NovaPodRetail floors and rental asks diverged again this week. A US buyer looking at an RTX 5090 was being shown about $9,400 for a card. The same name on this marketplace, checked November 3, started at $0.26 an hour, with 648 offers and 32 GB. Those are not competing prices for the same decision.
I wrote down the asks I would actually click. They will move. If you are reading this in a later week, open the card.
The asks I pulled on November 3
These are the live minimums on NovaPod, not a median of every cloud. Offer counts are what the market list returned with the price.
An H100 SXM was $1.21 an hour, median about $1.68, 65 offers, 80 GB. An H200 was $2.76, 82 offers, 141 GB. The H200 NVL row was actually cheaper at the floor, $2.29, with 76 offers, same 141 GB. A B200 was $4.38, median about $6.30, 51 offers, 192 GB. A B300 was $6.95, 15 offers, 288 GB. The GB300 NVL72 slice was $8.40 to start and $11.20 in the middle, four offers.
| GPU | VRAM | From | Median | Offers |
|---|---|---|---|---|
| RTX 4090 | 24 GB | $0.09/hr | $0.35/hr | 419 |
| RTX 5090 | 32 GB | $0.26/hr | $0.46/hr | 648 |
| A100 SXM4 | 80 GB | $0.30/hr | $0.46/hr | 158 |
| H100 SXM | 80 GB | $1.21/hr | $1.68/hr | 65 |
| H200 NVL | 141 GB | $2.29/hr | $2.91/hr | 76 |
| H200 | 141 GB | $2.76/hr | $2.76/hr | 82 |
| B200 | 192 GB | $4.38/hr | $6.30/hr | 51 |
| B300 | 288 GB | $6.95/hr | $6.95/hr | 15 |
How that sits against other public indexes
A November cloud index at gpu.ai had H200 SXM secure pricing around $5.29 an hour and B300 around $8.99, with the RTX 5090 marked out of stock on that board. A November 2 market-median file from another tracker put H200 near $4.25 an hour, H100 SXM near $2.40, B200 near $8.75, and B300 near $10.00. I am not auditing their scrapers. I am saying the direction matches: Hopper is a few dollars, Blackwell Ultra is several dollars more, and a 32 GB GeForce is a different shelf.
Our H200 floor at $2.76 is under both of those public marks. Our B300 floor at $6.95 is under the $8.99 and $10 marks I just cited. That can change before lunch. The useful habit is to read offer count with the dollar. Fifteen B300 offers is a real book. It is not eighty-two.
Which hour I would buy
A 7B to 14B model, a LoRA that already fits, or a batch of images: start on the RTX 4090 at $0.09 or the 5090 at $0.26. You are paying for 24 or 32 GB and a lot of offers. Stop the pod when the batch is out.
A 70B that is memory-bound on 80 GB: the H100 SXM at $1.21 is the first serious row. Move to the H200 at $2.76 when the context or the batch does not fit in 80 GB. The extra 61 GB is the product. The brand name is not.
B200 at $4.38 is the 192 GB step. B300 at $6.95 is the 288 GB step. I would not rent a B300 because a slide said Blackwell Ultra. I would rent it when the job already spills H200. The GB300 NVL72 at $8.40 is a rack slice, four offers, and an estimate-adjacent scarcity. Read the card before you treat it as idle metal.
The street card and the hour are not the same purchase
Nine thousand four hundred dollars for a 5090, the October US floor from the retail survey, is about 36,000 hours at $0.26. Nobody runs one card that many hours without a reason. The people paying $9,400 want the card on a desk. The people paying $0.26 want the hour. If you need the hour, do not finance the desk.
Billing here is still the ask times the seconds the pod runs. The deposit is crypto. The number on the card is not a monthly seat. I checked these asks on November 3, 2026. Tomorrow's list is allowed to disagree.