GPU pricing, card by card
GPU prices are quoted per hour, but the card sets the scale: a consumer card such as the RTX 4090 rents for a fraction of an A100, which rents for less than an H100 or B200. For one card, the spread comes from the tier, the region and what the hourly rate includes. The bill also carries what the rate leaves out, such as disk, volumes, data transfer and time spent loading. rent.dev's price board lists hourly GPU rates by card, normalised per GPU, and GPU pods run on your own key at the GPU provider rent.dev connects to.
rent.dev: one API for the sandboxes and GPU pods an agent needs.
The most-offered cards, smallest memory first
| GPU | Memory | Offers | Lowest per GPU hour | Lowest at | Highest per GPU hour | Rate read |
|---|---|---|---|---|---|---|
| Tesla V100 16GB | 16 GB | 8 | $0.122 | Verda, spot | $0.244 | 2026-11-11 |
| g6 | 24 GB | 8 | $0.8048 | AWS | $3.3968 | 2026-11-11 |
| g5 | 24 GB | 8 | $1.006 | AWS | $4.096 | 2026-11-11 |
| RTX A6000 48GB | 48 GB | 8 | $0.4305 | Verda, spot | $0.861 | 2026-11-11 |
| RTX 6000 Ada 48GB | 48 GB | 12 | $0.779 | Verda, spot | $1.714 | 2026-11-11 |
| L40S 48GB | 48 GB | 12 | $0.999 | Verda, spot | $2.198 | 2026-11-11 |
| g6e | 48 GB | 8 | $1.861 | AWS | $7.57719 | 2026-11-11 |
| A100 SXM4 80GB | 80 GB | 13 | $1.1695 | Verda, spot | $4.09625 | 2026-11-11 |
| H100 SXM5 80GB | 80 GB | 16 | $2.5425 | Verda, spot | $12.29 | 2026-11-11 |
| RTX PRO 6000 96GB | 96 GB | 12 | $1.376 | Verda, spot | $3.0275 | 2026-11-11 |
| H200 SXM5 141GB | 141 GB | 12 | $3.252 | Verda, spot | $7.155 | 2026-11-11 |
| B200 SXM6 180GB | 180 GB | 12 | $4.6145 | Verda, spot | $10.1525 | 2026-11-11 |
| B200 CC SXM6 180GB | 180 GB | 12 | $4.70625 | Verda, spot | $9.415 | 2026-11-11 |
| B300 SXM6 268GB | 268 GB | 12 | $6.025 | Verda, spot | $13.26 | 2026-11-11 |
| B300 CC SXM6 268GB | 268 GB | 8 | $6.145 | Verda, spot | $12.29125 | 2026-11-11 |
Questions
- How does cloud GPU pricing compare by card and hour?
- Rank cards by the work first, then by price. Memory decides what fits: a 24 GB consumer card runs smaller models, while an 80 GB data-center card holds larger ones without splitting them across GPUs. Compare per-GPU hourly rates for the same card and memory size, and remember that a faster card at a higher rate can finish a job for less. Taking each card at its lowest rate on rent.dev's price board, generated 2026-11-14, the cheapest is the gtx1050 2Gi PCIe (2 GB) at $0.02 per GPU hour and the dearest the GB300 (288 GB) at $18.00 per GPU hour.
- What hidden fees affect GPU pricing?
- Persistent volumes and container disk bill while they exist, often after the GPU has stopped. Data transfer out of the provider, public addresses and minimum billing increments add more. Idle time is the largest: a GPU left running between jobs bills at the full rate. Price the whole job, including storage and the time spent pulling images and loading weights.
- Where can students get cheap GPU credits?
- Start with what costs nothing: university clusters and research computing centres, the free GPU tiers of hosted notebook services, and the education and research credit programs that large clouds and GPU vendors run for students and academic labs. Hackathons and open-source programs hand out credits too. Read when a credit expires and what happens when it runs out, since some programs move to normal billing once it is spent.
- Is it cheaper to rent a GPU or buy one?
- Renting wins for occasional or bursty use and for cards you could not justify owning. Buying wins when a card would be busy most hours for a year or more, once power, cooling and the work of running it are counted. Many teams rent for experiments and buy only for a steady base load.