The B200 GPU rental price is a Blackwell number, not a cheaper H200. On NovaPod's November 5, 2026 book the B200 started at $4.38 an hour. The median was $6.30. There were 51 offers. Each one listed 192 GB of HBM3e.
People searching this phrase want the hourly ask they will actually be billed. The floor is one host. The median is the middle of the book. Per-second billing means the clock stops when you stop the pod. A copied screenshot from another month is not today's offer.
The B200 GPU rental price against the cards next to it
The H200 on the same book started at $2.76 an hour, median $2.76, 82 offers, 141 GB. The B300 started at $6.95, 15 offers, 288 GB. The B200 sits between them: more memory than the H200, less than the B300, and a floor $1.62 above the H200.
Fifty-one offers is a real list and a thinner one than the H200's 82. I would read the median, $6.30, if the $4.38 host is gone or looks wrong. Budgeting only the floor is how a Blackwell job surprises the balance.
| Card | Memory | Floor | Median | Offers |
|---|---|---|---|---|
| H200 | 141 GB HBM3e | $2.76/hr | $2.76/hr | 82 |
| B200 | 192 GB HBM3e | $4.38/hr | $6.30/hr | 51 |
| B300 | 288 GB HBM3e | $6.95/hr | Not recorded | 15 |
What that hourly ask is paying for
A B200 is the Blackwell datacenter GPU: 192 GB of HBM3e at about 8 TB/s, NVLink 5 at about 1.8 TB/s, and FP4 on the Transformer Engine. A common node shape next to that memory is about 30 vCPUs and 184 GB of host RAM. The H200 stays on Hopper compute and stops at 141 GB and 4.8 TB/s.
The extra 51 GB matters when a model or a context already failed on the H200. FP4 matters when the serving stack can use it. Neither matters if the weights already fit in 141 GB and the job is not waiting on Blackwell numerics.
Blackwell GPU rental as a phrase covers more than this card. A B300 is 288 GB and started at $6.95. A GB300 NVL72 slice started at $8.40. Those are rack-scale rows. This page is the single B200 ask.
When the B200 is the wrong rental
Do not rent it to replace a 24 GB or 32 GB card. An RTX 4090 on this book started at $0.09 an hour, and an RTX 5090 at $0.26. Those are different jobs. A low hourly number on a card that cannot hold the model is a failed run.
Do not rent it because an H100 quote felt high. The H100 SXM floor was $1.21 for 80 GB. If 80 GB is enough, the B200 GPU rental price is about three and a half times that floor for memory you will not touch.
Stop the pod when the run ends. An idle Blackwell box is still a rental. The ask is the rate billed per second while the machine is up.
Questions
What is the B200 GPU rental price?
On NovaPod's November 5, 2026 book it started at $4.38 an hour. The median was $6.30, across 51 offers, each with 192 GB. The live marketplace row is the number you will be billed.
How does that compare with the H200 rental price?
The H200 started at $2.76 an hour the same day, with 141 GB and 82 offers. The B200 floor is $1.62 higher for 51 more gigabytes, Blackwell compute, and FP4.
Is the B200 the cheapest Blackwell GPU rental?
It was the Blackwell datacenter card with the lowest floor I recorded that day, at $4.38. It was not the cheapest GPU. The RTX 4090 floor was $0.09. A B300, with 288 GB, started at $6.95.
Does the price include a monthly contract?
No. NovaPod bills the ask per second from a crypto balance. USDT, USDC, SOL, ETH, and BNB can fund it. Stopping the pod stops the clock.
- B200 rental page
- H200 rental page
- Cheapest cloud GPU for Blackwell
- H100 vs H200 rental price
- Live GPU marketplace