Five steps
- 1. Size the model. Parameters × bytes per parameter (2 for 16-bit, 1 for 8-bit, 0.5 for 4-bit) plus 20–50% for cache and activations.
- 2. Pick the GPU. Up to 96 GB per card: RTX PRO 6000. Training and NVLink: H100. Large-model inference: H200. Frontier scale: B200.
- 3. On demand or reserved. Per-minute billing for tests; 3–36 month terms for anything that runs all month.
- 4. Pick a region. Near your users or where your data must stay.
- 5. Deploy. Use a PyTorch/CUDA image or your own container; our engineers help with the first deployment.
Not sure which GPU? Read how to choose a GPU for your AI workload or send us your workload.

