# GPUBank

> Cloud GPUs rented by the second, priced all-in (vCPU, RAM, NVMe, one public IPv4 and traffic included), in Paris, Frankfurt, Ashburn and Singapore. The meter starts after the health check passes and stops the second you stop. Persistent disk is the only other line: $0.07 per GB-month.

Operator: GPUBank SAS, Paris, France. Support: support@gpubank.io.

## GPUs and on-demand prices (USD per hour, all-in)

- B200 SXM (180 GB HBM3e, NVLink 1.8 TB/s): $4.90/h, $0.001361/s. 28 vCPU, 283 GB RAM, 2 TB NVMe included. Best for: Training and frontier inference.
- H200 SXM (141 GB HBM3e, NVLink 900 GB/s): $3.25/h, $0.000903/s. 24 vCPU, 276 GB RAM, 2 TB NVMe included. Best for: Large-context inference, 70B+ fine-tuning.
- H100 SXM (80 GB HBM3, NVLink 900 GB/s): $2.45/h, $0.000681/s. 20 vCPU, 125 GB RAM, 2 TB NVMe included. Best for: Training, fine-tuning, production inference.
- H100 PCIe (80 GB HBM2e, PCIe Gen5): $2.05/h, $0.000569/s. 16 vCPU, 188 GB RAM, 1 TB NVMe included. Best for: Single-GPU training and inference.
- RTX PRO 6000 (96 GB GDDR7 ECC, PCIe Gen5): $1.50/h, $0.000417/s. 16 vCPU, 188 GB RAM, 2 TB NVMe included. Best for: 70B inference in 4-bit, video generation.
- A100 SXM (80 GB HBM2e, NVLink 600 GB/s): $1.15/h, $0.000319/s. 16 vCPU, 125 GB RAM, 1 TB NVMe included. Best for: LoRA, QLoRA, batch inference.
- L40S (48 GB GDDR6 ECC, PCIe Gen4): $0.95/h, $0.000264/s. 16 vCPU, 94 GB RAM, 1 TB NVMe included. Best for: Image and video, Whisper, embeddings.
- MI300X (192 GB HBM3, Infinity Fabric): $2.30/h, $0.000639/s. 24 vCPU, 240 GB RAM, 2 TB NVMe included. Best for: Large models in one card, ROCm.
- RTX 5090 (32 GB GDDR7, no NVLink): $0.65/h, $0.000181/s. 9 vCPU, 35 GB RAM, 500 GB NVMe included. Best for: FLUX, SDXL, video, small LLMs.
- RTX 4090 (24 GB GDDR6X, no NVLink): $0.45/h, $0.000125/s. 8 vCPU, 41 GB RAM, 500 GB NVMe included. Best for: ComfyUI, SDXL, LoRA on 8B models.

## Regions

- EU West · Paris (eu-paris): H200 SXM, H100 SXM, RTX PRO 6000, A100 SXM, L40S, RTX 5090. Compute, volumes, backups, logs and support stay in the EU.
- EU Central · Frankfurt (eu-frankfurt): B200 SXM, H100 SXM, H100 PCIe, A100 SXM, L40S, RTX 4090. Compute, volumes, backups, logs and support stay in the EU.
- US East · Ashburn (us-ashburn): B200 SXM, H200 SXM, H100 SXM, H100 PCIe, RTX PRO 6000, A100 SXM, L40S, MI300X, RTX 4090. US residency. EU disks are never replicated here unless you move them.
- APAC · Singapore (ap-singapore): H200 SXM, H100 SXM, A100 SXM, L40S, RTX 5090. APAC residency, low latency to South-East Asia and Australia.

## Billing rules

- Per-second billing on every instance type; no minimum.
- Launch and health check are not billed; a failed start is not billed.
- Stopped instances are billed for the disk only. Delete the disk and nothing is billed.
- Interruptible capacity: about 40% below on-demand, five minutes of notice, stopped never deleted.
- Reserved capacity: fixed price for a week or a month, physically allocated.

## Machine-readable

- Price feed: https://gpubank.io/prices.json

## Legal

- Terms of service: https://gpubank.io/terms
- Privacy policy: https://gpubank.io/privacy