# GPU servers

> NVIDIA H200, B200 and L40S servers with CUDA images, NVLink and 400 Gbps InfiniBand for training clusters.

Source: https://docs.avenlith.com/en/gpu-servers  
Category: AI & GPU  
Last updated: 2026-09-22

## Configurations

| Plan | GPUs | GPU memory | Best for |
| --- | --- | --- | --- |
| gpu-l40s-1x | 1 × L40S | 48 GB | Inference, rendering |
| gpu-l40s-4x | 4 × L40S | 192 GB | Mid-size inference |
| gpu-h200-8x | 8 × H200 SXM | 1,128 GB | Training, 70B+ models |
| gpu-b200-8x | 8 × B200 | 1,440 GB | Frontier-scale training |

## Deploy

```bash
avenlith gpu list-stock
avenlith server create --name train-01 --plan gpu-h200-8x --region mow1 \
  --image ubuntu-26.04-cuda-13
nvidia-smi
```

Images with the `-cuda` suffix include the NVIDIA driver, CUDA 13, cuDNN, NCCL and the container toolkit.

## Clusters

8-GPU servers in MOW1 and FRA1 can be connected with non-blocking 400 Gbps InfiniBand. Request a cluster of up to 64 nodes from our AI team.

> **Note:** > GPU capacity is limited. Reserve servers for 3 or 12 months to guarantee availability and save up to 35%.