Guide5 sections1 min read
The short answer
The cheapest GPU is the smallest one that runs your job well. Fit the model in memory first, then compare cost per completed job, then bid below the current offer if you can wait.
Right-size the memory
Estimate memory with the formula in the VRAM guide. A model that fits in 24 GB does not need an 80 GB GPU; quantizing to INT8 or INT4 can bring a model down to a smaller and cheaper card.
Compare cost per result
Measure runtime on two or three candidates with the same software. Total cost equals hourly rate × hours: a faster GPU can be cheaper per job even at a higher rate.
Bid instead of taking the first offer
On EverXCloud, you can place a bid at the price you are willing to pay per GPU-hour. It is matched when an offer meets it, which suits jobs that can start later.
Stop paying when the job ends
Download results and release the instance as soon as the job finishes. Review persistent storage separately, as it can remain billable. Accounts are prepaid, so spending never exceeds your balance.
Live prices on EverXCloud
| GPU | Memory | From, per GPU-hour |
|---|---|---|
| NVIDIA GeForce RTX 3090 24GB | 24 GB | $0.22 |
| NVIDIA GeForce RTX 4090 24GB | 24 GB | $0.33 |
| NVIDIA L4 24GB | 24 GB | $0.44 |
| NVIDIA A10 24GB | 24 GB | $0.24 |
| NVIDIA A100 SXM 40GB | 40 GB | $0.34 |
| NVIDIA L40S 48GB | 48 GB | $0.79 |
Best live offer on the exchange, excluding tax. Prices are indicative during pre-launch.
Frequently asked questions
- What is the cheapest GPU for AI?
- The one with enough memory for your model at the lowest cost per completed job. 24 GB GPUs such as the RTX 3090, RTX 4090, L4 or A10 are often the cheapest options for models that fit.
- Can I set my own price?
- Yes. Place a bid at the price per GPU-hour you are willing to pay; it is matched when an offer meets it.
- Are the prices shown final?
- No. During pre-launch, EverXCloud prices are indicative. They are shown per GPU-hour, excluding tax, and become rentable on launch day.