tightness, plus a higher rate |
What I liked once a machine was up
Per-second billing is the feature I notice. A two-hour eval does not become a day rate because I forgot the calendar. Crypto funding meant I was not waiting on a card that the bank wanted to "review." The template list, vLLM and the rest, got me to a first request without installing CUDA on an empty disk.
Stopping the pod is the whole trick. I say that because I have left one running. The clock does not care that I went to get coffee.
The parts that annoyed me
Scarcity is not a NovaPod invention. If the B200 you want is not on the list, no blog post creates it. The October checks already said Blackwell and H200 fail about a quarter of stock looks.
You need $100 on the account before a small test. That is clumsy when you only wanted twenty minutes. There is no card button. And the public $4.38 spot average will not match the machine you click. Write down the ask you accepted. Do not paste this table into a runbook and reuse it next week.