GPUs and pricing
One fixed hourly price per GPU model, billed per second while your server runs. The price includes the machine’s CPU cores and RAM in proportion to your GPUs, network traffic and your disk.
| GPU | VRAM | Architecture | $/hour | ≈ $/month 24/7 |
|---|---|---|---|---|
| Consumer | ||||
| NVIDIA RTX 3060 | 12 GB | Ampere | 0.08 | 58 |
| NVIDIA RTX 4060 Ti 16GB | 16 GB | Ada | 0.12 | 86 |
| NVIDIA RTX 3080 Ti | 12 GB | Ampere | 0.13 | 94 |
| NVIDIA RTX 4070 Ti SUPER | 16 GB | Ada | 0.18 | 130 |
| NVIDIA RTX 5070 Ti | 16 GB | Blackwell | 0.22 | 158 |
| NVIDIA RTX 4080 / SUPER | 16 GB | Ada | 0.25 | 180 |
| NVIDIA RTX 5080 | 16 GB | Blackwell | 0.30 | 216 |
| NVIDIA RTX 3090 | 24 GB | Ampere | 0.25 | 180 |
| NVIDIA RTX 4090 | 24 GB | Ada | 0.40 | 288 |
| NVIDIA RTX 5090 | 32 GB | Blackwell | 0.65 | 468 |
| Workstation | ||||
| NVIDIA RTX A4000 | 16 GB | Ampere | 0.15 | 108 |
| NVIDIA RTX A5000 | 24 GB | Ampere | 0.22 | 158 |
| NVIDIA RTX A6000 | 48 GB | Ampere | 0.45 | 324 |
| NVIDIA RTX 6000 Ada | 48 GB | Ada | 0.75 | 540 |
| NVIDIA RTX PRO 6000 Blackwell | 96 GB | Blackwell | 1.40 | 1008 |
| Datacenter | ||||
| NVIDIA A10 | 24 GB | Ampere | 0.35 | 252 |
| NVIDIA L4 | 24 GB | Ada | 0.35 | 252 |
| NVIDIA A40 | 48 GB | Ampere | 0.40 | 288 |
| NVIDIA L40 | 48 GB | Ada | 0.75 | 540 |
| NVIDIA A100 40GB | 40 GB | Ampere | 0.80 | 576 |
| NVIDIA L40S | 48 GB | Ada | 0.85 | 612 |
| NVIDIA A100 80GB | 80 GB | Ampere | 1.10 | 792 |
| NVIDIA H100 80GB | 80 GB | Hopper | 2.00 | 1440 |
| NVIDIA H200 141GB | 141 GB | Hopper | 2.60 | 1872 |
| NVIDIA B200 180GB | 180 GB | Blackwell | 4.00 | 2880 |
Current prices and live availability are always in the dashboard and at GET /v1/gpus (API). A price change applies to new servers only; running servers keep the price they were launched with.
Several GPUs
Section titled “Several GPUs”A server with N GPUs costs N × the GPU price. For example, 4× RTX 4090 = $1.60/hour. See Multi-GPU.
How billing works
Section titled “How billing works”- Per second, only while the server is running. Pulling the image, downloading the model and starting are free.
- Nothing is charged while a host machine is offline.
- Launching needs a balance for at least one hour; at zero balance the server stops automatically.
- Stopped servers cost nothing; their data is kept for 7 days.
Example costs
Section titled “Example costs”| Task | Setup | Cost |
|---|---|---|
| Chat with Qwen3 8B through vLLM for a workday | 1× RTX 4090, 8 h | $3.20 |
| 1,000 SDXL images in ComfyUI (~3 s each) | 1× RTX 3090, ~1 h | $0.25 |
| LoRA fine-tune of a 7B model | 1× A100 80GB, 3 h | $3.30 |
| Serving Llama 3.3 70B for a day | 2× H100, 24 h | $96.00 |